|
|
Log in / Subscribe / Register

Debian weighs eight options in vote on LLM usage

By Joe Brockmeier
August 19, 2026

The Debian Project is voting on the usage of large language models (LLMs) to make contributions to the project. The first proposal, sent in late July by Matthias Geiger, would expressly forbid any contributions to Debian that are created by or with the assistance of LLMs. That kicked off a firestorm of discussion and a flood of alternate proposals. Debian developers are now voting on eight proposals in total that range from banning LLM-assisted contributions to explicitly approving them, as well as the standard "none of the above" option that would leave Debian with no agreed policy.

Debian and AI

While other distributions, such as Fedora and Gentoo, have settled their policies regarding AI-assisted contributions, Debian has held off on making a decision—though not for a lack of discussion. The topic has come up several times over the years without any resolution.

In 2024, Tiago Bortoletto Vaz worried that Debian was "already facing negative consequences" due to the use of AI-generated content and wanted to gather input from Debian developers ahead of a possible general resolution (GR) on the topic. The discussion led him to conclude that the project was "far from consensus on an official Debian position" regarding the use of generative AI, and he did not pursue a GR.

In February 2026, Lucas Nussbaum opened another discussion on the topic by proposing that Debian allow AI-assisted contributions. He said that people who admit to using AI were being attacked by "a group of people that vehemently rejects AI", and he wanted to find a middle ground for Debian to move forward. The resulting discussion was, in his words, "generally civilized and interesting", and he thought that a GR might not be needed after all.

It's time

Last month, Geiger decided that one is needed after all. He said that it is "time for Debian to make a statement regarding generative AI and LLM usage". He opted to skip ahead to officially submitting his GR proposal, rather than testing the waters as others had done, because the topic had already been discussed in detail. Once the dam broke, a flood of proposals followed. In total, there are eight proposals for Debian's developers to consider, summarized below, plus "none of the above".

  1. No LLM contributions to Debian via Social Contract. Geiger's option: it would amend Debian's Social Contract to expressly forbid any contributions to Debian created with the use of LLMs or other generative-AI tools. It does not apply to upstream projects that use LLMs or AI-related software (thus, packaging an application for AI-related development would be unaffected).
  2. Allow AI-assisted contributions with conditions. Nussbaum's option: it expressly allows AI-assisted contributions as long as the contributor follows a set of conditions, including assuming full responsibility for technical merit, security, license compliance, etc. It also requests, but does not require, disclosure if "a significant portion of a contribution is generated or substantially assisted by a tool". The form of disclosure is left to the contributor.
  3. Reject LLMs as far as practical, update code of conduct. Proposed by Ian Jackson, this option would amend Debian's code of conduct to forbid the use of LLMs to create "messages to humans" such as bug reports, emails, discussions on Salsa, and even blog posts aggregated on Planet Debian. It requests that contributors "avoid the use of LLMs in their Debian work", but concedes that a complete ban on LLM output is impractical.
  4. Accept AI contributions for Debian-specific work. This option, proposed by Pierre-Elliott Bécue, expressly allows contributions created with or by generative-AI tools, with a set of conditions placing responsibility on the submitter.
  5. Responsible use of generative AI. Under Marc Haber's option, Debian would issue a statement describing its position as neither endorsing or prohibiting the use of generative-AI tools in Debian, and says that use of such tools "is neither exempt from nor subject to special rules beyond the standards already expected of Debian contributors".
  6. A cautious approach to generative AI. Proposed by Tobias Frost: if adopted, the project would issue a statement describing its position on generative AI as encouraging contributors to "avoid the use of generative AI where practical and to prefer human authorship, collaboration, and technical understanding over AI-generated output". However, it does not forbid the use of generative AI, and affirms that the project trusts its developers to uphold Debian's "high quality values".
  7. Debian is created by humans. This proposal, put forward by Gard Spreemann, was inspired by the GCC project's AI policy and the rust-lang/rust policy from the Rust language team. It disallows output of generative-AI tools as direct contributions, but allows contributors to use such tools "as an [assistive] tool to explore, research, analyze, critique, etc."
  8. Avoid the use of LLM: climate destruction is a deal breaker. Holger Levsen's option would have Debian issue a position statement against LLMs on environmental grounds as well as "significant ethical, legal, technical, and social concerns". However, it acknowledges LLM usage can be "hard if not impossible to detect". Therefore, it is a position statement only: there would be no prohibition of LLM usage for Debian contributions.

Debian uses the Condorcet method for voting to find the majority-preferred option, so voters will rank their choices in order of preference rather than needing to choose only one. Because Geiger's option would modify the Social Contract it requires a three-to-one majority to pass.

Discussion

It is difficult these days, if not impossible, to participate in open-source projects without being immersed in the debates about LLM usage. The positions and arguments deployed in Debian's discussion will likely be familiar to most readers. For example, regarding code copyright, Geiger argued that there is "huge uncertainty" about the copyright status of LLM outputs. He failed to see how LLM-generated code could comply with the Debian Free Software Guidelines (DFSG): "If you let a function be generated where you can't be 100% if you don't have any license obligations due to the training data".

Others were not persuaded that LLMs posed a risk to Debian or violated the DFSG. Russ Allbery pointed out that many of Debian's upstream projects, including the kernel, "have decided that none of these concerns matter and they're going to ignore all licensing issues around LLM output". The ship had sailed, he said. Ted Ts'o said that he was "perfectly happy to certify that I believe there is no copyright issue". He argued there was no more copyright uncertainty around LLM output than humans who might write code after searching for examples on Stack Overflow.

Simon Richter said that he expected the US legal system would decide that LLM outputs were "unrelated to training data and not copyrightable". However, he reached that conclusion not based on its legal merits, but because "the expensive lawyers" would be working against a finding that LLM output was not distributable due to copyright problems.

The topic of quality, or lack thereof, was also explored at length. Bécue's first draft of a ballot option, which became "Accept AI contributions for Debian-specific work", originally included specific prohibitions against using an AI assistant to push commits or upload packages. Nussbaum complained that this added arbitrary restrictions against personal workflows. He asked what the problem was with asking a local agent to commit and push to a branch.

In the ensuing back-and-forth, Bécue said that he used LLMs daily, but believed that "the more automated and fast-paced a tool is, the more safety nets are required". He was not arguing that humans don't do low-quality work, but he wanted to prevent automated systems pushing a higher volume of it. Haber responded that people should be able to use LLMs to do low-quality work if they were going to do so anyway:

But why should we not let us be assisted in doing crap if we're going to do crap anyway? Bad people are going to do AI assisted crap, good people are going to produce less crap if they are allowed to use assistance of their choice.

Bécue said that would lead to producing "10 times the crap 10 times faster, which has a significant impact and produces more strain" on the project. In another message, he said that most of the people he knew using generative AI "read less than half the code it writes" and warned that they were already having difficulty in trying to write code on their own.

Matthias Urlichs, however, did not share Bécue's concerns about quality. He said that his own packaging had improved a lot since he started letting an LLM do it, and he doubted that Debian would see 100 times more crap:

Conversely, I can immediately think of a couple of use cases where LLMs would add some tangible value. Like, chasing down reproducible-build failures. Or filter [Intent to Package bugs] for existing software that already does what the new package wants to add. Or help scan packages in the NEW queue. Or help keep our clones of Rust, Go, Python and/or Node packages up to date – which frankly is one of the more tedious and un-rewarding jobs in Debian.

Compilers and calculators

Ts'o commented on the concerns about people who depend on LLMs forgetting how to code. He acknowledged that calculators led to a decline in the number of people who could do math in their heads, and that compilers had definitely led to fewer people who were good at writing assembly language. "But I think it's more about people being less proficient at certain skills that are no longer as important, and less about some more general claim about 'cognitive decline'."

Adrian Bunk brought up the debhelper tool for building Debian packages as an example closer to home for Debian. He said that there was now a clear lack of knowledge among Debian developers about the internals of building Debian packages, which meant that they were "often lost how to debug and fix issues in their packages". Bunk and Ts'o submitted these examples of skill loss as arguments for, not against, the use of LLMs.

However, Bécue remained unconvinced. He replied to Ts'o that his calculator comparison "has some merit, but is not a good one anyway". He worried that people would turn to LLMs without having first developed their ability to think critically about the problems, which left them "both vulnerable to a rug pull, but also to be fed garbage without ever realizing it".

Environmental impact

The environmental impact of LLMs was a dominant theme in the discussion. Bas Wijnen said he could not understand how others could agree that LLMs "destroy our ecosystem" but conclude that it is not a deal-breaker when considering their use. The only response he had seen on that point was that "other activities are also bad (some even worse) for the climate", which he did not see as a reasonable argument.

All of our users live on this planet, so protecting that planet is a matter of life and death for them. In other words, caring about something as big as climate change during our Debian work is not (as was suggested elsewhere in the discussion) a violation of the social contract. On the contrary, I would argue that the social contract requires us to care about climate change and prevent it as much as we can.

Nussbaum replied that he was not dismissing the topic of climate change, but claimed that singling out LLMs "sounds like AI-bashing, not like a well-reasoned discussion about the impact on climate change". He said that Debian could reduce its environmental impact in other ways, such as building fewer packages, stopping large-scale QA efforts that require many rebuilds, and suggested ways Debian could reduce its "travel-induced CO2" by changing how DebConf is organized. "Why would Debian encourage me to travel to India, South Korea, Argentina or Japan for Debconf, but at the same time, decide that it is unacceptable to use AI assistance"?

In a postscript to a message about free-as-in beer LLMs, Ts'o suggested that the concerns about environmental impact were merely "an excuse to justify a position that someone has already has made for other reasons". Didier 'OdyX' Raboud responded that Ts'o's type of argument was "either ill-advised, or in bad faith". His concern was not with ecological impact on a per-query basis, but with the effect of the LLM industry overall:

AI companies are not a mandatory feature of our societies. We don't have to allow AI companies to destroy books to train LLMs [0]. We don't have to allow building datacenters to put strain on local electricity or water networks, or increase the local temperature. In a summer with massive drought and fires burning through Europe and North America, with (public) funding desperately missing in firefighting, climate action, etc, deciding to invest 30 B€ in AI Gigafactories [1] _is_ a choice that we "as a species" are making, and one that I don't think is right at all.

It is unclear whether the dozens of messages exchanged between Debian developers have used more, or less, energy than asking an LLM for help building a package; it is, however, clear that few, if any, minds were changed during the discussion.

Vote

Debian has dealt with divisive decisions before, of course. For instance, the project struggled with the question of choosing a default init system for some time before it was decided by the Debian Technical Committee in February 2014; The topic reared its head again as a GR in October 2014, the outcome of which was that a GR was not required.

The init debate—in the form of "init-system diversity" was rekindled in 2018, and voted on as a GR at the end of 2019. Ultimately, Debian chose systemd with support for exploring other alternatives. While the init debates brought out strong feelings, and words, it does not seem to have dealt great harm to Debian in the long run.

Considering the number of options, and the fact that several are similar in nature, it is difficult to guess which of the proposals is likely to win. If history is any guide, though, the likely outcome is that the project will dust itself off and get back to work after the vote—though the topic may well bubble up again before all is said and done.

The voting period began on August 15, and will run through August 28. While there were quite a few people participating in the discussion, there were far fewer developers who weighed in than likely voters; according to the tally sheet for the vote, more than 130 developers have cast their ballot already out of more than 1,000 eligible voters.

Whatever option wins, one hopes that the project is better for it in the long run.


to post comments

Ethical considerations, but not the right ones

Posted Aug 19, 2026 18:15 UTC (Wed) by josh (subscriber, #17465) [Link] (126 responses)

I appreciate that many in Debian are evaluating this on ethical grounds. But I'm always sad to see how dominant the environmental message is, when that just seems like an irrelevant distraction compared to everything else. Many, many things consume energy, and we're improving in how we generate energy, and I don't think LLM usage is particularly unusual in that regard.

Opposition to LLMs acquaints us with strange bedfellows. And while I'd normally appreciate the support for the no-LLM position regardless of rationale, the problem is that the rationale doesn't hold up when the circumstances change. If every use of LLMs was powered by solar/wind/etc, LLMs would still be a problem on ethical grounds that have nothing to do with the environment.

I put much more stock in "LLMs ripped off all of our work in violation of its license, and then use that to compete with us and replace us". As well as "we haven't solved AI alignment, and need to stop training new frontier models until we have, or everyone dies".

This is quite a field of options, and I hope we end up on one of the ones that rejects AI-generated code. Personally, I'd favor 7. Ending up with one of those that just makes a statement, rather than actually taking action, would be disappointing.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 18:27 UTC (Wed) by bluca (subscriber, #118303) [Link] (56 responses)

> LLMs ripped off all of our work in violation of its license

1) That did not happen, licenses only have effect under copyright law, and training is done in derogation to copyright law, so no license violation can take happen in the first place
2) What exactly happened to "information yearns to be free"? Did you always side with the MPA and RIAA and the Disney Corporation on copyright maximalism, or did you just recently come to agree with them?

