Login
You're viewing the functional.cafe public feed.
  • Sep 13, 2026, 9:19 AM

    I am not a lawyer. I have studied IP law in various forms guided by solicitors, barristers, and professors of law of my acquaintance since I was a teenager (including reading through big piles of case histories and commentaries) but not in a formal setting (mostly I learned that I would rather invent the things in the patents than draft the patents, so changed career direction). I’ve also spent a surprising amount of my career talking to copyright and patent lawyers (almost never trademark or trade-secret specialists). I have enough of a lay-person’s understanding of the topic that I was able to spot that a clause in the contract from my US publisher was unenforceable in the state that they claimed jurisdiction because it hinged on an aspect of copyright law that the USA delegates to states and which the state in question did not have relevant laws. Their lawyers subsequently confirmed and fixed this. But I a not lawyer and this is not legal advice.

    I have two objections to the EFF’s position on ‘AI’ model training. One is technical, one is social.

    The technical one first.

    Imagine I rip a DVD and transcode it to MPEG-4 video. This is lossy recompression. The new copy is not identical to the original. It is a derived work. There was no transformative step. Specifically, losing fidelity of reproduction is not a transformative step.

    Video CODECs take advantage of redundancy. Simple CODECs build predictive patterns for redundancy in a single frame (for example, is this all one colour or a gradient? Store just that fact not every pixel). More complex ones look at the previous frame and compare it to the current one and use that redundancy. Most modern ones do this in both directions. Effectively, they create a cube of voxels, where one dimension is time, and try to find redundancy in the cube.

    For copyright law, this doesn’t matter. Lossy compression is not a transformative step, no matter how complex the compression.

    A lot of compression schemes (rarely for video, mostly because it doesn’t make sense for video unless you have a lot) also support special cases for large quantities of redundancy across a data set. For example, if you wanted to compress English Wikipedia with ZSTD, you would use the dictionary mode. It will come up with a list of the words (or even common phrases such as ‘citation needed’) and Huffman encode them so that each page has a short encoding for referencing them. This makes each page smaller than it would be if you compressed it individually.

    Compressing Wikipedia like this is not a transformative step.

    Now, imagine that you create a video compression CODEC that does this. You buy a copy of every DVD or BluRay disk available and compress them together such that you have a large dictionary of all common compressed sequences. Given a prefix of a film, each film in the input set would be reproducible with some loss of quality (not necessarily the same level of quality). Similarly, if you provided an initial vector that was not something in the training set then you’d get out video that might be similar to one of the inputs, might be similar to many, or might not be obvious to a human is close to either.

    This is the crux of the argument. Deep neural networks are functionally equivalent to lossy compression schemes. The inference or generation step in ‘generative AI’ is an initial vector and a random seed that decompresses the data that might be there. If nothing from the training (input) set exactly matches (or if the random seed moves away from that path) then you’ll get something new, possibly something that’s recognisable as a lossily compressed version of the input data.

    Note, in particular, that a lot of compression schemes now do take advantage of neural networks. They are one of the most efficient known ways of generating a specialised lossy compression scheme over arbitrary data. The law typically doesn’t care what specific technology an action uses, only about the outcome. In this case, that doesn’t matter: exactly the same underlying technology, used in exactly the same way, covers both ‘AI’ and compression. If one is legal then so is the other because they are the same process.

    The EFF’s argument hinges on the idea that this lossy compression is a transformative step. Not only is that an idea that is not supported in case or statute law, there is case law that makes it clear that lossy compression of a work is not transformative.

    Their argument would be internally self consistent if it also argued that Netflix does not owe royalties on any of the third-party videos it streams (and that they can buy BluRays on Amazon, recompress them, and then stream them without paying royalties). But there is so much case and statute law that this is not the case that they didn’t make this claim.

    Instead, they tried to claim that this is permitted if you call the system ‘AI’ even though it is settled law that it is not permitted if you do not call the system ‘AI’.

    Second, the social aspect. Copyright law in the USA draws its legitimacy from this line in the Constitution:

    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.

    So does patent law. Patent law is more of a mess because the USA was founded just before James Watt bribed various MPs to subvert the patent system, but (due to later treaties) the modern US patent system inherits from that subversion (Watt, like Edison, was a bit of a dick).

