8

There's a lot of hubub recently about ML models memorizing their training data, and the copyright implications of that.

How did this used to work before fancy technology was brought into it? If I buy a book of poetry at the store, I have one copy. If I memorize it, now there physically must be a copy in my head, right? I can sell the book and still have the poems. If I go around teaching people all the poems until they too have memorized them (but not in the context of a public performance), aren't I committing copyright infringement and making a bunch of unauthorized copies in people's heads?

1
  • Comments have been moved to chat; please do not continue the discussion here. Before posting a comment below this one, please review the purposes of comments. Comments that do not request clarification or suggest improvements usually belong as an answer, on Law Meta, or in Law Chat. Comments continuing discussion may be removed.
    – Dale M
    Jan 26 at 9:18

2 Answers 2

17

Memorizing a work is not copyright infringement.

Infringement requires doing something that is an exclusive right of the copyright owner.

One of the exclusive rights of a copyright owner is the right to make a "copy" or "reproduction." This concept is defined to various degrees of precision, depending on the jurisdiction, but for one example, in the U.S., "copies" are defined as:

material objects, other than phonorecords, in which a work is fixed by any method now known or later developed, and from which the work can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device.

Many jurisdictions, and the Berne Convention itself, do not define "reproduction" or "copy." But the WIPO Guide to the Copyright and Related Rights Treaties Administered by WIPO and Glossary of Copyright and Related Rights Terms (2003) at 55 says:

The text of the Berne Convention does not contain any complete and explicit definition of “reproduction”. Certain elements of the concept of reproduction may, however, be identified in it. A good example is the clarification offered in Article 9(3) [sound or visual recordings deemed to be reproductions] . . . which makes it obvious that it is not a condition that, on the basis of the reproduction, the copy of the work be directly perceivable; it is sufficient if the reproduced work may be made perceivable through appropriate equipment.

Memorization is not one of the exclusive rights held by a copyright owner, and memorization does not result in a "copy" as defined in the U.S., nor in any jurisdiction's conception of a "reproduction."

18
  • 2
    I don't think it's as easy as in this answer. Digital copies on computer hard drives somehow do satisfy the definition, although the "physical" substance of the copy is rather intangible (the state of electromagnetic charge of the disk surface/cells). The copy in a human brain is intangible likewise. The principal difference is that brains can't be read as easy.
    – Greendrake
    Jan 25 at 14:49
  • 6
    @Greendrake The brain is not a physical object from which memories can be perceived, reproduced, or communicated by anyone other than the owner. Transcribing a book into a unique language unknown to anyone but the person who invented is not generating a copy, nor is creating a photocopy of a book which turns out to be illegibly smudged. A physical object is not a copy if no person or machine can perceive that it is, in fact, a copy. It's not the "intangibility" of bits or memories that differentiate the copyright issue, it's the ability to decode them. Jan 25 at 15:24
  • 3
    @NuclearHoagie That seems quite incorrect. Are you getting this from a legal authority, or is it mostly just a guess?
    – bdb484
    Jan 25 at 16:28
  • 5
    @bdb484 9th Circuit Court: "the material object must, in some manner, take on the physical aspects of the protected work such that the “copy” of the work may be perceived by an observer". My examples are cases where it is impossible for the "work" to perceived as such by an observer. I don't see how you could be guilty of copyright infringement if no one can perceive that your work is, in fact, a copy of another. Jan 25 at 16:46
  • 4
    I agree with that, but I don't see it as supporting any of your previous statements. If you're unsuccessful in making a copy, you haven't made a copy, even if the whole world knows you tried. If you're successful in making a copy (even in a form that others don't recognize), you've made a copy, even if no one knows.
    – bdb484
    Jan 25 at 18:52
-6

The issue with AI training is they didn’t buy the book of poetry

If you change your analogy to printing out a book of poetry you scraped from the internet and then memorising it, it is the print out that is the copyright infringement, not the memorisation.

6
  • 2
    An additional issue is that the copy/derivative work which was created, the AI model, to create further derivative works when the AI model is used to create something else (e.g., the response from ChatGPT). For LLMs, it's been shown that they can, under some conditions, recreate the text that was used to train them, or substantial portions of that text. That latter fact makes it quite clear that the AI model is, in fact, a derivative work "from which the work can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device."
    – Makyen
    Jan 25 at 23:10
  • 7
    I don't know if the headline is really a useful analogy. Anyone can go to a library and start memorizing books without buying anything. The legal problem, both for the human who memorizes physical library books and the AI trained on the digital library which is the internet, is in the later reproduction of the memorized/trained material. IMO, whether or not the human or AI trainer owns a copy of the original may not be particularly relevant.
    – josh3736
    Jan 25 at 23:54
  • 6
    I believe that there certainly are at least some cases where a model is trained on a legitimately obtained copy of some works (i.e. that they did "buy the book of poetry"), but without a permission or license to make further copies, and the question should be treated in the context of this "cleaner" scenario, where you do not have an infringement already before the "memorization" is even attempted.
    – Peteris
    Jan 26 at 0:03
  • 4
    Downvoting. The fact that they didn't buy it is completely irrelevant to the question; buying it wouldn't give them a right to copy it. Jan 26 at 10:05
  • 1
    @josh3736 AFAIK Internet libraries are illegal (unless authorized by the copyright holder) - the right to lend a copy only applies to physical copies (for some reason).
    – user253751
    Jan 26 at 15:05

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .