September 2026 - ghc-devs

Re: GHC LLM policy vs copyright
by Carter Schonwald 02 Sep '26

02 Sep '26

yes, because code is math, and sometimes theres only one proof, and copyright doesnt protect factual stuff just the ultra soecific exact sentence (except when its literally the only sane way to say soemthing). there is a huge amount of law / case law / history for how this works, and fortunately literally here means verbatim naively these two remarks sound contradictory, but its really a reflection of the anthropogenic complexity of humanity On Tue, Sep 1, 2026 at 1:56 PM <code(a)jaro.addy.io> wrote: > If you define “exact literal content” so broadly as to allow for > translations > of prose, why wouldn't it also apply to translating software to different > languages or using different naming conventions and other minor code style > elements? > > On 9/1/26 19:50, Carter Schonwald 'carter.schonwald at gmail.com' wrote: > > translations are explicitly reusing the exact literal content, and > theres > > explicitly carve outs in modern law around translated works > > > > > > math and food recipes are explicitly not protected by copyright law. > > > > i hope that helps clarify? > > > > On Tue, Sep 1, 2026 at 1:41 PM <code(a)jaro.addy.io <mailto: > code(a)jaro.addy.io>> > > wrote: > > > > > building dimilar code for a similar task in a > > > different language is definitely not infringing unless its also > copying > > all > > > the text from the source application like some sort of text rpg. > > What if we compare this to literature? Surely a translation of a > book into > > another language is still considered a derivative work, even if it > reuses none > > of the “literal text”. > > > > >