    This intent goes back to the invention of the printing press, where publishers made copies of books in large quantities without paying authors. This removed the incentive to write books. Allowing authors to control distribution rights put that incentive back. This short paragraph covers about a hundred years of the evolution of legal thought in this space, please forgive the many oversimplifications.

    The EFF brief references this motivation but twists it. OpenAI and Anthropic are the equivalents of the post-Gutenberg printers. They are taking the work of creative individuals and producing output that competes directly with the products of those authors. This is precisely the situation that copyright law in the USA exists to prevent.

    Yet the brief twists this to say that a lossily compressed duplication of original work is actually the kind of creativity that this was intended to cover.

    This hinges on the idea that writing news, or other creative works, is a trivial commodity, whereas mechanically compressing them into a system that can lossily reproduce them is a key contribution to society.

    Even if I agreed with their other arguments (I do not, I believe that they either misunderstand or deliberately misrepresent the technology and the relationship to other settled law), this is such a profoundly anti-human viewpoint. The idea that human creativity exists to feed poor-quality technological reproductions of that creativity is incompatible with any possible society that I would want to live in and I would struggle to find common ground with people who aim to create such a world.

    💬 26🔄 533⭐ 674

Replies

  • Sep 13, 2026, 9:40 AM

    Don’t you hate it when you write a long post, proofread it, post it, see it in your feed, and immediately spot a typo in the first line?

    (Fixed now, it said ‘firms’ when it should have said ‘forms’)

    💬 11🔄 6⭐ 119
  • 💬 1🔄 9⭐ 56
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 2
  • 💬 0🔄 0⭐ 4
  • 💬 0🔄 0⭐ 0
  • Sep 13, 2026, 9:58 AM

    @david_chisnall Or worse, as you try to delete the offending letter causing the typo focus moves for reasons known only to the universe to another part of the page and the backspace key causes the web browser to navigate to the previous page.

    Even worse when this happens as you have nearly completed your long and carefully crafted toot.

    I am going to start to using Mousepad to write my toots then copy-pasting them into the compose pane.

    💬 3🔄 0⭐ 9
  • Sep 13, 2026, 10:00 AM

    @david_chisnall If I ever get my hands on a time machine I will go back in time to find the person/people/team who thought that using the "Backspace" key to navigate back was a good idea and persuade them to change career before they make that fateful decision.

    This feature is the bane of my internet life.

    💬 0🔄 0⭐ 8
  • 💬 1🔄 1⭐ 8
  • Sep 13, 2026, 1:13 PM

    @david_chisnall A nice feature, although potentially opens you up to other people hitting the back button to see what you have been typing and where?

    It would be better to just have a way to disable the "Backspace to go back" feature.

    💬 1🔄 0⭐ 2
  • Sep 13, 2026, 2:07 PM

    @the_wub @david_chisnall I don't know about other browsers, but that feature exists in Firefox:

    • open about:config
    • search for browser.backspace_action
    • set it to 2

    More specifically, the value 0 makes backspace navigate back in history (shift-backspace goes forward), 1 makes backspace scroll up (shift-backspace scrolls down), and 2 disables it.

    💬 1🔄 4⭐ 7
  • Sep 13, 2026, 3:18 PM

    @barubary @david_chisnall Interesting!

    LibreWolf, it is already set to "2".

    PaleMoon, it is already set to "2" .

    SeaMonkey, it is already set to "2".

    Firefox ESR, it is already set to "2".

    I will have to look further/post in the forums to see if there is an equivalent option in Vivaldi. I could not find anything obvious under Settings or vivaldi://flags .

    💬 3🔄 0⭐ 2
  • 💬 0🔄 1⭐ 4
  • 💬 1🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 1
  • Sep 13, 2026, 2:47 PM

    @the_wub @david_chisnall If the conversation is deep and serious enough I am not above composing in a text editor and only cutting-and-pasting a final product for tooting once I am satisfied with it. The conversation approaches that level.

    💬 0🔄 0⭐ 2
  • Nov 20, 2026, 12:46 AM

    @the_wub @david_chisnall this is a result (in part) of the misuse of event bubbling. Backspace (on a page not a text field) was wired to "previous" (this was a mistake). Backspace on a text field when there is no character to delete can bubble (this was a mistake), causing "previous". But bubbling itself, that's my invention. My friend took it to Netscape and implemented it there, alongside jwz.