> and then use that to compete with us and replace us

Welcome to capitalism. Instead of ranting about evil LLMs taking your job, I'd recommend joining a trade union and unionizing the workplace. Historically much, much more effective for protecting workers rights than complaining on social media. Not as great for dopamine hits as likes and thumbs ups, though.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 18:58 UTC (Wed) by josh (subscriber, #17465) [Link] (41 responses)

That did not happen

You show up on every article about LLMs to make this point, repeatedly. It seems like you are letting law determine ethics, rather than the other way around.

I am objecting, on ethical grounds, to LLMs copying Open Source work and not respecting its license. I do not care that some jurisdictions have made rulings deciding in violation of all sense that that mass ripoff was "fair use" or equivalently allowed, it was still wrong and should never have been allowed.

To use the US's four-pillar fair-use test as an example (because I think it's instructive as to why this makes no sense, not desiring to derail into a tangent on different jurisdictions): the purpose and character of the use is transformative but fully commercial, the nature of the copyrighted work is "everything ever written by humans", the amount and substantiality of the portion taken was "all of it", and the effect of use on the potential market is massively destructive.

What exactly happened to "information yearns to be free"?

Asymmetry happened. I think copyright should not exist. In a world in which copyright did not exist, for anyone, I would not have copying-related objections to AI (though I'd still have a half-dozen other objections), and on net I would celebrate. As long as copyright exists, I do not want a world in which AI training can rip off everyone's work (from Open Source software to documentation to literature to random comments like this) and ignore licenses, while it's still a license infringement to copy macOS or to virtualize macOS on non-Apple hardware.

Instead of ranting about evil LLMs taking your job

I'm not "ranting", and I'm not complaining about competition. I'm complaining about using my own work and the work of everyone I know to compete with us.

unionizing

Often a good idea for a whole host of other reasons, but not going to stop people from using LLMs and spewing slop everywhere; in practice, there are too many developers who have fully drunk the AI koolaid.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 19:21 UTC (Wed) by mb (subscriber, #50428) [Link] (25 responses)

I'm complaining about using my own work and the work of everyone I know to compete with us.

I don't understand why you publish your work as Open Source and then expect other people not to take it and compete with you. That doesn't really make any sense. Before LLMs it was already possible to learn from your code and then compete with you. No compliance to copyright licenses was/is required, because nothing is copied. And even under copyright it is possible with most Open Source licenses to take (copy) your work and compete with you.

I am objecting, on ethical grounds, to LLMs copying Open Source work and not respecting its license.

Most licenses have no ethical rules and definitions. How would I know what your ethical grounds are, if you don't write it down in the legal document?

Of course you are free to restrict your software with additional ethical restrictions. But then it's essentially proprietary software at the end.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 19:44 UTC (Wed) by corbet (editor, #1) [Link] (20 responses)

This line of argument, IMO, is attacking a bit of a strawman.

In general, in our community, we have agreed that free-software licenses are important. We have put considerable effort into those licenses as a result.

There is a significant subset of our community who feel that the current use of free software for model training is a violation of the free-software license they have chosen for their work. You may not agree with that feeling. As witnessed by the flood of court cases around model training, the actual legal situation is unclear, despite the claims that some might make.

Somebody who feels that their license has been violated for the production of proprietary software, and who thinks that is ethically wrong, has not become a copyright maximalist. Their position may or may not be legally correct, but it is not inconsistent.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 20:46 UTC (Wed) by mb (subscriber, #50428) [Link] (18 responses)

>As witnessed by the flood of court cases around model training, the actual legal situation is unclear

This is what I hear all the time, but this doesn't make much sense to me.
The pure number of cases doesn't say much more than that there are companies with lots of money and confidence that they are right.
There are lots of cases about shoplifting. But that doesn't make the situation around shoplifting unclear.

What matters instead is: What are the results/rulings from these cases? How many of these cases went pro/con AI training and for what reasons. If they are still going: What direction are they going?
Is there some kind of unbiased summary of both sides (pro and con rulings) available?

>a significant subset of our community who feel

Sure. I was on that side, too, some time ago.
It felt "obvious" to me that AI training was "stealing".
I do understand that this is one possible feeling one might have.

But I stopped feeling the topic and started reasoning about the topic which resulted in me changing my mind.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 20:52 UTC (Wed) by bluca (subscriber, #118303) [Link]

In the US where there's no specific law and it hinges on fair use doctrine, the highest profile case I think was Anthropic's from last year:

https://www.theguardian.com/technology/2025/jun/25/anthro...

In the EU the law is pretty clear-cut so I don't know if anyone is even attempting a court case. The 2019 Copyright Directive has a pretty clear-cut carveout for data mining training data from public data, and the AI Act directly and explicitly references it, so it would be hard to even imagine what case could be brought. It might happen at some point that someone with enough money to spend tries though.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 0:01 UTC (Thu) by NYKevin (subscriber, #129325) [Link] (16 responses)

Broadly speaking, I would summarize the situation like this in the US (and someone else can speak for Europe):

Firstly, there are three different legal questions we need to consider:

1. Does the training process infringe the training inputs by illegally copying them?
2. Is AI-generated content a derivative work of some or all training inputs?
3. Is the AI model a derivative work of some or all training inputs?

Two separate district courts have held, in response to (1), that training is fair use. But I would call one of those cases more of a draw than a victory for the AI companies. The judge's tone was extremely skeptical, and he seemed to hint that he would have gone the other way if the plaintiffs had pleaded specific facts about the market effects of AI. The other case was much more favorable to the AI companies, and its opinion was worded as if this was not at all a close question. If I had to guess, I would say that this is the sort of thing that will continue to be litigated a lot, and possibly go before the Supreme Court at some point in the future.

For (2), cases are much rarer because they are much harder to file. You cannot litigate derivative works as a class action, because the relevant legal test is substantial similarity, a direct comparison between the plantiff's work and the defendant's. If every plaintiff is suing over a different work, then every comparison needs to be different, so all of those cases have to be litigated individually. The New York Times filed a case which is broadly similar to this theme, but technically not about any of the questions listed above (it's a purely factual matter of whether a significant number of users bypassed NYT's paywall by asking ChatGPT for copies of NYT columns, after NYT lawyers showed that this is a thing that users hypothetically could have done).

I think there was some early litigation around (3), but I don't recall it going anywhere. This is, again, a derivative work theory, so it has the same problems as (2), but even worse. Nobody knows how to do a substantial similarity comparison involving an AI model. You could perhaps argue that the model must "know" things that appear in its outputs, and do the comparison on that basis. But here your options are limited by 17 USC 102(b) - you can't sue the model creator because the model has a general understanding of factual information (or in the case of image models, shading, color theory, etc.), but only if it contains some specific copyrightable element, such as a longer passage of verbatim text, or an image of Mickey Mouse. Regurgitation studies might be relevant here, depending on the model.

It's also important to understand the potential consequences of each of those questions. If the answer to (1) is "yes," then training would become illegal in a given jurisdiction, and you might get significant monetary damages, but that's (probably) it. They could still train in some friendlier jurisdiction and import the model. If the answer to (2) is "yes" with respect to a given output, you can probably get damages and destruction of that output, but again there's little or no recourse against the model as such. Only (3) entitles anyone to demand the destruction of the model, but since nobody is litigating (3), it probably isn't going to happen any time soon.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 5:53 UTC (Thu) by pbonzini (subscriber, #60935) [Link] (9 responses)

I think nobody is litigating (3) because even if anyone does the defense can assert fair use and fall back onto (1). The question itself is more or less proven mathematically.

If you try to use an LLM as a token-level arithmetic compressor, you get *better* results on quirky text like poetry or the quake invsqrt, which is exactly the opposite of what you expect from a compressor. It only takes a few words for the LLM to lock onto the original text and reproduce the next words with probability ~1. There is no explanation for this other than the presence of lossless "pockets" in the overall-lossy compression process performed by training.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 6:32 UTC (Thu) by collinfunk (subscriber, #169873) [Link] (5 responses)

> I think nobody is litigating (3) because even if anyone does the defense can assert fair use and fall back onto (1). The question itself is more or less proven mathematically.

Wouldn't the original, for lack of a better term, author have a pretty strong argument against the asserted fair use if the LLM-generated program acted as a replacement for it such that it harms the original (monetarily, commercially, or to it's license)? My layperson understanding is that is one of the stronger arguments against assertions of fair use.

Not to say that I disagree with your larger point, though. I'm sure people are a bit hesitant to make these arguments. Having your case turn into commonly cited case law doesn't sound too fun. :)

Ethical considerations, but not the right ones

Posted Aug 20, 2026 10:20 UTC (Thu) by pbonzini (subscriber, #60935) [Link] (4 responses)

The New York Times's lawsuit against Microsoft and OpenAI is probably the most visible example of (3), i.e. alleging that the model is a derivative work of anything that it is able to reproduce verbatim.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 15:25 UTC (Thu) by NYKevin (subscriber, #129325) [Link] (3 responses)

As I said, the NYT has not made any allegations regarding the model. They have accused OpenAI of distributing copies of NYT columns to end users without a license. This has nothing to do with the model and is purely a question of whether the NYT's columns appear in the chat logs.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 18:09 UTC (Thu) by pbonzini (subscriber, #60935) [Link] (2 responses)

See https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_De..., page 32: "101. Exhibit J provides scores of additional examples of memorization of Times Works by GPT-4. Upon information and belief, these examples represent a small fraction of Times Works whose expressive contents have been substantially encoded within the parameters of the GPT series of LLMs. Each of those LLMs thus embodies many unauthorized copies or derivatives of Times Works."

Ethical considerations, but not the right ones

Posted Aug 20, 2026 18:56 UTC (Thu) by NYKevin (subscriber, #129325) [Link] (1 responses)

The section you quote is factual allegations. The claims begin on page 60, and are much more specific about the laws they invoke.

I will admit, I was mistaken, the NYT is in fact also alleging violations of types (1) and (3). Paragraph 162 is a claim of type (3):

> On information and belief, by storing, processing, and reproducing the GPT models trained on Times Works, which GPT models themselves have memorized, on Microsoft’s supercomputing platform, Microsoft and the OpenAI Defendants have jointly directly infringed The Times’s exclusive rights in its copyrighted works.

However, I will point out that "on information and belief" is a rather significant hedge. It basically means "we don't have firm evidence of this yet, but we're hoping something comes up in discovery."

Ethical considerations, but not the right ones

Posted Aug 21, 2026 5:56 UTC (Fri) by pbonzini (subscriber, #60935) [Link]

Yes, I agree it's a bit of a "throw it at the wall and see if it sticks".

Ethical considerations, but not the right ones

Posted Aug 20, 2026 14:20 UTC (Thu) by kleptog (subscriber, #1183) [Link] (2 responses)

> If you try to use an LLM as a token-level arithmetic compressor, you get *better* results on quirky text like poetry or the quake invsqrt, which is exactly the opposite of what you expect from a compressor.

That's not surprising at all. It's even expected. There is an argument to be made that compression is related to intelligence. If you compress the string "e^{i\pi}+1=0" you would expect it to be smaller than if you tried "e^{i\pi}-1=0". The text "1234+8765=9999" would compress better than "1234+8765=8888" because the former is consistent with the model and the latter isn't. Grammatically correct text should compress better than text with grammatical and spelling errors. True statements about the world should compress better than false ones.

Poetry is interesting, because once you constrain the rhyme and rhythm, the word choices contract enormously. So there is a good chance it would compress extremely well.

> It only takes a few words for the LLM to lock onto the original text and reproduce the next words with probability ~1. There is no explanation for this other than the presence of lossless "pockets" in the overall-lossy compression process performed by training.

Or the text wasn't very imaginative. Grammatical English only has an information density of 1-2bits per character. It's not uncommon for people to be able to quote large sections of favorite movies, books or plays and we don't consider that copyright infringement. Unless it's a performance for money. It could point to overfitting of the model and it may indicate you could reduce the number of parameters while retaining fidelity. But the fact a model can reproduce some text is not in and of itself an issue. What matters is if it a person picks it up and passes it off as theirs to earn money from it.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 15:01 UTC (Thu) by pbonzini (subscriber, #60935) [Link] (1 responses)

> So there is a good chance [poetry] would compress extremely well. [...] Or the text wasn't very imaginative.

Just try asking an LLM to play a "what's the next word" game. You'll get something like

User: I would like to play a game, "Tell me the next word". Ok?

LLM: I'm in! Give me the first word or phrase whenever you're ready.

User: Shall I

LLM: compare

User: go on

LLM: thee

Literally, giving a verb and a pronoun was enough to lock the LLM onto a famous poem that starts with that expression. No metric or rhyme constraints yet and the LLM was not even told the game was about poetry.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 11:58 UTC (Fri) by kleptog (subscriber, #1183) [Link]

That sounded like a good experiment so I gave it a shot:

User: I would like to play a game, "Tell me the next word". Ok?