3 2

Re: GHC LLM policy vs copyright
by Carter Schonwald 02 Sep '26

02 Sep '26

No. Please do some background reading before making these bald assertions. wikipedia or any search engine. also depends on which countries laws. i can only speak confidently about US ip law. but international treaties kinda unify how copyright works, mostly. theres a thing called moral rights outside the us, i dont really understand it and perhaps thats what youre thinkjng of, idk theres absolutely no mysticism needed, but you need to find some background reading for thr regions law. law logic is not like formal methods. itd like deciding who gets hen in the disagreement based on applicable live statutes and or in english style court systems: any precedent in related disputes. but it can involve a dispute frkm 150 yrs ago ans how it was decided also pragmatically: every copyright dispute that holds any water tends to involve naked theft. show me these open source gpl/agpl/closed source but in training sets for coding that have capabilitirs that would work verbatim in ghc. i reallly reallly want to look at these compilers!! On Tue, Sep 1, 2026 at 4:07 PM <code(a)jaro.addy.io> wrote: > Re 1: I think it is quite a tall order to require a concrete example of > copyright infringement. > > I don't see how 2 & 3 are related to my messages in this thread. I did not > mention morality nor did I assume no laws were written as jokes or out of > spite. If you think I did, please explain your reasoning more clearly and > point > to specific parts of my responses. > > Also, if the law is really so complex, contradictory, and sometimes > written or > enforced as a joke or out of spite, shouldn't we be extra careful to avoid > even > a tiny risk of being sued? > > On 9/1/26 21:25, Carter Schonwald via ghc-devs 'ghc-devs at haskell.org' > wrote: > > 1) i want haskell native examples of the actual worrys, not abstract ones > > > > 2) people have literally written enforced laws as a joke or spite, > please do > > some supplementary reading to ground your conjecture in legal writings, > even > > from wikipedia > > > > 3) morals and law are not the same. llms do conflate thise all the time. > i > > totally beleive you have better moral values thanapplicable law, but > thats not > > law. > > > > > > On Tue, Sep 1, 2026 at 2:37 PM <code(a)jaro.addy.io <mailto: > code(a)jaro.addy.io>> > > wrote: > > > > It's true that sometimes there is only one way to implement > something and I > > know copyright does not cover such cases. However, I think that is > quite rare > > in software development, especially once we start looking at > combinations of > > two or three functions. > > > > > and fortunately literally here means verbatim > > > > > > naively these two remarks sound contradictory, but its really a > > reflection of > > > the anthropogenic complexity of humanity > > > > I remain utterly confused about what you mean with “literally” and > “verbatim”, > > and thus unconvinced by your arguments. Like, are you saying that > renaming all > > variables and functions and changing all for-loops to while-loops > (using C for > > this example) is a valid way to avoid copyright infringement because > it is no > > longer a literal or verbatim copy? I refuse to believe that. > > > > On 9/1/26 20:12, Carter Schonwald via ghc-devs 'ghc-devs at > haskell.org > > <http://haskell.org>' wrote: > > > yes, because code is math, and sometimes theres only one proof, > and > > copyright > > > doesnt protect factual stuff just the ultra soecific exact > sentence (except > > > when its literally the only sane way to say soemthing). > > > > > > there is a huge amount of law / case law / history for how this > works, and > > > fortunately literally here means verbatim > > > > > > naively these two remarks sound contradictory, but its really a > > reflection of > > > the anthropogenic complexity of humanity > > > > > > On Tue, Sep 1, 2026 at 1:56 PM <code(a)jaro.addy.io > > <mailto:code@jaro.addy.io> <mailto:code@jaro.addy.io > > <mailto:code@jaro.addy.io>>> > > > wrote: > > > > > > If you define “exact literal content” so broadly as to allow > for > > translations > > > of prose, why wouldn't it also apply to translating software > to > > different > > > languages or using different naming conventions and other > minor code > > style > > > elements? > > > > > > On 9/1/26 19:50, Carter Schonwald 'carter.schonwald at > gmail.com > > <http://gmail.com> <http:// > > > gmail.com <http://gmail.com>>' wrote: > > > > translations are explicitly reusing the exact literal > content, > > and theres > > > > explicitly carve outs in modern law around translated works > > > > > > > > > > > > math and food recipes are explicitly not protected by > copyright law. > > > > > > > > i hope that helps clarify? > > > > > > > > On Tue, Sep 1, 2026 at 1:41 PM <code(a)jaro.addy.io > > <mailto:code@jaro.addy.io> > > > <mailto:code@jaro.addy.io <mailto:code@jaro.addy.io>> > > <mailto:code@jaro.addy.io <mailto:code@jaro.addy.io> > > > <mailto:code@jaro.addy.io <mailto:code@jaro.addy.io>>>> > > > > wrote: > > > > > > > > > building dimilar code for a similar task in a > > > > > different language is definitely not infringing > unless > > its also > > > copying > > > > all > > > > > the text from the source application like some sort > of > > text rpg. > > > > What if we compare this to literature? Surely a > translation of a > > > book into > > > > another language is still considered a derivative > work, even > > if it > > > reuses none > > > > of the “literal text”. > > > > > > > > > > > > > > > > >