    💬 1🔄 1⭐ 1
  • 💬 0🔄 1⭐ 1
  • 💬 0🔄 0⭐ 3
  • 💬 0🔄 0⭐ 0
  • Sep 13, 2026, 11:59 AM

    @david_chisnall
    Now imagine the embarrassment when the post is on a platform where it cannot be edited, while you are an editor, commenting on the topic of editing and grammar. (Yes, this has happened to me... more than once.🤦‍♀️)

    Even editors need an editor when we write.

    So don't feel bad at all.

    💬 1🔄 1⭐ 10
  • 💬 1🔄 9⭐ 14
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 1
  • 💬 0🔄 0⭐ 2
  • 💬 0🔄 0⭐ 1
  • Sep 14, 2026, 12:51 PM

    @david_chisnall the occasional typo here and there is now a badge of honor, an sign to me and everyone else that the author has blood coursing through their veins, that the words are not cheap and were punched by flesh

    💬 0🔄 3⭐ 3
  • 💬 1🔄 0⭐ 8
  • Sep 13, 2026, 9:54 AM

    @bjn There’s a lot of steps but the short version is that he had some patents on key parts of steam engines and, every time they were about to expire, bribed politicians to extend them, much as Disney did with copyright in the USA. Some of these extended the terms of all patents, others special cases Watt’s patents.

    This probably both slowed down the Industrial Revolution and ensured a specific concentration of wealth during the process.

    It’s a case study I wish was required reading for MPs entering parliament because they keep repeating the same model, often through ignorance rather than corruption (though sometimes both).

    EDIT: He also had a habit of using the patents to push competitors out of business and then hire buy their assets and cheaply and hire and underpay their assets. He was probably an inspiration for Edison and Gates.

    💬 0🔄 4⭐ 32
  • Sep 13, 2026, 10:02 AM

    @david_chisnall I feel conflicted about the technical argument: lossy compression is not a binary but a spectrum. The amount of loss makes the result vary smoothly from "perfect reproduction" to "completely unrecognizable random corruption". IMO, law needs stable categories to determine outcomes, and when none are present it needs to create them, so I wouldn't be surprised if "gzip" landed on the opposite side than "Claude".

    💬 3🔄 1⭐ 9
  • Sep 13, 2026, 10:05 AM

    @omixann

    I’d be happy with an argument (which would be consistent with case law) that said:

    If the outputs of a machine-learning system can be something that would meet the existing tests for being a derived work of one of the inputs, then the machine-learning system is a derived work of the training data.

    That would be entirely consistent with existing law and would make OpenAI and Anthropic liable for a few trillion dollars of wilful copyright infringement.

    💬 2🔄 10⭐ 30
  • Sep 13, 2026, 10:12 AM

    @omixann

    To add a bit more (sorry for the double reply):

    The EFF brief did not try to capture any of this nuance. It could have proposed tests for where on that spectrum you are. It could have argued that the interpolation between different works is a transformative step (though there is case law that says this is not the case if you don’t call the interpolation ‘AI’) if done with a sufficiently large data set. It could have made arguments about the size of the models versus the size of the input (note: the weights of ChatGPT are larger than a losslessly compressed copy of Wikipedia and Project Gutenberg, and compression ratios generally improve the more data you have).

    It didn’t do any of these things and instead made specious arguments about ‘AI’ being special and magic and totally different from almost identical applications of exactly the same technology.

    💬 1🔄 7⭐ 25
  • Sep 13, 2026, 10:27 AM

    @david_chisnall once again I hope my message wasn't understood as an specific endorsement of what the EFF said, but more of a reflection on the nature of LLMs and their interactions with law and society in general. I think they have their reasons to be dismissive on any support for copyright (given its current shape vs. what it was meant to be historically), but I'm more than willing to accept that's preventing them from making a better argument than a knee-jerk reaction.

    💬 0🔄 0⭐ 7
  • Sep 13, 2026, 10:12 AM

    @david_chisnall unfortunately I'm not in a position to support or contest this, so assuming your verdict is correct ("even setting a reasonable threshold, OpenAI and Anthropic must be violating it") I'd agree on the outcomes.

    Not like I'm rooting for these two anyway, like, fuck them, I could probably trace most if not all of my grievances with LLMs back to these two.

    💬 1🔄 0⭐ 6
  • Sep 13, 2026, 10:18 AM

    @omixann

    In the case for which this brief was filed, NYT’s initial filing showed that the chatbots would regurgitate almost word-for-word copied of NYT articles. So the test of ‘is enough of the input preserved in the model to appear in the output’ has already been met. The definition of a derived work isn’t binary, but there’s plenty of precedent that tweaking the occasional adjective is not sufficient to stop something being a derived work, so it’s easy for a court to conclude that the output is a derived work of the input.