LLM: Yes. I'll give a word, you reply with the next word that comes to mind. Then I'll continue. Cloud → ?

User: I want to start with "Shall i"

LLM: go

User: first

LLM: or

etc...

It really depends on which model. And some randomness of course. Although, after several tries the only next word I got was "go". Even if I phrase it in different ways.

(As an aside, I still find it magic to have a computer be able to understand commands like this.)

Ethical considerations, but not the right ones

Posted Aug 21, 2026 4:00 UTC (Fri) by marcH (subscriber, #57642) [Link] (5 responses)

> 2. Is AI-generated content a derivative work of some or all training inputs?

> For (2), cases are much rarer because they are much harder to file. You cannot litigate derivative works as a class action, because the relevant legal test is substantial similarity, a direct comparison between the plantiff's work and the defendant's. If every plaintiff is suing over a different work, then every comparison needs to be different, so all of those cases have to be litigated individually.

Yes, obviously.

Is a song "inspired" by another song violating the copyright of the former? Same question for a book, painting and everything else - including computer code. The answer is obviously: IT DEPENDS. Pretty much everyone understands that there will never be a single answer there. There is zero debate: it always depends.

So why is anyone entertaining the idea that AI creations could have a more systematic copyright answer? Based on what logic?

Ethical considerations, but not the right ones

Posted Aug 21, 2026 11:35 UTC (Fri) by alx.manpages (subscriber, #145117) [Link] (4 responses)

> So why is anyone entertaining the idea that AI creations could have a more systematic copyright answer? Based on what logic?

Because LLMs don't have creativity. You may agree or disagree with that claim, but at least it's a legitimate claim.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 12:50 UTC (Fri) by marcH (subscriber, #57642) [Link] (1 responses)

> Because LLMs don't have creativity.

This is debatable because "creativity" is a fuzzy concept in the first place - long before LLMs. Exactly why copyright is case-by-case. But the creativity concept is not even needed here:

- Some generative AI created unique works that do not have "substantial similarity" with anything else.
- Some generative AI created other works that _do_ have "substantial similarity" with something else.

So, the question "Are LLM products derived works?" is case by case. End of story.

Antisocial networks always demand black or white answers now and hate slow and complex justice but they can't possibly win this one.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 12:55 UTC (Fri) by marcH (subscriber, #57642) [Link]

Forgot this sorry: to make this question even more complex and less predictable, "creations" are usually not just the product of AI but rather a collaboration between AIs and human(s). Even more grey.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 18:21 UTC (Fri) by rgmoore (✭ supporter ✭, #75) [Link] (1 responses)

Because LLMs don't have creativity. You may agree or disagree with that claim, but at least it's a legitimate claim.

In the USA, that isn't just a theory. By law, a work must have a human creator to be considered to have enough creativity to qualify for copyright protection. Works that are purely a work of happenstance, are created by non-human animals, or are created solely by machines lack sufficient creativity to qualify for copyright. Of course that doesn't have anything to say about whether those works violate someone else's copyright. An LLM can violate copyright by returning a copyrighted work in response to a prompt even though it's output wouldn't be eligible for copyright.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 19:45 UTC (Fri) by NYKevin (subscriber, #129325) [Link]

> Of course that doesn't have anything to say about whether those works violate someone else's copyright. An LLM can violate copyright by returning a copyrighted work in response to a prompt even though it's output wouldn't be eligible for copyright.

Yes, but this goes both ways. The fact that an LLM's output might be ineligible for copyright doesn't automatically create an infringement either. If the output is not substantially similar to any particular input, then there can be no infringement (in the case of that one output, anyway).

Ethical considerations, but not the right ones

Posted Aug 20, 2026 1:24 UTC (Thu) by SLi (subscriber, #53131) [Link]

Hmm. I think the major problem here is that the argument almost always sounds like it's a legal argument. I think the ethical side is debatable by laypeople. What is problematic is conflating the two and claiming it's illegal, which, at least to me, almost all of these arguments seem to do.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 19:48 UTC (Wed) by josh (subscriber, #17465) [Link] (3 responses)

> I don't understand why you publish your work as Open Source and then expect other people not to take it and compete with you.

I expect them to, but I expect them to have to respect the license in doing so.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 6:40 UTC (Thu) by burki99 (subscriber, #17149) [Link] (2 responses)

Just to clarify: If there was a model only trained exclusively on GPL-2 code and the output would carry a GPL-2 copyright notice, would you object? (Possibly with a link to a combined very long list of all authors included in the training data to give proper individual credit, even if it is very hard to assess the individual share.)

Ethical considerations, but not the right ones

Posted Aug 20, 2026 12:44 UTC (Thu) by dskoll (subscriber, #1630) [Link]

I would not object to such an LLM on copyright-violation grounds. (I still have many reservations for other reasons about LLMs being used to write software.)

Ethical considerations, but not the right ones

Posted Aug 25, 2026 13:59 UTC (Tue) by jond (subscriber, #37669) [Link]

That kind-of thing would alleviate one of my major problems with the technology. Although I am led to believe that the corpus size of any such licence-constrained training process would be too small to yield a useful model.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 20:02 UTC (Wed) by bluca (subscriber, #118303) [Link] (13 responses)

> You show up on every article about LLMs to make this point, repeatedly.

Yes, because this blatantly incorrect point keeps getting raised by the luddites for shock effect, so I'll keep correcting it. You are entitled to your own opinion, not your own laws. It matters not one bit whether you like it or not, training models on publicly accessible works is legal and there's nothing wrong with it, as copyright is a (fairly recent) human construct that exists solely in legislation, and has no moral or ethical grander meaning behind it. It's good that its reaches are limited and that there are carveouts.

It's also a very good thing that it worked out this way, otherwise only mega-corporations would be able to afford paying for the corpora needed to train models. It's actually great that the law for once is on the side of everyone, instead of Disney and other mega-corps, and that absolutely anybody has the same rights to train their own models on their own hardware using the same corpora.

> I think copyright should not exist.

And yet here you are, doing the work of Disney for them, arguing for the losing side of copyright maximalism. So assuming you really do think copyright should not exist, you are not doing a very good job at it I'm afraid.

> I'm not "ranting", and I'm not complaining about competition. I'm complaining about using my own work and the work of everyone I know to compete with us.

Nah, this is pretty much the definition of ranting. You are missing the woods for the trees, and are angry at the effects of capitalism without actually realizing it, and pointing at the finger instead of the moon. Capitalism will capitalism, as it has always capitalism'ed. It doesn't matter at all what tools it uses, as the tools are not the problem, they never were since the 1800s, they are not today, they will not be tomorrow.

Join your local green party, join your local trade union, and that will do something useful and tangible to stop injustice perpetrated by out-of-control capitalism. Keep ranting about muh evil LLMs, and the capitalists will rejoice, as you will be wasting your time tilting at windmills, and going against fellow workers that you should ally with, instead of antagonising.

It was inevitable

Posted Aug 19, 2026 20:13 UTC (Wed) by corbet (editor, #1) [Link] (5 responses)

...that, Luca, we would have to tell you to stop this.

You can argue your point without calling people names, creating strawmen, etc. Do not make us ask you again.

It was inevitable

Posted Aug 19, 2026 20:21 UTC (Wed) by bluca (subscriber, #118303) [Link] (4 responses)

Nobody has been "called names", I have no idea what you are talking about

It was inevitable

Posted Aug 19, 2026 20:22 UTC (Wed) by corbet (editor, #1) [Link] (3 responses)

"Luddites" is not calling names? Be serious.

It was inevitable

Posted Aug 19, 2026 20:42 UTC (Wed) by bluca (subscriber, #118303) [Link] (2 responses)

No, I'm pretty sure it's not actually. In the colloquial meaning it represents a group of people who lash out against technological changes due to perceived negative effects, despite those effects not being due to those changes, but instead brought by the capitalist class exploitations. You may not like it to hear it, but it doesn't mean it's "calling names". I don't like being told that I "spew slop everywhere" and that I "drank the AI koolaid", but that's not calling me names either.

It was inevitable

Posted Aug 21, 2026 4:35 UTC (Fri) by marcH (subscriber, #57642) [Link]

If you can't admit that calling others "luddites" is calling them names, then you are really not interested in being taken seriously. Too bad, there is also some interesting stuff in what you write; it's sad most people will miss it.

> I don't like being told that I "spew slop everywhere" and that I "drank the AI koolaid", but that's not calling me names either.

There is a grammatical difference (these are not nouns/"names") but I agree the effect is pretty much the same.

Two wrongs

Posted Aug 21, 2026 23:00 UTC (Fri) by warrax (subscriber, #103205) [Link]

Do you think that others calling you names is a justification for calling other unrelated people[0] names? This isn't difficult.

[0] ... who have *not* called you names.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 20:59 UTC (Wed) by collinfunk (subscriber, #169873) [Link] (1 responses)

> Join your local green party

I would caution this statement. The Green Party in the US is an insane joke of an organization who beleives that Russia should be allowed to continue its invasion of Ukraine without being contested. My understanding is that European green parties are less unhinged [1], but I'd assume not all are sane to the same degree.

[1] https://europeangreens.eu/news/us-elections-european-gree...

Ethical considerations, but not the right ones

Posted Aug 19, 2026 21:05 UTC (Wed) by bluca (subscriber, #118303) [Link]

Right, that's sad to hear, sorry I was really not familiar with such details with the US situation. In Europe, generally speaking, local or national parties affiliated with the Greens/EFA EU parliament group are sound

Ethical considerations, but not the right ones

Posted Aug 19, 2026 21:39 UTC (Wed) by josh (subscriber, #17465) [Link] (4 responses)

> training models on publicly accessible works is legal and there's nothing wrong with it

Those are two very different statements. It is, currently, legal in some jurisdictions and contested/unclear in others. None of those legal situations determines whether it's wrong; that is a question purely in the realm of ethics, not legality, and we are not going to resolve our fundamental disagreement on the ethics of LLMs by pointing to what's legal where.

>> I think copyright should not exist.

> And yet here you are, doing the work of Disney for them

Nope, you are misrepresenting my position. I am arguing that as long as copyright exists, it must be applied symmetrically, not favoring large companies over others. By all means, destroy copyright, but do not create carveouts that *disproportionately benefit a handful of large companies rather than individuals*.

> It doesn't matter at all what tools it uses, as the tools are not the problem

That is a position. It is not and will never be my position. LLMs are "not just a tool" any more than nuclear weaponry or leaded gasoline is "just a tool"; technology is not morally neutral.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 21:47 UTC (Wed) by josh (subscriber, #17465) [Link]

Also:

> You are missing the woods for the trees, and are angry at the effects of capitalism without actually realizing it

Do not presume to tell me what my own concerns are, or project your positions onto me. You are factually incorrect, and your perceived root cause is not my root cause.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 23:57 UTC (Wed) by mjg59 (subscriber, #23239) [Link] (1 responses)

> None of those legal situations determines whether it's wrong; that is a question purely in the realm of ethics, not legality

I agree, but when the ethical software movement suggested writing licenses taking that into account they received a lot of pushback. I'm not sure free software has an entirely consistent view on this.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 6:19 UTC (Thu) by josh (subscriber, #17465) [Link]

I think it's very consistent. If we try to address a threat to Open Source software by using licenses that aren't Open Source anymore, that feels like clearcutting a forest to protect it from fire.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 11:28 UTC (Thu) by bluca (subscriber, #118303) [Link]

> Those are two very different statements. It is, currently, legal in some jurisdictions and contested/unclear in others.

Exactly, it is legal, and yet you said the opposite, which is what I was responding to:

> LLMs ripped off all of our work in violation of its license

A violation of a license is a legal issue. You cannot demonstrate that any such violations has occurred, as so far every legal case, and every applicable law, says it has not. And hence I corrected that point, and will continue to do so as long as it keeps being erroneously made. You are entitled to hold the view that it is immoral to train LLMs on open source software, you are not entitled to the view that it is a license violation to do so. The former is an opinion, the latter is a legal matter.

> Nope, you are misrepresenting my position. I am arguing that as long as copyright exists, it must be applied symmetrically, not favoring large companies over others. By all means, destroy copyright, but do not create carveouts that *disproportionately benefit a handful of large companies rather than individuals*.

No, you are arguing for access to information to be made asymmetrical and thus favor mega-corporations. If an explicit license was required to train models as you wish it to be the case, only the richest entities in the world could afford to do so and train large models. The fact that the law (or absence thereof in some jurisdiction) allows _anybody_ to train LLMs on any publicly available corpora in derogation to copyright, makes it symmetric and favors everyone, especially the small players. In your ideal world, only mega-corporations would be able to create LLMs. No local models, no open weight models, none of that would exist anymore. For *once* the law works in favor of everyone, and against the mega-corps, and I am extremely happy about it and wish it would go even further.

> technology is not morally neutral.