2 1

Re: GHC LLM policy vs copyright
by Carter Schonwald 01 Sep '26

01 Sep '26

1) i want haskell native examples of the actual worrys, not abstract ones 2) people have literally written enforced laws as a joke or spite, please do some supplementary reading to ground your conjecture in legal writings, even from wikipedia 3) morals and law are not the same. llms do conflate thise all the time. i totally beleive you have better moral values thanapplicable law, but thats not law. On Tue, Sep 1, 2026 at 2:37 PM <code(a)jaro.addy.io> wrote: > It's true that sometimes there is only one way to implement something and > I > know copyright does not cover such cases. However, I think that is quite > rare > in software development, especially once we start looking at combinations > of > two or three functions. > > > and fortunately literally here means verbatim > > > > naively these two remarks sound contradictory, but its really a > reflection of > > the anthropogenic complexity of humanity > > I remain utterly confused about what you mean with “literally” and > “verbatim”, > and thus unconvinced by your arguments. Like, are you saying that renaming > all > variables and functions and changing all for-loops to while-loops (using C > for > this example) is a valid way to avoid copyright infringement because it is > no > longer a literal or verbatim copy? I refuse to believe that. > > On 9/1/26 20:12, Carter Schonwald via ghc-devs 'ghc-devs at haskell.org' > wrote: > > yes, because code is math, and sometimes theres only one proof, and > copyright > > doesnt protect factual stuff just the ultra soecific exact sentence > (except > > when its literally the only sane way to say soemthing). > > > > there is a huge amount of law / case law / history for how this works, > and > > fortunately literally here means verbatim > > > > naively these two remarks sound contradictory, but its really a > reflection of > > the anthropogenic complexity of humanity > > > > On Tue, Sep 1, 2026 at 1:56 PM <code(a)jaro.addy.io <mailto: > code(a)jaro.addy.io>> > > wrote: > > > > If you define “exact literal content” so broadly as to allow for > translations > > of prose, why wouldn't it also apply to translating software to > different > > languages or using different naming conventions and other minor code > style > > elements? > > > > On 9/1/26 19:50, Carter Schonwald 'carter.schonwald at gmail.com > <http:// > > gmail.com>' wrote: > > > translations are explicitly reusing the exact literal content, > and theres > > > explicitly carve outs in modern law around translated works > > > > > > > > > math and food recipes are explicitly not protected by copyright > law. > > > > > > i hope that helps clarify? > > > > > > On Tue, Sep 1, 2026 at 1:41 PM <code(a)jaro.addy.io > > <mailto:code@jaro.addy.io> <mailto:code@jaro.addy.io > > <mailto:code@jaro.addy.io>>> > > > wrote: > > > > > > > building dimilar code for a similar task in a > > > > different language is definitely not infringing unless > its also > > copying > > > all > > > > the text from the source application like some sort of > text rpg. > > > What if we compare this to literature? Surely a translation > of a > > book into > > > another language is still considered a derivative work, even > if it > > reuses none > > > of the “literal text”. > > > > > > > > > >

2 1

Re: GHC LLM policy vs copyright
by code＠jaro.addy.io 01 Sep '26

01 Sep '26

Also, we can consider a case much closer to software: reverse engineering. Sometimes reverse engineering is done through the elaborate clean-room process. Why would people do that if just changing the language (so not using assembly, I guess) would be enough to avoid copyright problems? On 9/1/26 19:41, Jaro Reinders 'j.s.reinders at posteo.net' wrote: >> building dimilar code for a similar task in a >> different language is definitely not infringing unless its also copying all >> the text from the source application like some sort of text rpg. > What if we compare this to literature? Surely a translation of a book into > another language is still considered a derivative work, even if it reuses none > of the “literal text”. > >

2 1

Re: GHC LLM policy vs copyright
by Carter Schonwald 01 Sep '26

01 Sep '26

translations are explicitly reusing the exact literal content, and theres explicitly carve outs in modern law around translated works math and food recipes are explicitly not protected by copyright law. i hope that helps clarify? On Tue, Sep 1, 2026 at 1:41 PM <code(a)jaro.addy.io> wrote: > > building dimilar code for a similar task in a > > different language is definitely not infringing unless its also copying > all > > the text from the source application like some sort of text rpg. > What if we compare this to literature? Surely a translation of a book into > another language is still considered a derivative work, even if it reuses > none > of the “literal text”. > >

