24 ms·

It's not a fair use question, it's a question about models reproducing the article text almost verbatim. (IANAL)

by solomatov 3y ago

brigadier132 3y ago
If it's fair use reproducing the article text verbatim is fine.
- andy99 3y ago
  What is "it"? Training can be fair use, i.e. updating weights incrementally based on predicted next token probabilities. And I (not a lawyer) think that if a broadly trained foundation model can recall some verbatim text, that doesn't mean the model is infringing. It seems like the lawsuit here is talking about specific NYT related functionality, like retrieving the text in new articles. That essentially has nothing to do with training and running a foundation LLM, it's about some specific shenanigans with NYT content, and it's legal status would appear to have nothing to do with whether training is fair use.
  - gremlinunderway 3y ago
    Good luck trying to explain "updating weights incrementally based on predicted next token probabilities" to completely non-technical lawyers and judges.
    - lesuorac 3y ago
      Good thing they don't have to. As I've said before, this slight of hand to talk about the case as-if its about training is a great move by OpenAi; however the case is more than just about training. NYT is alleging that OpenAI maintains a database of NYT works and refers to it verbatim (per-prompt) as part of the ChatGPT software and not via some pre-trained wieghts. This is akin to having to explain to non-technical lawyers and judges how crypto works. In the FTX case it became irrelevant because you can just nail them on fraud for using deposited funds for non-allowed reasons.
      - simonw 3y ago
        I didn't see NYT allege that - their lawsuit explains pre-training pretty accurately I thought.
        lesuorac 3y ago
        Its buried on page 37 - #108. There probably are other examples in the lawsuit but this is sufficent. > Synthetic search applications built on the GPT LLMs, including Bing Chat and Browse with Bing for ChatGPT, display extensive excerpts or paraphrases of the contents of search results, including Times content, that may not have been included in the model’s training set. The “grounding” technique employed by these products includes receiving a prompt from a user, copying Times content relating to the prompt from the internet, providing the prompt together with the copied Times content as additional context for the LLM, and having the LLM stitch together paraphrases or quotes from the copied Times content to create natural-language substitutes that serve the same informative purpose as the original. In some cases, Defendants’ models simply spit out several paragraphs of The Times’s articles. https://www.courtlistener.com/docket/68117049/1/the-new-york-times-company-v-microsoft-corporation/ https://www.courtlistener.com/docket/68117049/1/the-new-york...
        simonw 3y ago
        Oh I see - yeah, that's the part of the lawsuit that's about Bing and ChatGPT Browse mode retrieval augmented generation. It's a separate issue from the fact that the model can regurgitate it's NYT training data. There's a section on page 63 which helps clarify that: Defendants materially contributed to and directly assisted with the direct infringement perpetrated by end-users of the GPT-based products by way of: (i) jointly-developing LLM models capable of distributing unlicensed copies of Times Works to end-users; (ii) building and training the GPT LLMs using Times Works; and (iii) deciding what content is actually outputted by the GenAI products, such as grounding output in Times Works through retrieval augmented generation, fine-tuning the models for desired outcomes, and/or selecting and weighting the parameters of the GPT LLMs. So they are complaining about models that are capable of distributing unlicensed copies (the regurgitated training data issue), the fact that the models were trained on NYT work at all, and the fact that the RAG implementation in Bing and ChatGPT Browse further creates "natural-language substitutes that serve the same informative purpose as the original".
- solomatov 3y ago
  Yep, you seem to be right. Google stores the quotes from pages, and it's fair use. Again, I am not a lawyer, and didn't think about this.
- simonw 3y ago
  This is argued extensively in the lawsuit document. A key argument the NYT is making is that part of the definition of fair use is not producing a product that competes with the original. They argue that ChatGPT et al DO compete with the original, in a way that harms the NYT business model. One example they give: ChatGPT can reproduce recommendations made by the Wirecutter, without including the affiliate links that form the Wirecutter's main source of revenue - page 48 of https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec20...
  - shkkmo 3y ago
    There are plenty of other services that strip affiliate links from content for users, such as ad blockers. Notably, in both those cases, the user is specifically asking for what wirecutter thinks. To me, that makes the infringing behavior clearly the fault of the user of the tool, not the tool itself. When I saw this lawsuit announced, I assumed that large portions content were being reproduced in response to generic queries. That isn't the case, every example I've seen from this lawsuit has prompts where the user specifically asks for the content. To me, any fault and liability rests on the user here.
dkjaudyeqooe 3y ago
Fair use is a key defense for OpenAI. The article also mentions the idea that the model is merely extracting (uncopyrightable) facts, which is interesting, but might be a tough one to prove since LLMs have no way of establishing what are facts, and don't return facts by design.