Means of production are means of production. They can be used by the capitalist class to oppress and exploit workers, or they can be used by workers to provide for their communities. Capitalism did not start to exploit workers when ChatGPT was released, and it would not stop exploiting workers if your wish came true and a genie snapped their fingers and disappeared all LLMs from the safe of the earth. Destroying mechanical looms did not stop capitalism's exploitations, and destroying LLMs won't either. Only unionizing, severe regulations and taxation have historically worked to rein in the excesses of capitalism, and I don't see any reason why that should change now.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 1:21 UTC (Thu) by SLi (subscriber, #53131) [Link]

I think the issue is that "violation of license" inherently means breaking the law, not ethics. Licenses are legal documents. If you mean you'd prefer it was illegal to use your code that way without taking a position on whether it is, you could also say so.

(FWIW, I also fall on the side that considers more copyright restrictions harmful.)

Ethical considerations, but not the right ones

Posted Aug 20, 2026 6:49 UTC (Thu) by gspr (subscriber, #91542) [Link] (9 responses)

> 2) What exactly happened to "information yearns to be free"? Did you always side with the MPA and RIAA and the Disney Corporation on copyright maximalism, or did you just recently come to agree with them?

To turn this question back on you: What happens if LLMs reach a stage where they can ingest an MPAA member's blockbuster movie, or a RIAA member's bestselling song, and output its own "interpretation/rewrite", detached from the original's copyright? Will the little man/woman be granted the same fair use leeway then?

This is the asymmetry that worries a lot of us. I, too, grew up with a mindset of "information yearns to be free". But I've come to the same realization as the FSF, namely that copyright law can in fact be a mechanism to protect that freedom. In the potentially asymmetric future where e.g. the GPL can be laundered off by an LLM, but "freeing" proprietary software or other IP by passing them through an LLM will be forbidden, you must surely agree we are *a lot* worse off?

Ethical considerations, but not the right ones

Posted Aug 20, 2026 9:10 UTC (Thu) by joib (subscriber, #8541) [Link] (1 responses)

> In the potentially asymmetric future where e.g. the GPL can be laundered off by an LLM, but "freeing" proprietary software or other IP by passing them through an LLM will be forbidden, you must surely agree we are *a lot* worse off?

OTOH, given a binary artifact and any documentation about it, an LLM can probably relatively quickly produce a reverse engineered piece of software that works the same. So one could argue LLM's help bridge the "proprietary software moat". Interesting times..

Ethical considerations, but not the right ones

Posted Aug 24, 2026 6:49 UTC (Mon) by gspr (subscriber, #91542) [Link]

Absolutely. It's not all bad!

Ethical considerations, but not the right ones

Posted Aug 20, 2026 11:08 UTC (Thu) by bluca (subscriber, #118303) [Link] (6 responses)

> To turn this question back on you: What happens if LLMs reach a stage where they can ingest an MPAA member's blockbuster movie, or a RIAA member's bestselling song, and output its own "interpretation/rewrite", detached from the original's copyright? Will the little man/woman be granted the same fair use leeway then?

You can already do that. There are endless scores of clones of the same stories, tropes, etc. You cannot write a new "Game of Thrones" novel starring "Ned Stark and Jon Snow", but you can certainly write yet-another "gritty low-fantasy political-murderous-gory novel". Should it go further? Yes, copyright should be abolished. Any step that weakens its hold on information is a good step in the right direction.

> This is the asymmetry that worries a lot of us. I, too, grew up with a mindset of "information yearns to be free". But I've come to the same realization as the FSF, namely that copyright law can in fact be a mechanism to protect that freedom. In the potentially asymmetric future where e.g. the GPL can be laundered off by an LLM, but "freeing" proprietary software or other IP by passing them through an LLM will be forbidden, you must surely agree we are *a lot* worse off?

I don't see why that would be the case, in fact you can already reverse-engineer any proprietary program and open source the result. You cannot release a new "Microsoft Office 365", but can totally develop and open source an office suite that is bit-by-bit compatible with Microsoft Office 365 file formats. I don't know how good LLMs are at reading machine code and spitting out sources, as I'm not involved in that side of computing, but I do hope they get really good at it soon.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 12:27 UTC (Thu) by anselm (subscriber, #2796) [Link] (1 responses)

in fact you can already reverse-engineer any proprietary program and open source the result. […] I don't know how good LLMs are at reading machine code and spitting out sources, as I'm not involved in that side of computing, but I do hope they get really good at it soon.

I don't think it is actually the case that reverse-compiling a proprietary binary and open-sourcing the result is obviously legal. Certainly under an EU copyright regime, analysing a proprietary binary is only allowed under certain restrictions – typically, to enable interoperability of one's own code with the binary in question in the absence of sufficiently detailed documentation. This does not appear to endorse the idea of using an LLM to generate sources for, say, Office 365 or Photoshop binaries wholesale and publishing the result under an open-source license. (It is also safe to assume that Microsoft and Adobe would be less than enthusiastic.)

Ethical considerations, but not the right ones

Posted Aug 20, 2026 14:57 UTC (Thu) by bluca (subscriber, #118303) [Link]

Reproducing 1:1 down to branding? Likely not, which is what I meant with 'You cannot release a new "Microsoft Office 365"'. But it doesn't need to be, being able to reverse engineer and reimplement the needed functionality is what actually matters, and you can already do that. IIRC there are already pretty decent binary -> sources tools that reverse engineering folks use, and if LLMs can further improve that, that sounds all great to me

Ethical considerations, but not the right ones

Posted Aug 24, 2026 6:47 UTC (Mon) by gspr (subscriber, #91542) [Link] (3 responses)

> You can already do that. There are endless scores of clones of the same stories, tropes, etc. You cannot write a new "Game of Thrones" novel starring "Ned Stark and Jon Snow", but you can certainly write yet-another "gritty low-fantasy political-murderous-gory novel".

This is the pre-LLM status quo. It's akin to using some software, being inspired, and making your own version. This is entirely uncontroversial.

Imagine instead a world in which you can feed Game of Thrones to an LLM and have it produce e.g. a cartoon version or some other little tweak. How would the IP holders then react?

> Should it go further? Yes, copyright should be abolished. Any step that weakens its hold on information is a good step in the right direction.

I think it's important that you lead with that motivation. How would you feel if the LLM future we discuss comes to pass, yet copyright remains, but is in practive only enforceable by the powerful? Coming from an "information wants to be free" mindset, won't we then be strictly worse off than we are today? I think there's a real danger that that's where your path will take us. You may well disagree, but I actually think a stricter interpretation of copyright to restrict LLM-washing is the way to go. We, the free software movement, have long ago adapted to this world and we play it well. We should be careful what we wish for when we wish for revolutionary change.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 7:34 UTC (Mon) by taladar (subscriber, #68407) [Link] (2 responses)

Have you considered that your stricter interpretation vs. the powerful's looser interpretation of Copyright is already exactly the future you fear, just self-imposed?

Ethical considerations, but not the right ones

Posted Aug 24, 2026 7:52 UTC (Mon) by gspr (subscriber, #91542) [Link] (1 responses)

I'm not sure I understand you. Could you elaborate?

Ethical considerations, but not the right ones

Posted Aug 25, 2026 7:48 UTC (Tue) by taladar (subscriber, #68407) [Link]

If you interpret Copyright strictly and limit your own actions more but the powerful are interpreting it loosely and limit their own actions less you end up exactly in the situation you fear where the powerful can do a lot more than you on Copyright matters without anyone having to change a single line in Copyright law and without any powerful person even having to do so much as suing you to achieve that outcome.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 10:54 UTC (Mon) by Phantom_Hoover (subscriber, #167627) [Link] (3 responses)

> That did not happen, licenses only have effect under copyright law, and training is done in derogation to copyright law, so no license violation can take happen in the first place

Copyright law in practice is created and enforced by extremely protracted and expensive civil suits, so there are vast grey areas where nobody knows for sure what ‘derogation to copyright law’ means in practice. Google thought they were complying with copyright law when they copied the Java API definitions and that case got flip-flopped on appeal all the way up to the Supreme Court — and they avoided ruling on the central question! The question of whether LLM training is really legal has been grinding its way through the courts for years now (the NYT case being the biggest example) and will continue to do so for years to come (unless the parties involved settle, which will leave the legal question unanswered). So you’re straightforwardly wrong to be presenting the licence compliance of LLM training as a known fact.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 11:02 UTC (Mon) by bluca (subscriber, #118303) [Link] (2 responses)

> So you’re straightforwardly wrong to be presenting the licence compliance of LLM training as a known fact.

No, as that is the situation as of today in the US. There were multiple rulings that reinforced it in the past couple years. Yes, there will probably be more, and at some point it will likely end up in front of the federal supreme court, but right now that is a correct representation of the status quo to the best of my knowledge.
In Europe, which I care about much more since that's where I live and work, the Copyright Directive and the AI Act explicitly allow training on public data, so I'm not sure how it could be any more straightforward.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 14:26 UTC (Mon) by anselm (subscriber, #2796) [Link] (1 responses)

In Europe, which I care about much more since that's where I live and work, the Copyright Directive and the AI Act explicitly allow training on public data, so I'm not sure how it could be any more straightforward.

For one, training may be allowed (presumably, e.g., in order to do inference based on the training data), but being allowed to read public data to train your model doesn't automatically mean you're also allowed to let your model regurgitate that data (possibly even minus any license grants in the original) as output to queries, which is a different kettle of fish as far as copyright is concerned. Even in the EU, LLMs are not supposed to serve as copyright laundering machines.

It turns out that this is something which is explicitly disallowed by the EU's “Code of Practice for General-Purpose AI Models”, which most of the important AI outfits subscribe to. In theory they agree to take measures to ensure that models do not generate “outputs that reproduce training content protected by Union law on copyright and related rights in an infringing manner”, but in practice, many examples have shown that this doesn't seem to work all that well.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 14:53 UTC (Mon) by bluca (subscriber, #118303) [Link]

> in practice, many examples have shown that this doesn't seem to work all that well

This just doesn't happen though in any real-world case. The only time someone manages to make it happen, for very small snippets, is if they are _actively_ trying to make it happen to prove a point, and then again that had a higher chance of succeeding with earlier models than with more recent ones, where it quickly falls apart because of the guardrails that were put in. In reality, when using claude code or copilot or similar, the output is driven by and resembles the local context (ie: the local repository you are working on).

So, not a actual issue.

Ethical considerations, but not the right ones

Posted Aug 19, 2026 23:19 UTC (Wed) by rgmoore (✭ supporter ✭, #75) [Link] (63 responses)

LLMs ripped off all of our work in violation of its license

I keep seeing this, and I keep going back to GNU's First Freedom:

The freedom to study how the program works, and change it to make it do what you wish

It seems to me that feeding source code into an LLM as part of its training corpus falls squarely into the category of "study[ing] how the program works". It sucks that other people want to use your code in ways you didn't think of and don't approve of, but you agreed to it when you released it under a Free Software license.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 2:08 UTC (Thu) by dskoll (subscriber, #1630) [Link] (61 responses)

I've released a number of programs under the GPL. I would have no issue with someone training an LLM on those programs providing the output of the LLM had to be GPL licensed also.

I know there are people who claim that this doesn't hold because it's just as if a person studied the code and then independently wrote some other software. I don't buy it; I think the outputs of an LLM are in fact a derivative of their inputs. It has to be so because the outputs could not even exist without the inputs, whereas a person could write a program without studying my source code.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 9:06 UTC (Thu) by joib (subscriber, #8541) [Link] (59 responses)

> I think the outputs of an LLM are in fact a derivative of their inputs. It has to be so because the outputs could not even exist without the inputs, whereas a person could write a program without studying my source code.

Similarly, one could argue that the code I write is a derivative work of the textbooks and lectures I crawled through in university.

Well, clearly we don't think that, but it seems to me a closer approximation of an LLM than just copy-pasting code found on the interwebs.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 12:43 UTC (Thu) by dskoll (subscriber, #1630) [Link] (58 responses)

Similarly, one could argue that the code I write is a derivative work of the textbooks and lectures I crawled through in university.

One could argue that. But there are two differences: (1) textbooks and lectures were created with the specific purpose of training people to be software developers. Software products were not; they were created to perform some function on a computer.

And (2), unless your university education system was very different from mine, you were extremely unlikely to have read and internalized every single line of source from a large production piece of software. Most likely, you studied fragments of software, or learned in general terms about software design, algorithm design, and data structure design.

LLMs ingest the entire source code corpus of production-ready software. That's a very different mechanism from how humans learn.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 1:34 UTC (Fri) by intelfx (subscriber, #130118) [Link] (10 responses)

One could argue that. But there are two differences: (1) textbooks and lectures were created with the specific purpose of training people to be software developers. Software products were not; they were created to perform some function on a computer.