2 1

Re: GHC LLM policy vs copyright
by Carter Schonwald 01 Sep '26

01 Sep '26

fortunately i am learned a clear analogue of your hypothetical is: fantasy swords and their nomenclature. describing a sword and reference common knowledge about decorative elements, materials, geoemtry and fabrication copyright applies to literal text of a creative work. and oracle v google makes this interp sound for software. building dimilar code for a similar task in a different language is definitely not infringing unless its also copying all the text from the source application like some sort of text rpg. in technical domains, the word compiler means compiler and its used in more than one project without issue. i really want to know what other prjects have zonking, knot typing , the same coding styke etc. its a genuinely complex idiosyncratic code base! and no auto regressive model will write and valdiate a plausible patch without reading / loading dome files, which by the power of autoregression will look like ghc flavored haskell. can someone actually try to manufacture a synthetic instance of a llm trace that verbatime copies gpl code from dome other haskell project? i cant see thst sneaking in given auto regression and haskell corpus. itd be very interesting to be incorrect, and it would help ground this dialogue. cauee its easier to prove an exists rather than forall On Tue, Sep 1, 2026 at 12:33 PM <code(a)jaro.addy.io> wrote: > You seem to assume that cross-language copyright infringement is > impossible and > that copyright infringement must involve either very local things like > naming > conventions or global things like architecture. > > I am not learned in software IP law, but the above assumptions seem > plainly > false from the little I do know. > > It could just be a few functions from another project, possibly in a > different > programming language, adapted to Haskell and the GHC naming conventions. > Also, > there are plenty of compilers or just general libraries written in Haskell > which contain code that could also be used in GHC. > > The fact that there is little comparable code in the training set could > also > mean the LLM is more likely to copy a significant part of a single project > as > there is less code to “learn” from. > > Cheers, Jaro > > On 9/1/26 17:26, Carter Schonwald via ghc-devs 'ghc-devs at haskell.org' > wrote: > > this copyright thread confuses me, because …. what Haskell code thats > public > > could these model echo into a ghc edit that isnt gonna be basically ghc > code? > > the architecture and naming conventions and code style is quite > distinct and > > theres very few salient hs code bases that would actually be salient, > and > > translating an algorithm or data structure from one programming language > does > > not create copyright risk wrt the translated code except to the extent > string > > literals or specific data is copyrighted… > > > > i would love examples that ground the ip risk with > > > > now if we were talking abiut a bsd open source nosql db or web server, > there’s > > would be lots to worry about agpl / gpl/ that commercial one that > converts to > > apache after k years. in the case of ghc the very style and architecture > just > > creates a risk of a model generating old no longer correct ancient ghc / > > haskell code! > > > > caveats: my family is basically all software ip lawyers and ive had to > navigate > > quite complex ip nuances in my work over time. > > > > On Sun, Aug 23, 2026 at 2:34 AM andrew.lelechenko--- via ghc-devs <ghc- > > devs(a)haskell.org <mailto:ghc-devs@haskell.org>> wrote: > > > > Hi Tom, > > > > > don't these two adjacent sentences contradict one another? > > > > I don't see how. Is there "an immediate or visible danger"? No. Is > there a > > way to eliminate any potential danger around LLM usage? No, "until > the > > matter is actually tested by courts". Is it worth to spend five > digits on > > some sort of "middle way", which might alleviate but not mitigate a > > potential but unspecified danger? Also no. > > > > Best regards, > > Andrew > > _______________________________________________ > > ghc-devs mailing list -- ghc-devs(a)haskell.org <mailto: > ghc-devs(a)haskell.org> > > To unsubscribe send an email to ghc-devs-leave(a)haskell.org <mailto: > ghc- > > devs-leave(a)haskell.org> > > > > >