It's not a fair use question, it's a question about models reproducing the article text almost verbatim. (IANAL)

brigadier132 3y ago

If it's fair use reproducing the article text verbatim is fine.

andy99 3y ago
What is "it"? Training can be fair use, i.e. updating weights incrementally based on predicted next token probabilities. And I (not a lawyer) think that if a broadly trained foundation model can recall some verbatim text, that doesn't mean the model is infringing. It seems like the lawsuit here is talking about specific NYT related functionality, like retrieving the text in new articles. That essentially has nothing to do with training and running a foundation LLM, it's about some specific shenanigans with NYT content, and it's legal status would appear to have nothing to do with whether training is fair use.
- gremlinunderway 3y ago
  Good luck trying to explain "updating weights incrementally based on predicted next token probabilities" to completely non-technical lawyers and judges.
  - lesuorac 3y ago
    Good thing they don't have to. As I've said before, this slight of hand to talk about the case as-if its about training is a great move by OpenAi; however the case is more than just about training. NYT is alleging that OpenAI maintains a database of NYT works and refers to it verbatim (per-prompt) as part of the ChatGPT software and not via some pre-trained wieghts. This is akin to having to explain to non-technical lawyers and judges how crypto works. In the FTX case it became irrelevant because you can just nail them on fraud for using deposited funds for non-allowed reasons.
    - simonw 3y ago
      I didn't see NYT allege that - their lawsuit explains pre-training pretty accurately I thought.
      - lesuorac 3y ago
        Its buried on page 37 - #108. There probably are other examples in the lawsuit but this is sufficent. > Synthetic search applications built on the GPT LLMs, including Bing Chat and Browse with Bing for ChatGPT, display extensive excerpts or paraphrases of the contents of search results, including Times content, that may not have been included in the model’s training set. The “grounding” technique employed by these products includes receiving a prompt from a user, copying Times content relating to the prompt from the internet, providing the prompt together with the copied Times content as additional context for the LLM, and having the LLM stitch together paraphrases or quotes from the copied Times content to create natural-language substitutes that serve the same informative purpose as the original. In some cases, Defendants’ models simply spit out several paragraphs of The Times’s articles. https://www.courtlistener.com/docket/68117049/1/the-new-york-times-company-v-microsoft-corporation/ https://www.courtlistener.com/docket/68117049/1/the-new-york...
        simonw 3y ago
        Oh I see - yeah, that's the part of the lawsuit that's about Bing and ChatGPT Browse mode retrieval augmented generation. It's a separate issue from the fact that the model can regurgitate it's NYT training data. There's a section on page 63 which helps clarify that: Defendants materially contributed to and directly assisted with the direct infringement perpetrated by end-users of the GPT-based products by way of: (i) jointly-developing LLM models capable of distributing unlicensed copies of Times Works to end-users; (ii) building and training the GPT LLMs using Times Works; and (iii) deciding what content is actually outputted by the GenAI products, such as grounding output in Times Works through retrieval augmented generation, fine-tuning the models for desired outcomes, and/or selecting and weighting the parameters of the GPT LLMs. So they are complaining about models that are capable of distributing unlicensed copies (the regurgitated training data issue), the fact that the models were trained on NYT work at all, and the fact that the RAG implementation in Bing and ChatGPT Browse further creates "natural-language substitutes that serve the same informative purpose as the original".
solomatov 3y ago
Yep, you seem to be right. Google stores the quotes from pages, and it's fair use. Again, I am not a lawyer, and didn't think about this.
simonw 3y ago
This is argued extensively in the lawsuit document. A key argument the NYT is making is that part of the definition of fair use is not producing a product that competes with the original. They argue that ChatGPT et al DO compete with the original, in a way that harms the NYT business model. One example they give: ChatGPT can reproduce recommendations made by the Wirecutter, without including the affiliate links that form the Wirecutter's main source of revenue - page 48 of https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec20...
- shkkmo 3y ago
  There are plenty of other services that strip affiliate links from content for users, such as ad blockers. Notably, in both those cases, the user is specifically asking for what wirecutter thinks. To me, that makes the infringing behavior clearly the fault of the user of the tool, not the tool itself. When I saw this lawsuit announced, I assumed that large portions content were being reproduced in response to generic queries. That isn't the case, every example I've seen from this lawsuit has prompts where the user specifically asks for the content. To me, any fault and liability rests on the user here.

dkjaudyeqooe 3y ago

Fair use is a key defense for OpenAI. The article also mentions the idea that the model is merely extracting (uncopyrightable) facts, which is interesting, but might be a tough one to prove since LLMs have no way of establishing what are facts, and don't return facts by design.