I'm sorry, but this is absurd. Nothing stops me from studying code of any number of Free Software programs, and later applying that knowledge as I see fit (including for the purposes of getting a job with an express intention of writing proprietary software). Free Software does not say "you are free to use this source code for any purpose other than studying it", it says "you are free to use this source code for any purpose" (as long as you are not copying it or deriving from it, which is a different argument, and not the one we are discussing here). If you do not wish to grant your users the latter set of freedoms, do not release your code as Free Software, for it is non-free.

And (2), unless your university education system was very different from mine, you were extremely unlikely to have read and internalized every single line of source from a large production piece of software. Most likely, you studied fragments of software, or learned in general terms about software design, algorithm design, and data structure design.

Perhaps it was.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 12:21 UTC (Fri) by dskoll (subscriber, #1630) [Link] (9 responses)

Nothing stops me from studying code of any number of Free Software programs, and later applying that knowledge as I see fit

Except that the license of the software, if it is GPL, does prevent you from reproducing large chunks of the program and distributing them under a license other than the GPL.

(as long as you are not copying it or deriving from it, which is a different argument, and not the one we are discussing here)

No, that's exactly the argument we're discussing here. It is my contention that any output of an LLM is necessarily derived from its training set. It cannot be any other way; without the training set, the LLM could not produce anything. Whereas, a human being who has only seen fragmentary code examples and an overview of software design can produce large volumes of production-ready code.

It would be interesting to see if an LLM that was trained on exactly the same input materials as a computer science undergraduate could produce anything useful. My intuition says no.

Perhaps it was.

Come one. Now that's just being disingenuous.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 19:00 UTC (Fri) by intelfx (subscriber, #130118) [Link] (8 responses)

Except that the license of the software, if it is GPL, does prevent you from reproducing large chunks of the program and distributing them under a license other than the GPL.

Nobody was talking about reproducing large chunks, at least in this subthread. In this subthread, the argument was that LLMs' outputs being "derivative" of the input is allegedly more similar in spirit to how a human learns, rather than to being a "derivative work" in the copyright law sense.

Under the premise of this argument, an LLM ingesting a corpus of texts is transformative enough for it to count as learning, not copying (just like a human is free to learn from a Free Software program exactly the same as he would have learned from a textbook).

If you disagree, you are free to contest this position on its substance, but not by invoking some sort of nebulous difference between textbooks and programs: it does not exist, at least not under Free Software licensing.

No, that's exactly the argument we're discussing here. It is my contention that any output of an LLM is necessarily derived from its training set. It cannot be any other way; without the training set, the LLM could not produce anything. Whereas, a human being who has only seen fragmentary code examples and an overview of software design can produce large volumes of production-ready code.

And? It does not matter. For copyright law, the only thing that matters is the substantive similarity test (or other similar tests, or their equivalents in other jurisdictions). If you oppose LLMs on ethical grounds, then do so, and do not bring licensing into the mix: Free Software licenses are explicitly indifferent to ethics (DFSG §§5, 6).

In the adjacent subthread, Luca articulated his position (with which I agree wholeheartedly) very well:

You are entitled to hold the view that it is immoral to train LLMs on open source software, you are not entitled to the view that it is a license violation to do so. The former is an opinion, the latter is a legal matter.

If you are saying that "LLMs are ripping off our work", that is a defensible ethical position to make. But as soon as you start saying "LLMs are violating the licenses of our work by being a derivative of training inputs", it stops being an ethical position, and starts being an easily refutable incorrect interpretation of the copyright law.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 19:40 UTC (Fri) by dskoll (subscriber, #1630) [Link] (7 responses)

In this subthread, the argument was that LLMs' outputs being "derivative" of the input is allegedly more similar in spirit to how a human learns, rather than to being a "derivative work" in the copyright law sense.

OK. I disagree with that, though. We still don't really know how a human learns something, so I don't think we can say whether or not how LLM's are trained is "similar in spirit" to how a human learns.

And I think copyright law is wrong on this issue and is behind the times. It was written before LLMs existed and didn't anticipate them. In my opinion, any transformation of a work that is done purely mechanically by a machine with no or minimal human involvement (and a prompt certainly counts as "minimal") should be treated as a derivative work of its inputs.

I know that's now how it is. But IMO, from an ethical standpoint, that's how it should be.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 20:00 UTC (Fri) by mb (subscriber, #50428) [Link] (4 responses)

In my opinion, any transformation of a work that is done purely mechanically by a machine [..] should be treated as a derivative work of its inputs.

out = in XOR random

derivative work?

Ethical considerations, but not the right ones

Posted Aug 21, 2026 20:06 UTC (Fri) by dskoll (subscriber, #1630) [Link] (3 responses)

No, of course not, but that is not a transformation of in since any in is equally likely to end up as the same out. So out is totally independent of any particular in.

A transformation has to be dependent on the input. Your example is completely independent of it.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 20:13 UTC (Fri) by mb (subscriber, #50428) [Link] (2 responses)

Ok. Where do you draw the line? How much non-random bits in the stream make it a derivative work?
One constant zero in a billion random bits?
1% zero, 99% random?
50%, 50%?

I believe it's not so easy as "machine = derivative"

Ethical considerations, but not the right ones

Posted Aug 21, 2026 20:51 UTC (Fri) by rahulsundaram (subscriber, #21946) [Link]

> I believe it's not so easy as "machine = derivative"

Sure, it isn't that easy to determine direct copyright violation when humans are involved either. Google LLC v. Oracle America, Inc dealt with some of those nuances. It is entirely possible for court cases to show up where LLM usage itself maybe legal (New York Times Co. v. Microsoft Corp. et all will weigh in on that) but doing things along the lines of https://lwn.net/Articles/1061534/ is deemed illegal because it is materially different.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 22:59 UTC (Fri) by dskoll (subscriber, #1630) [Link]

I'm sure there are statistical measures of information and someone who knows information science a lot better than me can come up with them.

I'd say if more than around 1/3 (33%) of the in information is preserved in out, then it's a derivative.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 21:13 UTC (Fri) by marcH (subscriber, #57642) [Link] (1 responses)

> In my opinion, any transformation of a work that is done purely mechanically by a machine with no or minimal human involvement (and a prompt certainly counts as "minimal") should be treated as a derivative work of its inputs.

I think this should not matter much.

We must remember the rationale for inventing copyright! Similar to patents, copyright is an artificial and short-lived (ahem) monopoly for economical purposes. So, the main question is not who or what created something and how. Laws have defects too[*] but if you stick to the original copyright intent and purposes and to an "ideal" copyright world, then what matters is the _economical impact_ of derived works - irrelevant of _how_ they were created. Because consumers usually don't care how the output was done. Consumers may (and should) care whether the "inspiration" was legal or not, but they generally don't care what "type of Xerox machine" was used and that generally does not affect the market.

Of course, the less "creative" was the derivation process, then the more likely is the output infringing. It's not like you should completely ignore the copy/derivation/inspiration processes. Also, intentionality matters. But they are implementation details and definitely not be the main copyright concern.

[*] especially in a legislation where infinite money is "free speech"

Ethical considerations, but not the right ones

Posted Aug 21, 2026 22:48 UTC (Fri) by rgmoore (✭ supporter ✭, #75) [Link]

We must remember the rationale for inventing copyright! Similar to patents, copyright is an artificial and short-lived (ahem) monopoly for economical purposes.

The economic argument may have been the original argument for copyright, but it isn't the only one. International copyright law starting with the Berne Convention recognizes moral rights as well. Critically, one of the moral rights recognized by the Berne Convention is the right "to object to any distortion, modification of, or other derogatory action in relation to the said work, which would be prejudicial to the author's honor or reputation." For example, an artist who sculpted a nude statue could claim a prudish person putting clothes on it violated their moral rights by distorting the work. I don't know if this specific moral right would allow a software author prevent people from feeding the source code into an LLM, but it's at least worth thinking about. I still think releasing under a license that explicitly allows people to make derivative works would probably count as disclaiming those moral rights, but it would have to be litigated.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 17:23 UTC (Fri) by Cyberax (✭ supporter ✭, #52523) [Link] (46 responses)

> One could argue that. But there are two differences: (1) textbooks and lectures were created with the specific purpose of training people to be software developers.

LLMs also can learn by reading books.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 19:35 UTC (Fri) by dskoll (subscriber, #1630) [Link] (45 responses)

Can they, though? I would love to see an LLM produce production-ready software, but that has only been trained on undergraduate computer science textbooks and lecture notes.

I have my doubts that it would produce much useful output.

Ethical considerations, but not the right ones

Posted Aug 22, 2026 13:35 UTC (Sat) by anselm (subscriber, #2796) [Link] (43 responses)

I would love to see an LLM produce production-ready software, but that has only been trained on undergraduate computer science textbooks and lecture notes.

I have my doubts that it would produce much useful output.

Probably, but in that respect I wouldn't expect too much from actual human CS undergraduates, either.

Ethical considerations, but not the right ones

Posted Aug 22, 2026 13:45 UTC (Sat) by dskoll (subscriber, #1630) [Link] (42 responses)

I think the good undergraduates would do better than the best LLMs.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 7:25 UTC (Mon) by taladar (subscriber, #68407) [Link] (41 responses)

Do good CS undergraduates who only read the class supplied materials and have had no contact with IT/CS related materials outside of school even exist?

Ethical considerations, but not the right ones

Posted Aug 24, 2026 11:29 UTC (Mon) by mathstuf (subscriber, #69389) [Link]

Indeed…I rather think we have had more of a Ramanujan kind of thing where some of the geniuses have come from far afield of "traditional" education. But seeing as universities still treat VCS and collaboration as a side effect of group projects rather than something worthy of teaching on its own[1], there are definitely important things missing from a barebones curriculum.

[1] Sure, they're teaching "CS" and not "software engineering" (under any acceptable name due to the legal power of "engineer" in places), but I'm not aware of a large contingency of no-VCS-using job prospects using a CS degree.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 13:18 UTC (Mon) by dskoll (subscriber, #1630) [Link] (36 responses)

No, probably not. But I also doubt any undergraduate has ingested gigabytes of source code from tens of thousands of projects, the way LLMs are trained.

I would love to see the output of an LLM that gets the same training experience as a person. But I don't know if it's even possible to run such an experiment.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 13:53 UTC (Mon) by intelfx (subscriber, #130118) [Link] (35 responses)

But I also doubt any undergraduate has ingested gigabytes of source code from tens of thousands of projects, the way LLMs are trained.

I would love to see the output of an LLM that gets the same training experience as a person. But I don't know if it's even possible to run such an experiment.

No, because a person learns and builds their world model by continually ingesting vast amounts of information, in multiple modalities, starting at birth. By the time a person learns Computer Science it is more like fine-tuning their brain to obtain specific domain expertise, while an extensive and mostly complete world model is already in there.

An LLM has no such luxury; it can only learn on language. It is reasonable that it would require much more of it.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 14:09 UTC (Mon) by dskoll (subscriber, #1630) [Link] (34 responses)

Yes, which reinforces my doubt that LLM training is "just like" a human studying something. I think it is not, and I think the difference makes LLM outputs derivative works of their inputs, despite the fact that copyright law in the United States (wrongly IMO) seems to assert that isn't the case.

Ethical considerations, but not the right ones

Posted Aug 25, 2026 7:55 UTC (Tue) by taladar (subscriber, #68407) [Link] (33 responses)

Consider it similar to a blind human studying movies and other visual art forms. They would need a lot more to learn everything about a single movie than the 90 minutes a seeing person requires to just watch it.

Ethical considerations, but not the right ones

Posted Aug 25, 2026 13:01 UTC (Tue) by dskoll (subscriber, #1630) [Link] (32 responses)

90-minute movies nowadays are rarer than hens' teeth. 🙂

But I still don't think this is a good analogy. I think the way LLMs are trained is really fundamentally different from how humans learn. And that's why I believe copyright law should consider the output of an LLM to be derived from its inputs, while (except in cases of obvious plagiarism) I don't think the output of humans should be considered to be derived from the educational materials they used to learn their skills.

Humans can come up with something fundamentally new. For example, I don't think that LLMs trained on the version-control systems available prior to Bitkeeper would ever have come up with a distributed version-control system.

Ethical considerations, but not the right ones

Posted Aug 25, 2026 15:33 UTC (Tue) by kleptog (subscriber, #1183) [Link] (1 responses)

> And that's why I believe copyright law should consider the output of an LLM to be derived from its inputs, while (except in cases of obvious plagiarism) I don't think the output of humans should be considered to be derived from the educational materials they used to learn their skills.

For what purpose though? Copyright law is written to achieve certain policy objective. The US has text like "to promote the progress of science and useful arts". In the EU there are also strong interests in protecting authors and enabling them to obtain remuneration from their works. How does declaring the model output a derivative work of the training inputs achieve those objectives?

If the issue is the compensation of authors for their works, that's a separate issue. EU Law has some mechanisms for that that don't require declaring the model a derivative work.

I do wonder though. If people (in the US) expect to get paid for works used in training and they don't expect Congress to create an actual remuneration scheme, I can see why trying to convince a court to declare the model a derivative work would be attractive since it gives the courts mechanism to work with. That doesn't make it a good choice.