2 1

GHC LLM policy vs copyright
by damiancoding＠gmail.com 01 Sep '26

01 Sep '26

Dear all, I have thus far been perusing the LLM policy discussion from the bleachers. A lot of conversation has been had about the ethical and practical reasons to support or abolish LLM use for GHC. However, one really important point seems to have been all but omitted from the conversation, and that is the legal aspect. GHC, like any other project using any licensing at all, enjoys protections based in the legal system of every country it is being used and developed in. That is no small thing and is essentially what allows GHC to exist in the legal sense. A body of work untethered from this is like a person without citizenship - no one will advocate for it, and no government will back up legitimate claims around it by enforcing laws and regulations. It is the contention of some legal scholars that having a significant (not large or overwhelming, merely significant) amount of AI based contributions in a body of work can lead to lack of recognition of its copyright in court. This opinion is already backed up by case law relating to art: https://www.theverge.com/policy/887678/supreme-court-ai-art-copyright Such a status would put the project's protections in danger legally and could turn to foil the efforts of GHC contributors over the past numerous years. Of course, art is not code, and I would say many people's code is not art... but absence of proof is not proof of absence. This is merely legal limbo, not legal certainty, and early indications suggest the tree may fall on the house. In fact, who is to say that at least in some jurisdictions, AI companies couldn't go rogue and turn around and say anything that's been generated does not belong to you intellectually - it belongs to the model, and therefore to its owner, the AI company? Of course, there are explainers and assurances, but none of this is legally tested. As someone who has advised on, joined, and started lawsuits in more than a handful different legal systems, I can tell you that they are all very idiosyncratic and there is no common-sense to any of them. Local law can be extremely surprising and in fact one place's laws can contradict another place's laws. There are also secondary effects as well. For example, can you get public funding for such a project? Can you be recommended by government entities and various organizations? Can you use or contribute to such work as a university employee, researcher, or student? This last one is particularly important for a project that lives and dies by the ability of researchers to contribute to it. I may have simply missed it in the deluge, but I have found that there has been little to no discussion of the effect of LLM contributions on the copyright of GHC. Perhaps before entering this "Brave New World", we should see if the ice is too thin to walk on. Especially during war time, there has been a multitude of legal documents that were proof of citizenship to one area or another. Sometimes they looked like the blue British passport that is all the rage these days, sometimes they were a tattered piece of paper in an envelope of you were lucky, with nothing more to certify its authenticity than a stamp, and a block-letter name written by someone who barely ever held a pen, at an outpost in the deep forest. This may have been all that a person, coming from a country that may not even exist anymore due to changes in war, has to go on for legal recognition as a citizen of any country. Many times such documents were ad-hoc or provisional and there wasn't even a way to check them for authenticity or correctness, if the checking was being done by a person a thousand miles away from the document's origin and who is not even able to read the alphabet let alone the language. Very often such documents would not be recognized; it was as much dependent on the politics of the day (does the Soviet Union recognize the Croatians?) as on whether you had a pack of cigarettes to share with the militia man patrolling the path you were on. A solid passport from a powerful country got you out of trouble. The legal system around LLM contributions is changing. Very little of it is proven and what has been proven so far shows that they enjoy no protection. For better or worse whatever imaginary lines on the ground we imagine and agree upon within our little tribe will be completely disregarded and obliterated by the machine of war that sets out the order as the legal system solidifies. To understand how this can lead to disappointment, just ask people in Finland whether they are nostalgic for their lost territory. I believe the integration of LLM based output into GHC is an issue that requires prudence far beyond "community consensus", "robust conversations", and "good proposals". It can clearly have ramifications of legal, compliance, civic, policy, and otherwise government-related nature, should be treated seriously as such, and therefore the project should source informed analysis and recommendation from those whose job it is to do so. This in particular means we should be hiring and listening to impartial professionals and not sounding off as developers who don't know the first thing about such matters. I say all this as someone who uses AI on the daily and has worked at a frontier AI lab that has been acquired by interests related to the US government. I think GHC is more important than being able to work on it with specific techniques. Everyone employing asbestos had the best intentions and could never have known where it would lead, but out of this period in history we have learned to be careful about employing new technology. Currently, the only people talking about this on this list are asbestos users and carpenters, and no one's thought to ask the doctors or physicists. With best regards, Damian

14 45

Re: LLM policy for GHC
by Teo Camarasu 01 Sep '26

01 Sep '26

Comparisons between LLMs and calculators were made earlier in this thread. That reminded me of this paper https://arxiv.org/abs/2507.19960 which explores these sorts of comparisons. I find that it provides a helpful framework for analysing the relationship between LLMs and cognition. Cheers, Teo

5 4

Re: GHC LLM policy take 2
by Tom Ellis 01 Sep '26

01 Sep '26

On Wed, Jul 29, 2026 at 12:42:45PM +0000, code(a)jaro.addy.io wrote: > > On 29 Jul 2026, at 14:24, Tom Ellis via ghc-devs <tom-lists-ghc-devs-2026(a)jaguarpaw.co.uk> wrote: > > Perhaps, or perhaps people use a better, but more expensive, model > > less? > > I expect most people to use a subscription like Claude Max or OpenAI > Pro, which encourages using as much as the plan allows. Subscriptions typically come with a number of credits that refresh every so often. More expensive models use more credits. Subscription use is not "unlimited". https://help.openai.com/en/articles/12642688-using-credits-for-flexible-usa… Tom

11 22