    The question then is if A is fed to B and B produces C, and C meets the legal definition of being a derived work of A, then is B also a derived work of B? And, if so, does it qualify for fair-use protections?

    I’m not a fan of NYT or Disney either, but they’re on the correct side in this specific dispute.

    To be honest, I’d be moderately happy with a ruling that says the model itself is not a derived work of the input but any output might be. That would make the legal risk of using the output in any commercial setting unacceptable for any company.

    💬 1🔄 5⭐ 20
  • Sep 13, 2026, 10:23 AM

    @david_chisnall

    > In the case for which this brief was filed, NYT’s initial filing showed that the chatbots would regurgitate almost word-for-word copied of NYT articles.

    I won't argue against that, and I hope my first message wasn't an endorsement of that.

    > To be honest, I’d be moderately happy with a ruling that says the model itself is not a derived work of the input but any output might be.

    Oh, I actually like this one.

    💬 0🔄 0⭐ 7
  • Sep 13, 2026, 10:06 AM

    @david_chisnall your social objection resonates much more with me (and it's the kind of advocacy I've been promoting so far among peers wrt. AI): no matter what reasonable arguments can be made about the nature of LLMs, their harms to society are factual. Much like with the printing press, the technology just clashed with how the world around it worked, and it damaged the very people it fed from. (cont)

    💬 1🔄 1⭐ 7
  • Sep 13, 2026, 10:08 AM

    @david_chisnall Hence why copyright was created: an artificial valve to let genuine creativity be profitable while leveraging the advantages of the new technology. In a similar vein, I think trying to shoehorn LLMs into our existing paradigm of "copyright" and "transformative work" is a fruitless endeavor, it just won't fit. We need either an amendment for copyright that answers to LLMs' impact on society and needs no justification, or the replacement of copyright with some different framework.

    💬 1🔄 0⭐ 3
  • Sep 13, 2026, 11:46 AM

    @omixann @david_chisnall
    No, we don't.
    It's really just a wonky retrieval system using statistics to build an output from the sources.

    It's copyright violation even though it's using a massive amount of compute that's damaging the environment. They don't want to pay for the content because already the systems can't make money.

    💬 1🔄 1⭐ 5
  • Sep 13, 2026, 12:16 PM

    @raymaccarthy @david_chisnall I need you to elaborate on your "wonky retrieval system" definition given how my argument hinges on the fact that lossy compression lies in a spectrum between perfect reproduction and perfect noise.

    I won't get into the "environment impact" of LLMs because that's tangential to the technology itself, and says more about who is profiting from them than about LLMs themselves (and also excludes the advancements in local models outside the Silicon Valley hype train).

    💬 1🔄 0⭐ 2
  • Sep 13, 2026, 2:29 PM

    @omixann @david_chisnall
    "my argument hinges on the fact that lossy compression lies in a spectrum between perfect reproduction and perfect noise."
    Lossy Compression isn't on such a spectrum.
    1. Never gives perfect reproduction.
    2. The compressed information, though lossy, is never noise, but always recognisable.
    All lossy audio, video & image compression.
    There is a special one for scanned text (initially images). It creates a table of image glyphs rather than doing good OCR, e.g. Topaz.

    💬 0🔄 0⭐ 0
  • Sep 13, 2026, 3:33 PM

    @raymaccarthy @david_chisnall you're either missing the point or willingly refusing to engage with my argument. The continuum between a lossless copy of your data and nothingness _is_ what lossy encoding is! The reason why you can tweak quality settings when exporting something in a lossy format is precisely to choose where to land in that spectrum.

    💬 1🔄 0⭐ 2
  • Sep 13, 2026, 3:36 PM

    @raymaccarthy @david_chisnall Compression algorithms are made to preserve fidelity so they land closer to the lossless end. LLMs are not concerned with that, so they stray further away from it (which is obvious when comparing their compression ratios). Law cannot operate on such fuzzy boundaries so it needs to put labels to things, set boundaries, and the boundary can very well land between "JPG" and "Qwen", there is no technical impediment against it.

    💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Sep 13, 2026, 4:12 PM

    @omixann @david_chisnall
    That is not what lossy compression means.
    You are inventing a new definition.
    There is no quality setting on lossy compression that gives perfect or nothing. Only very good to poor.

    💬 1🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0
  • Sep 13, 2026, 4:47 PM

    @raymaccarthy @omixann

    By definition, lossy compression is imperfect but there’s a long history of combining lossy compression with a correction dictionary to turn it into lossless compression. Some of the oldest examples of this involve machine learning.

    One of my favourite examples is the hyphenation dictionary in LaTeX. This takes a large number of examples of words with correct hyphenation and builds short Markov chains that predict when words should be hyphenated. The second stage in the process then runs the predictor over every word in the input set and records the ones that are incorrectly predicted as a dictionary of special cases.

    When you do hyphenation it first looks at the special cases and, if not present there, it tries the ML model. This combination precisely compresses everything in the dictionary. The thousands of entries in the dictionary are compressed to a small ML model and around 70 outliers.

    If you feed it a word that wasn’t in the training set, it will usually get the hyphenation right. You can even run the model in reverse to generate plausible hyphenated words and many of them will be real words with correct hyphenation.

    The line between lossless compression, lossy compression, and generative machine-learning model is blurry.

    💬 1🔄 2⭐ 3
  • Sep 13, 2026, 4:57 PM

    @david_chisnall @omixann
    But any case you are making a copy, so the result can be shared or monetised, unless you have the rights.

    In any case only hyphenate for paper, not for ebook content. That will break search on the ereader and if hyphenation is used it's best done by the render for the current screen sise and user font size, margins, & line spacing..

    💬 1🔄 0⭐ 0
  • Sep 13, 2026, 5:00 PM

    @raymaccarthy @omixann

    Hyphenation lists aren’t creative works and so can’t be copyrighted, which means that the question arises. But the difference between this and an LLM is one of scale, not of kind. It consumes entire creative works in the training rather than individual words.

    If anything, there’s a stronger case that the simpler models are transformative creative works because they are created with a specific intent understood by the author, whereas LLMs are not.

    💬 1🔄 0⭐ 3
  • Sep 13, 2026, 5:04 PM

    @david_chisnall @omixann
    I wasn't referring to the hyphenation list but the source text!

    An LLM can regurgitate more of the source than UK or Ireland Copyright would allow.

    A hyphenation tool doesn't do that.

    LLMs are automated plagiarism tools.

    💬 0🔄 0⭐ 2
  • 💬 2🔄 6⭐ 12
  • 💬 1🔄 0⭐ 2
  • Sep 13, 2026, 2:03 PM

    @omixann @david_chisnall not as far as I am aware unfortunately. I was hoping that some law folks would at some point develop these arguments a bit further but it hasn't happened yet. I'm still hopeful though, especially now that the compression perspective seems to be getting the interest of more mainstream people (e.g. 3b1b)

    💬 0🔄 1⭐ 1
  • 💬 1🔄 0⭐ 0
  • Sep 18, 2026, 9:49 PM

    @iampytest1 maybe that anonymous lawyer should write an article explaining what part of Balaji's argument they believe is incorrect

    💬 0🔄 1⭐ 0
  • Sep 13, 2026, 10:13 AM

    @david_chisnall Thus far US courts have found the LLM training technology and its nature irrelevant to whether it is transformative, the application is. In Bartz v. Anthropic and Kadrey v. Meta, they were found to be transformative. In the case of Thomson Reuters v. Ross Intelligence, where Ross basically trained a competitor of Westlaw using Westlaw summaries, it was judged to be a substitute and not transformative.

    💬 2🔄 0⭐ 4
  • 💬 1🔄 0⭐ 3
  • Sep 13, 2026, 3:25 PM

    @david_chisnall yes, so they are being consistent:

    eff.org/document/amicus-brief-

    Regarding Ross, specifically, there is also precedent that mere "sweat of the brow" collection of facts like phone numbers in the Yellow Pages is not copyrightable in the US, only "expressive content" like commentary would be:

    en.wikipedia.org/wiki/Feist_Pu.

    and in fact that is cited by the EFF brief.

    In any case the paywalling of law in the US, e.g. PACER, is a scandal:

    open-public-records.com/articl

    You are supposed to abide by laws (or jurisprudence, which in Common Law is the same thing), and can only access them by paying.

    💬 0🔄 1⭐ 0
  • 💬 1🔄 0⭐ 0