Ethical considerations, but not the right ones

Posted Aug 25, 2026 15:58 UTC (Tue) by dskoll (subscriber, #1630) [Link]

For the purpose of protecting creators of works from the misuse of their works. I know the US doesn't really have the concept of "moral rights", but many other countries do.

You also left out some text after "to promote the progress...", which reads:

To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.

In my opinion, the exclusive right to works includes being allowed to control whether or not they are used to trail LLMs, for the reasons I've mentioned in other comments. Even if US courts disagree.

And I think allowing LLMs free rein to be trained on copyrighted material will, in the end, hurt the Progress of Science and useful Arts. It will flood the market with cheap output, meaning humans will be disincentivized to produce works, being unable to make a living at it. The self-published book industry, for example, is already collapsing under the weight of AI titles.

Ethical considerations, but not the right ones

Posted Aug 27, 2026 17:21 UTC (Thu) by Cyberax (✭ supporter ✭, #52523) [Link] (29 responses)

> But I still don't think this is a good analogy. I think the way LLMs are trained is really fundamentally different from how humans learn.

I don't think that this is a _fundamental_ problem. Shoving massive amounts of uncurated data is just the easiest approach right now. Humans learn iteratively: by ingesting information, trying to manipulate the world, and then updating the weights. Doing that with the current AI training infrastructure is not practical.

That's why it's a wrong question to ask. Would it change anything if AI models were trained differently? If some company starts training AI models in a virtual world where the AI agent can do iterative development.

We already have ample evidence that terabytes of data are not required for training because models can be distilled and quantized into a small fraction of the initial model.

Ethical considerations, but not the right ones

Posted Aug 27, 2026 18:39 UTC (Thu) by dskoll (subscriber, #1630) [Link] (28 responses)

Would it change anything if AI models were trained differently?

Speaking strictly from the perspective of protecting humans' creative works: yes, absolutely. If LLMs were trained using the same textbook and lecture notes as CS undergraduates, and then learned on their own iteratively, I would have no quarrel with their output from a copyright perspective.

(I have other problems with LLMs and the GenAI industry, but let's just concentrate on the copyright issue here.)

We already have ample evidence that terabytes of data are not required for training because models can be distilled and quantized into a small fraction of the initial model.

When we get to the point where even the initial terabytes of data are not required, as I said: I'll have no quarrel with LLM output. But IMO, even the distillation and quantization steps don't make things right. With current LLMs, the output is still completely derived from the input.

Ethical considerations, but not the right ones

Posted Aug 27, 2026 19:23 UTC (Thu) by mb (subscriber, #50428) [Link]

> If LLMs were trained using the same textbook and lecture notes as CS undergraduates, and then learned on their own iteratively

Well, that is actually not how humans learn, though.

Humans have eyes and other senses which capture massive amounts of data every minute. It is a massive data stream reaching the brain. Constantly all the time.

Computer science text books base many things on this massive lifetime experience. For an example take object oriented programming. We all know the animal-dog-cat class hierarchy example. This only makes sense with big amounts of pre-existing ingested data dating back to pre-school books and walks through the city observing cats and dogs.

Therefore, of course, it's impossible to become a computer programming expert merely by reading computer programming textbooks.
This doesn't work in humans and it doesn't work in AIs.

Ethical considerations, but not the right ones

Posted Aug 27, 2026 19:44 UTC (Thu) by Cyberax (✭ supporter ✭, #52523) [Link] (15 responses)

> Speaking strictly from the perspective of protecting humans' creative works: yes, absolutely. If LLMs were trained using the same textbook and lecture notes as CS undergraduates, and then learned on their own iteratively, I would have no quarrel with their output from a copyright perspective.

Why? The bits will be the same and just their color is going to change.

And conceptually, we have already done this with regular writing. That's really what "model distillation" is. A larger model is used to guide a smaller model so it can get feedback on its output and correct its behavior. This external model is really a simulation of the world.

I'm almost sure that you can take a basic LLM, feed it CS notes, and then have it iteratively try to apply its knowledge in a sandbox, giving it feedback. It will be a very inefficient way to train a model, but it will likely work. If somebody wants to try, there's Nematron from NVidia that can be used to train a model from scratch.

That's why I don't think that it makes sense to base the entire moral and legal framework on an assumption that training can't be performed differently.

Ethical considerations, but not the right ones

Posted Aug 27, 2026 21:16 UTC (Thu) by dskoll (subscriber, #1630) [Link] (14 responses)

The bits will be the same and just their color is going to change.

Because, like it or not, this makes a huge difference in copyright law. It happens a lot in music that two different composers come up with the same phrase or snippet of music, and lots of money rides on whether or not it was a coincidence or deliberate copying.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 6:36 UTC (Fri) by Cyberax (✭ supporter ✭, #52523) [Link] (10 responses)

In my mind, it's the outcome that matters more. The laws don't exist as some immutable things set in stone, they exist to further some goals.

So I don't think that just latching onto the copyright infringement angle is a good idea. It can be worked around (and likely it will be done soon). And/or the laws will be adjusted or interpreted in a way that allows the AI use.

Instead, we should focus on _consequences_ of the widespread availability of AI and try to manage them. Or maybe outlaw some kinds of AI models outright if their harm is deemed too large. I still don't have a firm opinion on the best course of action.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 12:23 UTC (Fri) by dskoll (subscriber, #1630) [Link] (9 responses)

I agree, and I think one of the goals of copyright law should be to encourage human creativity, which I think AI discourages

I no longer work as a professional software developer, having retired three years ago, but the people I know who are still in the business find it much less fulfilling than before, specifically citing AI as the major cause, and some actively hate the profession. I also hear stories from my daughter about how recent graduates find it hard to get jobs in the industry, and how fewer people are going into the profession.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 12:40 UTC (Fri) by bluca (subscriber, #118303) [Link] (8 responses)

It's really the opposite for everyone I know: it is so much more enjoyable to work in the field thanks to copilot&co which can take over the boring, rote and mindless tasks that are an unavoidable part of the job. E.g.: sifting through dozens of CI runs, each with hundreds of thousands of log lines, to investigate a test regression. Or to remain closer to the article's topic, to compile the debian/copyright file. I've _never_ met anyone who told me "yes, I enjoy manually scraping hundreds or thousands of source files by hand, correlating every git author with every file with every year of activity to manually write a policy-compliant debian/copyright that passes NEW queue review"

Ethical considerations, but not the right ones

Posted Aug 28, 2026 12:41 UTC (Fri) by dskoll (subscriber, #1630) [Link]

OK. I guess we travel in different circles, then. ¯\_(ツ)_/¯

Ethical considerations, but not the right ones

Posted Aug 28, 2026 13:57 UTC (Fri) by NAR (subscriber, #1313) [Link] (6 responses)

I've noticed that e.g. copilot is used in many cases where previously one might have written a (throwaway) shell/perl script to do the work. For those who don't know how to write that shell/perl script, copilot is a godsend as it saves manual work. For those who can write those script, they might feel like AI is taking away part of the enjoyable work.

There's an other issue when AI is used to generate production code. In this case the developer has to review the generated code and that might be a lot less enjoyable than writing the code in the first place.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 14:13 UTC (Fri) by geert (subscriber, #98403) [Link] (5 responses)

Exactly.

And how do you know the result you got from e.g. copilot is correct? OK, you're script might not be correct either.
Will you get the same response next week? At least your script should give you the same results next week.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 21:33 UTC (Fri) by NAR (subscriber, #1313) [Link]

On the other hand just today I had to do some refactoring on the test suite. The code change was fairly simple, but a little more complicated than that a simple sed command could solve. Creating a script that would handle refactoring the multiline calls seemed overkill as there were maybe 25-30 function calls across 6 or 7 files that needed to be changed - and I would never reuse that script again. I could do this manually (as 25-30 changes are not that much) or I could ask copilot to do this. And to my surprise (or I just managed to create the right prompt) it did a really good job, especially that it run the changed tests to check that everything works (I did not prompt to run the tests, so I was surprised to see it happening). Due to all those test runs it wasn't particularly fast, but maybe faster than me. Then I reviewed the code, noticed an other refactoring opportunity, created manually the new function (I felt it's simpler than explaining what that 5 lines long function would do), and prompted to refactor. There were two more similar steps. I was pleasantly surprised how well it went. During this refactoring I was able to work on an other task (that also wasn't complicated, so I didn't get overloaded with the context switches) - and without copilot running and evaluating the tests I don't think I would have been able to make such progress on the other task, because I would have need more of my "cycles" on the test runs and evaluations.

I also see that had we have a junior developer, this refactoring would have been an excellent opportunity for the junior developer to learn something...

Ethical considerations, but not the right ones

Posted Aug 28, 2026 21:42 UTC (Fri) by kleptog (subscriber, #1183) [Link] (3 responses)

You know it's correct because it's often far easier to check if the result is correct than it is to write the code. (see also NP-complete)

For example, this week we used Copilot to recreate a lost Terraform state file given the Terraform code and an export of the router config. Sure, I could have written the scripts myself, I estimate it would have taken two of us a few days. Instead we watched while the agent, with the occasional feedback from running 'terraform plan', reconstructed the state. It wrote scripts, analysed the results, saw where it screwed up, fixed itself and continued. And we knew it worked because 'terraform plan' confirmed no changes would be made. Took an hour. And a few $ in tokens.

We won't need these scripts ever again. If it ever happens again the situation is unlikely to be comparable.

When I get it to write code for production, it also writes the test cases and then verifies the code against the test cases. Validating test cases is much easier than validating the original code.

I'm one of those people who like getting things to work more than the act of coding. So AIs work well for me. I prompt it to make some changes, review it, hand edit the parts that are unclear or wrong, commit, rise and repeat. Another developer reviews it as well. I agree you need to keep the changes small and focussed, otherwise it gets harder for everyone.

I also like how it teaches me new things I didn't know about. Useful if you're not working with peers on the same thing.

Ethical considerations, but not the right ones

Posted Aug 29, 2026 7:36 UTC (Sat) by NAR (subscriber, #1313) [Link] (2 responses)

A couple of months ago I asked copilot to implement a small feature, including test cases. It would have took me an afternoon, but I wanted to see how copilot fares. Maybe 50 lines of code, but due to an unfortunate architectural decision, it's somewhat fragmented and hard to follow. Copilot created the code, the testcases, everything passed, coverage was 100%, looked good. I tried out in the test environment and a simple example worked. All of this was done in maybe two hours or less.

The next day a coworker reviewed the PR and said there's something fishy. I looked at the code again and after an hour I realised that instead of returning the result of the computation, it returned the input data (I'm simplifying things) - which happened to be correct most of the time, so even the tests passed, but it was still wrong (or the mocking was slightly incorrect? - I don't remember). Then I got to spend the afternoon to rewrite and fix it.

So eventually I spent more time on the task with copilot than I would have without copilot. My gut feeling now is that I need to write (at least some of) the tests first before I let copilot implement the task. Then maybe ask copilot to create more tests to increase coverage.

Ethical considerations, but not the right ones

Posted Aug 29, 2026 8:24 UTC (Sat) by mb (subscriber, #50428) [Link] (1 responses)

I need to write (at least some of) the tests first before I let copilot implement the task.

Pre-existing tests are extremely useful, but writing them by hand before doing an automated implementation is almost never the way to go.

I think you probably didn't set one of the many levers quite correctly. This is a learning process. These tools are complex and it takes a fair amount of time to lean how to use them correctly.

There are many things that could have been wrong here: Chosen a too weak LLM; Too small context window; Incorrect prompt; Incomplete prompt; Missing context information about something (e.g. your "unfortunate architecture"); etc, etc..

Remember that if you already know something is hard (e.g. arch is "hard to follow") it will also be hard for the LLM to figure out. Tell it about your pre-existing insights.

Ethical considerations, but not the right ones

Posted Aug 30, 2026 10:45 UTC (Sun) by NAR (subscriber, #1313) [Link]

This is one of the reasons I prefer LWN to follow the AI discourse. I'd like to learn how to use it well, learn from others how to use it effectively.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 7:31 UTC (Fri) by taladar (subscriber, #68407) [Link] (2 responses)

I would consider that one of the greatest flaws in copyright law that it can punish independent creation of things that look similar enough. Especially in constrained fields like music or algorithms.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 12:24 UTC (Fri) by dskoll (subscriber, #1630) [Link]

Yes, I agree that's a problem in music. You can't copyright an algorithm; there, I think the big threat is software patents.

Ethical considerations, but not the right ones

Posted Aug 28, 2026 14:12 UTC (Fri) by farnz (subscriber, #17727) [Link]

At least in theory, it is a requirement for something to be copyright infringement that there was some form of copying, not independent creation.

This is why "clean-room reverse engineering" exists in the software field; because you can prove that the people alleged to have copied your thing did not have access to it (and hence could not have copied it), you've demonstrated that there cannot be copyright infringement. Of course, proving that you didn't have access to the original is harder in fields like music - but you'll note if you look at infringement cases that the plaintiff states why they believe the copier could have copied the original (generally by pointing to the wide distribution of the original).

Ethical considerations, but not the right ones

Posted Aug 27, 2026 21:05 UTC (Thu) by kleptog (subscriber, #1183) [Link]

> When we get to the point where even the initial terabytes of data are not required, as I said: I'll have no quarrel with LLM output

I actually think we're going to end up in a different situation altogether. I think the actual models themselves are going to be relatively small and run on people's machines locally. The magic will actually be that the models will query for extra information on demand.

Look at how Google does it now: you do a query, it does a search using the Google algorithm and uses a model to summarise the results to answer your question. Hence you don't need the model to actually remember any of the information in the training set because it will get the information it needs on demand.

So in a sense all those data centres to run AI models aren't going to get used because we run them locally (a win for us) but these local models will need some external index of actual knowledge to be able to answer actual useful questions. This is the same problem with have with teaching students: do we expect them to memorise everything or just understand enough to lookup the details when needed? Are we going to all have a local model with a monthly subscription to the Google AI knowledge index?

This doesn't really address your question of needing terabytes of data to train an LLM, but I think the ROI for training an LLM from scratch is rapidly diminishing. Distilling from existing models is so much cheaper and more effective. The entire concept to "input corpus" doesn't really make a lot of sense anymore if a model isn't trained on data but on one or more other models. At some point we will have relatively small models in common use where determining the pedigree will be difficult. None of these processes is deterministic.

Ethical considerations, but not the right ones

Posted Aug 29, 2026 21:10 UTC (Sat) by aigarius (guest, #7329) [Link] (9 responses)

Textbooks are proprietary. If you are at any point considering that copyright survives the training process, then all software produced by an LLM (or a person) trained on those textbooks would become completely undistributable. And with that precedent established, everyone who has *ever* read a coding textbook (with a restrictive license) or any non-free software source code (like at work) would become automatically forbidden from ever contributing to any *real* free software as anything they produce would be a derivative work from those non-free sources and thus non-free.

Any legal finding that copyright survives training kills the whole free software ecosystem. Oh and the rest of the IT in the whole world as well.

Ethical considerations, but not the right ones

Posted Aug 29, 2026 21:39 UTC (Sat) by dskoll (subscriber, #1630) [Link] (6 responses)

Textbooks are proprietary, but I do not believe the output of a human programmer is a work derived from the textbook itself. A computer program does not bear much resemblance to a textbook.

Whereas I do believe that the output of an LLM trained on source code is derived from that source code itself.

Ethical considerations, but not the right ones

Posted Aug 29, 2026 21:51 UTC (Sat) by mb (subscriber, #50428) [Link] (5 responses)

>but I do not believe the output of a human programmer is a work derived from the textbook itself

Yeah, but why?

Ethical considerations, but not the right ones

Posted Aug 29, 2026 23:18 UTC (Sat) by dskoll (subscriber, #1630) [Link] (4 responses)

Does the source code of an application resemble a textbook? Not at all. A textbook might contain small snippets of code (probably too small to even be copyrightable) but consists mostly of explanations, discussion and problems. Now, if an LLM was trained on textbooks and spat out a textbook as output, then that would be a derived work.

Does the output of an LLM resemble its training inputs? Yes, quite a lot.

Ethical considerations, but not the right ones

Posted Aug 31, 2026 14:35 UTC (Mon) by geert (subscriber, #98403) [Link] (2 responses)

ISBN 0-262-24039-4 is the classical counter-example ;-)

https://en.wikipedia.org/wiki/Pretty_Good_Privacy#cite_no...

Ethical considerations, but not the right ones

Posted Aug 31, 2026 14:57 UTC (Mon) by anselm (subscriber, #2796) [Link]

That's not such a great example because IIRC, as a textbook it kinda sucks – it has lots of code but not that much (explanatory) text. Remember that its main raison d'être was to make it possible to export PGP from the US, not to teach people how cryptography works.

If you want an actual college-level textbook with both explanations and nontrivial code, consider Andrew S Tanenbaum's Operating Systems: Design and Implementation, which contains background material on operating system theory as well as significant amounts of code from a working operating system (MINIX).

Ethical considerations, but not the right ones

Posted Aug 31, 2026 14:57 UTC (Mon) by dskoll (subscriber, #1630) [Link]

Sure, and another example is Andrew Tanenbaum's book on Minix. But they are outliers, and AFAIK had specific licenses regarding reuse of the source code.

Ethical considerations, but not the right ones

Posted Aug 31, 2026 18:54 UTC (Mon) by Cyberax (✭ supporter ✭, #52523) [Link]

> Does the source code of an application resemble a textbook? Not at all.

Back when Internet was still young, many textbooks included pretty large source code listings. Up to and including full applications.

This very much included books on individual APIs or programming languages. E.g.: https://archive.org/details/programmingwindo0000petz_l2m9...

Ethical considerations, but not the right ones

Posted Aug 29, 2026 21:39 UTC (Sat) by pizza (subscriber, #46) [Link]

> Textbooks are proprietary.

Many are, not all.

...But any code in a textbook can easily be under a different license than the rest of the textbook.

(tl;dr: It's complicated)

Ethical considerations, but not the right ones

Posted Aug 29, 2026 22:16 UTC (Sat) by anselm (subscriber, #2796) [Link]

Textbooks are proprietary. If you are at any point considering that copyright survives the training process, then all software produced by an LLM (or a person) trained on those textbooks would become completely undistributable.

Copyright is about the concrete expression of ideas, not the ideas themselves. So if, for example, you have read about the Quicksort algorithm in a (proprietary) textbook you are then perfectly free to write your own independent program implementing Quicksort without having to worry about infringing upon the copyright on the textbook. The copyright on the textbook covers the (English?) explanation of Quicksort in the book, plus possibly a code sample that is printed alongside, but the copyright on the textbook does not then cover your own implementation of the algorithm in the book, since the algorithm itself is an abstract concept that cannot be copyrighted.

In other words, the copyright on the textbook prevents you from making and selling identical copies of the textbook. It does not prevent you from reading the textbook and applying the knowledge you gained from that to your own works.

Therefore, whether “copyright survives the training process” is not the issue at hand. Nobody argues that it does in general. But that doesn't mean that if, hypothetically, an LLM were to produce a sufficiently non-trivial piece of code that is substantially identical to some pre-existing copyrighted piece of code that the LLM was trained on, this LLM output wouldn't potentially infringe upon the copyright of the pre-existing code in the same way that it would if it had been produced by a human being.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 15:49 UTC (Mon) by kleptog (subscriber, #1183) [Link] (2 responses)

I'm always surprised by new CS graduates that have never touched GitHub and never used git outside of a few class projects.

When I did my degree way back they made sure I understood Pi and Lambda calculus, and programming supercomputers and SPARC workstations in assembly was super interesting. However, the direct value to my day-job was marginal at best. Everything I learnt about the software development process I learned on the job.

Now, Agile didn't exist at the time so they couldn't teach me that, and I don't expect universities to teach students how to use Jira given the pricing. But CS students who apply routinely have less experience in software engineering than I would hope. Not really a problem, but it just shows that a degree by itself isn't always a useful bar.

(My favourite course was by Tridge which included modifying a Linux kernel module. That one was, surprisingly, useful in $DAYJOB. All pre-git/bitkeeper though.)

Ethical considerations, but not the right ones

Posted Aug 24, 2026 20:08 UTC (Mon) by mathstuf (subscriber, #69389) [Link]

My favorite course was also taught by FOSS community members. They were also employees at my now-$DAYJOB. One project was to, basically, fix an issue in a project you use. Mine was to fix some (pre-LLM) AI corner case in Wesnoth related to recruiting (IIRC, it was that if the leader started in a recruiting spot and could reach another this turn, recruit, move, recruit). I had been doing FOSS for a long time before that (Fedora and the KDE SIG), but it was the first time a lot of classmates had actually engaged with the community around some piece of software they used every day.

Ethical considerations, but not the right ones

Posted Aug 25, 2026 18:40 UTC (Tue) by mbunkus (subscriber, #87248) [Link]

Three of my favorite practical experiences from uni (back around ~2000):

  1. Audio/video playback via network

This was a workshop running a whole fifth (?) semester. Given was a number of uncompressed images & correlating audio files (not interleaved in a container, all were separate). We had to write both a network server & a networt client & a simple protocol to make them communicate. The client would then have to play back audio & video in sync on a severely resource constrained Sun workstation — so resource constrained we could not make more than one memcpy of each video frame (and that one was from network receive buffer to X windowing system).

It taught us so much about low-level programming, network primitives, round-robin buffers, audio/video sync, network protocol design, and even Little vs Big Endian things.

  1. Programming in Scheme

This was during the first semester, the first actual hands-on programming, on HP-UX mainframes. It was also the first time I even encountered the idea of a functional programming language. It opened my mind to there being so much more than procedural or object-oriented programming, flexible techniques such as "code is data", "functions as first-class objects", why const is important (as in, no side effects in functions). Even if I havent't really done much functional programming over the years, I still use what I learned there daily.

  1. Software development workshop

This was during the… sixth or seventh semester, I think (this is a German university where the full Diplom was usually five years, back before the Bachelor reforms).

All ~140 or so CS students in that year had the same task: to build a game, or rather a strategy plugin for an existing game server that implemented a game. What we were given was a detailed specification about the interface the server offered (a strategy plugin was a loadable Java class) & the functionality we were supposed to implement.

Then we were divided into teams of about five or six people who each had to come up with their own implementation. We did not have access to the working server, we only had the specification to develop against. Meaning we couldn't actually test it with the final product — only with a mock we wrote ourselves.

What was really fun about it was that at the end of the semester all those strategy plugins would do battle against each other. Not only would this test whether or not we had written the plugin according to spec, it also tested if we came up with a winning strategy.

Furthermore it taught me the importance of communication (something I really wasn't good at in those days), the pitfalls or working in teams (one or two always slack off, others do way more than average, some cannot be relied on etc.), the value in having diverse experiences & ideas when designing solutions.

Today I use none of the languages or tools I used back then. But all of those lessons stuck hard.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 14:49 UTC (Mon) by paulj (subscriber, #341) [Link]

Ooh, that's a very good example.

And the human will then be able to further train and refine their own ability through their own practice of what they have learned. An area current AI tech is almost completely devoid in ability.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 18:53 UTC (Fri) by rgmoore (✭ supporter ✭, #75) [Link]

I think the outputs of an LLM are in fact a derivative of their inputs.

That depends on what definition of "derivative" you're talking about. In the ordinary lay person definition, you can certainly argue that the outputs of an LLM are derivative of their inputs. Of course you can make the same kind of argument about a person's output and their education, training, and experience. A lot of the way I write is derivative of the classes I took in junior high school where I was required to write compare and contrast essays.

The definition of "derivative" copyright lawyers care about is much stricter. Copyright law only considers a work to be a derivative if it contains substantial copyrightable elements of another work. Unless and until you can point to something specific that is copied from the input corpus, LLM output isn't derivative of its inputs in a legal sense. This is important because it's the legal consequences of copyright law that give the GPL its force. If the code isn't derivative of its GPLed inputs in the sense copyright law cares about, you have no legal right to demand those outputs be released under the GPL.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 15:02 UTC (Thu) by sionescu (subscriber, #59410) [Link]

> It seems to me that feeding source code into an LLM as part of its training corpus falls squarely into the category of "study[ing] how the program works".

It doesn't. Studying is done by humans. Feeding it into an LLM, thus augmenting it, is not studying.

Ethical considerations, but not the right ones

Posted Aug 20, 2026 0:34 UTC (Thu) by Wol (subscriber, #4433) [Link] (1 responses)

> I appreciate that many in Debian are evaluating this on ethical grounds. But I'm always sad to see how dominant the environmental message is, when that just seems like an irrelevant distraction compared to everything else. Many, many things consume energy, and we're improving in how we generate energy, and I don't think LLM usage is particularly unusual in that regard.

I'm just shocked how wasteful Computer Science is! Why are modern computers so much slower and more wasteful of resource than the old stuff?

> we're improving in how we *generate* energy

And if we weren't getting so much better in how we WASTE energy, we wouldn't be in the environmental mess we're in.

BigQuery regularly tells me how much data it needs to read to answer my queries. As far as I can tell it reads the *entire* *table* for every table I reference! What happened to old fashioned ideas like primary keys, and indices, to try and achieve some modicum of efficiency?

Pretty much every modern advance in Computer Science seems to me to be summed up as "let's throw more compute power at it" :-(

Cheers,
Wol

Ethical considerations, but not the right ones

Posted Aug 20, 2026 9:16 UTC (Thu) by joib (subscriber, #8541) [Link]

> BigQuery regularly tells me how much data it needs to read to answer my queries. As far as I can tell it reads the *entire* *table* for every table I reference! What happened to old fashioned ideas like primary keys, and indices, to try and achieve some modicum of efficiency?

I think we can safely assume that the people behind these warehouse scale data management systems are aware of previous work.

In physics, people who claim to have invented a theory of everything but clearly demonstrate that they understand neither relativity nor quantum mechanics are, entirely correctly, labeled as crackpots.

Advances in science and engineering are, as a rule, made by people intimately familiar with the status quo and the limitations of it.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 8:56 UTC (Fri) by sam.thursfield (subscriber, #94496) [Link] (1 responses)

> I'm always sad to see how dominant the environmental message is, when that just seems like an irrelevant distraction compared to everything else. Many, many things consume energy, and we're improving in how we generate energy, and I don't think LLM usage is particularly unusual in that regard.

You could make the same case about everything! In terms of energy used per mile travelled, an aeroplane is pretty efficient. But then, commercial airlines are huge consumers of kerosene, which drives oil drilling, which leads to venting of gas directly into the atmosphere, and that is one of the biggest drivers of the greenhouse effect...

An LLM could be solar powered, but the reality is that AI companies are literally setting up gas turbines to power them at the moment.

Ethical considerations, but not the right ones

Posted Aug 21, 2026 9:00 UTC (Fri) by bluca (subscriber, #118303) [Link]

> An LLM could be solar powered, but the reality is that AI companies are literally setting up gas turbines to power them at the moment.

_some_ AI companies are doing that yes, and that's a problem. A problem to be solved by enforcing strict regulations on _all_ industries regarding energy production, consumption, pollution, etc. etc. Gas turbines weren't invented by AI companies. The vast, vast majority of gas burned is not by AI companies. Make every planning application for any industrial site require building renewables capacity and grid upgrades to match their energy consumption, then when/if they go bust/close shop, we keep the energy infrastructure.

Ethical considerations, but not the right ones

Posted Aug 24, 2026 7:09 UTC (Mon) by tdz (subscriber, #58733) [Link]

> [...] I'm always sad to see how dominant the environmental message is, when that just seems like an irrelevant distraction compared to everything else.

Lucky you doesn't depend on a stable environment for your survival.

My (inconsequential) vote

Posted Aug 19, 2026 19:54 UTC (Wed) by dskoll (subscriber, #1630) [Link]

I'm just a Debian user, not in a voting position, but I like 1, 7 and 3 best, in that order.

Uncertain

Posted Aug 20, 2026 2:15 UTC (Thu) by ErikF (subscriber, #118131) [Link]

I have no horse in this race (I use Gentoo!), but I've seen over the years that statements about issue-of-the-<time period> topics tend to get stale quite quickly. Until the significant changes in AI slow down, I feel like putting a too-specific clause in a foundational operating document seems premature or else the probability of reconsideration and churn goes up. That said, I'm honestly torn about AI: it definitely has interesting positive possibilities in the future (it's quite good as a tool-assist IMO), but I also see how unchecked it could contribute to a measurable loss of knowledge and quality thought in the computing space (among many other fields).

Options are a mess

Posted Aug 20, 2026 10:25 UTC (Thu) by taladar (subscriber, #68407) [Link]

Is that just me or do those options read like a very messy first draft that could use some iteration to both clearly delineate differences and similarities between them and also just think through some of the burden put on the user under those rules.

To take just one example I just hate the kind of rule like mandatory LLM disclosure combined with unspecified form of disclosure, that just feels like someone didn't want to put in the work of agreeing on a form of disclosure once, instead leaving to to every little part of the community to hash out later inconsistently.

All decent proposals + some good news

Posted Aug 20, 2026 12:51 UTC (Thu) by jpeisach (subscriber, #181966) [Link] (7 responses)

I have varying opinions, and sometimes they change from time to time when it comes to LLMs, but basically, as long as you don't make it so future generations and current students (like me) won't be able to, uh, get a job, because the current people in charge decided to make usage of an LLM to replace a job acceptable, and then the more higher up role (like being a "Senior Software Engineer") would be out of reach to people just entering the job market... then that's fine.

The discussion around LLMs is always tense, but for optimism, I am faithful that humanity can figure out a solution to environmental impact (lessen it as much as possible). And there are, every now and then, some kind of project where an LLM would be okay to use. For example, I saw a post on social media about someone using it to make a printer driver for I think macOS, because the manufacturer didn't support it.

All decent proposals + some good news

Posted Aug 20, 2026 16:22 UTC (Thu) by rbranco (subscriber, #129813) [Link] (6 responses)

Clankers won't get our jobs. If anything, they're making regular folks venturing out of their confort zones writing code in languages they don't need to fully master.

You don't need to be Edgar Allan Poe to write in English. The same is now true for programming languages like Rust.

Rational criticism of LLMs is needed and that's why the Vegan criticism of AI should enter the quiet phase where Vegans are no longer harassing meat-eaters in the street. Now Vegan is just another option in the restaurant that deserves respect but it's just a choice among many.

All decent proposals + some good news

Posted Aug 20, 2026 19:08 UTC (Thu) by jpeisach (subscriber, #181966) [Link] (5 responses)

> Clankers won't get our jobs. If anything, they're making regular folks venturing out of their confort zones writing code in languages they don't need to fully master.

Well, it's the jobs for the manual code writing labor that are being taken out, and I guess the role up would be doing all the designing and the prompting. But certainly the oversaturation of the job market makes it unviable for people just entering the job market, and how else are you supposed to get the "work experience" for those higher roles?

(this is becoming off topic, so I'm just saying this as something to think about)

All decent proposals + some good news

Posted Aug 21, 2026 8:49 UTC (Fri) by rbranco (subscriber, #129813) [Link] (4 responses)

> Well, it's the jobs for the manual code writing labor that are being taken out, and I guess the role up would be doing all the designing and the prompting. But certainly the oversaturation of the job market makes it unviable for people just entering the job market, and how else are you supposed to get the "work experience" for those higher roles?

It's called creative destruction. It will create more jobs than it destroys.

Nobody writes Assembly anymore. People don't need to learn Rust to get stuff done.

Nobody writes assembly anymore

Posted Aug 21, 2026 15:17 UTC (Fri) by ebiederm (subscriber, #35028) [Link] (2 responses)

I beg to differ.

In performance and operationally critical pieces of code (implementation of locks) people still write assembly.

Even more people read assembly to understand why their code behaves the way it does.

The need to read assembly has never gone away.

As we climb higher in the stack the idea that we won't need people to understand the lower layers is preposterous, and dangerous nonsense.

Nobody writes assembly anymore

Posted Aug 21, 2026 15:38 UTC (Fri) by dskoll (subscriber, #1630) [Link] (1 responses)

My boyfriend writes entire graphical applications in assembly language. He even co-maintains an assembler and linker that are written in assembly. But people who do that are pretty unusual.

Unfortunately, he's a Windows developer. It's the big religious difference in our relationship.

Nobody writes assembly anymore

Posted Aug 21, 2026 20:38 UTC (Fri) by Klaasjan (subscriber, #4951) [Link]

That intriguing explanation begs the question: what architecture? (arm or amd, assuming 64bit). Or has assembly become portable across architectures since I last looked at it?

All decent proposals + some good news

Posted Aug 21, 2026 16:11 UTC (Fri) by mathstuf (subscriber, #69389) [Link]

Maybe a return of the QA department?

AI companies destroying books

Posted Aug 28, 2026 5:00 UTC (Fri) by dvdeug (guest, #10998) [Link] (8 responses)

"We don't have to allow AI companies to destroy books to train LLMs." We live in a capitalist society, where people and companies can buy books, and people can destroy their own property. There's a lot of things we could do; eminent domain Google Books' database, and make all the scans of public domain work widely available. We could put real funding into national book scanning projects that would make sure to make the scans available to everyone. We could make it clear (with funding) that major libraries should keep copies of works instead of continually disposing of unused works to make work for new books. But stopping AI companies from destroying books they own is very intrusive and doesn't do anything to make those works available today or in the future.

AI companies destroying books

Posted Aug 28, 2026 7:48 UTC (Fri) by anselm (subscriber, #2796) [Link] (7 responses)

Many countries require publishers to deposit copies of their published works with a national library of some sort. In theory, this should ensure that these works remain accessible to the public (even though this may mean having to physically visit the library in question).

In any case, “AI company X destroys book Y” usually means they find and buy a copy of book Y, scan that, and then dispose of the physical book, mostly because quirks of copyright law require them to do this (plus it would be a hassle to hang on to all those physical books – but otherwise they could just get books from the lending library at virtually no cost, scan them, and return them later). It doesn't mean that AI company X rounds up all copies of book Y to take it out of circulation for everyone else.

AI companies destroying books

Posted Sep 2, 2026 15:14 UTC (Wed) by dvdeug (guest, #10998) [Link] (6 responses)

The problem is, at least in the US, they don't keep all those books. A large number of them, but books that don't meet their needs are disposed of. For example, GURPS is a series by Steven Jackson Games with 200 or so entries; searching for that in the Library of Congress catalog only turns up 7 entries, with 42 books shoved into one entry.

They find a copy and destroy it, in part because the easiest way is to cut off the spine and put the pages into a loose-leaf scanner. There are non-destructive ways to scan some books, but they're slower and more expensive, and just don't work for certain books that are tightly bound.

It doesn't mean an AI company rounds up all copies. But books disappear by attrition. For a print run in the three or four digit range, how many copies destroyed by fire or mold or just thrown out do you need before an AI company or two buying copies and destroying them destroys the last copy, or destroys the last copy publicly available? I assume that any AI company has already used Anna's Archive as a basis, which includes everything the Internet Archive scanned and let people check out, and millions of other volumes, meaning anything they're scanning is not well archived.

AI companies destroying books

Posted Sep 2, 2026 15:45 UTC (Wed) by paulj (subscriber, #341) [Link]

Particulars like this are issues raised by the rise of the AI industry that could (and should) be addressed by society imposing requirements on that industry through regulations.

The fine details and the particular problems, such as these, are very much worthy of discussion.

AI companies destroying books

Posted Sep 3, 2026 8:21 UTC (Thu) by taladar (subscriber, #68407) [Link] (4 responses)

Sounds to me like anyone (AI company or otherwise) who first turns a physical copy of a book into a digital copy should be required to deliver a digital copy to some archives (ideally not just that one library of congress in the US, that seems like a single point of failure, just one defunding away from disaster). But wait, then all the authors and publishers are probably going to cry "theft" again.

AI companies destroying books

Posted Sep 3, 2026 10:05 UTC (Thu) by kleptog (subscriber, #1183) [Link] (3 responses)

From what I understand, it's in the "fair-use" doctrine in the US that allows this destructive scanning in the first place. The act of scanning itself it probably illegal in Europe. So they're buying books in Europe and shipping them to the US to be scanned under the more relaxed laws there. But even in the US, companies putting their scanned books in an archive for public benefit probably kills their "fair-use" defense.

In Europe, museums, libraries and such are allowed to scan their own archives for the public benefit, but the AI companies can't use that exemption. It would arguably be in the public interest to find a way for the scanned book of the AI companies to find its way to archival institutions but that would require a law change (Europe is largely Civil Law, lawsuits won't help you here).

The irony is that there are plenty of institutions that would love to collect every book and digitally archive it and would be legally allowed to do so, but they never have the funding. And the AI companies have money to burn but aren't allowed to do the archiving.

AI companies destroying books

Posted Sep 4, 2026 7:26 UTC (Fri) by taladar (subscriber, #68407) [Link] (2 responses)

One more data point for the argument that Copyright actually leads to cultural destruction (abandonware being the best known one on that side).

AI companies destroying books

Posted Sep 4, 2026 12:18 UTC (Fri) by dskoll (subscriber, #1630) [Link] (1 responses)

That can be fixed by limiting copyright terms to something reasonable like 25-30 years instead of the current (in the USA) life of author plus 70 years, or 95 years for a work made for hire.

Straying

Posted Sep 4, 2026 14:05 UTC (Fri) by jzb (editor, #7867) [Link]

As much as I support the idea of "fixing" copyright and capitalism... we've probably strayed past the "on-topic" line at this point, and certainly past the scope of the discussion. There's no path for Debian to do either of those things through a GR, sadly.

Can we end this here and return to regularly scheduled programming?

Not totally unconvinced

Posted Aug 28, 2026 9:38 UTC (Fri) by peb (subscriber, #123687) [Link]

However, Bécue remained unconvinced. He replied to Ts'o that his calculator comparison ""has some merit, but is not a good one anyway"". He worried that people would turn to LLMs without having first developed their ability to think critically about the problems, which left them ""both vulnerable to a rug pull, but also to be fed garbage without ever realizing it"".

Just to say that I acknowledge that the fears I have might prove untrue, I'm merely erring on the cautious side of the thing.


Copyright © 2026, Eklektix, Inc.
This article may be redistributed under the terms of the Creative Commons CC BY-SA 4.0 license
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds