> The SEC enforces laws that forbid people from misleading investors, regardless of whether fundraisers seek capital in public or private markets. The SEC often closes investigations without making formal accusations of wrongdoing.
> Some of the people familiar with the investigation described it as a predictable response to the former OpenAI board’s claim in its November statement. One of the people said that the SEC hasn’t pointed to any specific statement or communication by Altman that it has deemed misleading.
> The SEC’s civil investigation has been percolating in the background as OpenAI officials pitched investors as part of its recently closed tender offer, which valued the AI juggernaut behind viral chatbot ChatGPT at more than $80 billion.
> OpenAI is governed by a nonprofit. Investors in its for-profit arm include employees, venture capitalists, and Microsoft MSFT 0.06%increase; green up pointing triangle, which has committed $13 billion to the company in exchange for what is essentially a 49% stake in the earnings of its for-profit arm.
> At that time of the leadership turmoil, OpenAI executives started getting questions from regulators and law-enforcement entities such as the Manhattan U.S. Attorney’s Office about the board’s accusation of Altman’s lack of candor, The Wall Street Journal reported in November.
That's a quote from the article. WSJ puts ticker data after company references. Green up pointing triangle is the accessible tag for the image and was simply copy and pasted from the source content.
There is a better song to teach the kids - dont work for financial engineers.
If you have a purely creative, engineering or technical background you will end up under, or dependent on financial engineers soon enough.
To get out of that trap, either develop your own business and finance skills. Or work for people who dont have a pure business or finance background. A good mix is best.
And you can then live a life where you dont have to pay attention to content creators on wall st and build things you actually care about.
"I sometimes write “everything is securities fraud” as a shorthand for a particular weird argument about how modern US legal dynamics transmute every bad action taken by a public company into securities fraud.."
A bunch of companies pledged $1bln to a non-profit. A non-profit was turned into a de facto for-profit company. That’s what they should have been investigating since 2019.
I think Occam's Razor applies here, which is to say an inexperienced and idealistic board got cold feet and tried to remove Altman on principle. They failed because Altman is not only a political animal, but also one with friends in high places. Can't remove the helmsman at peak hype cycle and expect no consequences.
I don't understand why people refuse to see it. They had a nine-member board with three insiders but three independent directors resigned so they were split 3-3. The board was split on choosing new directors so they were at an impasse for months.
Then Altman cooked up an excuse to attempt to remove one of the independent directors and lied to other board members about it and that convinced Ilya Sutskever to vote to remove Altman from the board to restore its independence. Firing Altman was just fallout from that.
Edit: It occurs to me that Ilya could have resolved the situation by resigning from the board, but I suppose that was a non-obvious solution.
The problem with all the "Sam was bad" scenarios I've seen is: why did all the replacement CEOs chosen by the board side with Sam instead of the board?
My pet conspiracy theory on this is that members of the board were blackmailed by someone who wanted Sam gone, didn't want to admit whatever the embarrassing secret was, were therefore unable to give a convincing explanation to any of their chosen replacement CEOs.
But even I don't take this seriously, it's merely the only story I've got that doesn't seem to contradict any of the observed actions.
I can’t stop chuckling at how many people have talked me up about how innovative their structure was. It’s funny how trendy things are even when they make no sense.
This old pro, a copywriter, a Greek named Teddy, told me the most important thing in advertising is 'new'. It creates an itch. You simply put your product in there as a kind of calamine lotion.
Copyright infringement is mostly civil matter in the USA, not a criminal one, there is a very high bar to pass for criminal charges that doesn’t exist for civil ones. Also, I’m not convinced that organizations like the New York Times didn’t accidentally just discover entrapment via prompt injection when making their claims. The courts will sort that out quickly, however.
You give pieces of the article during your chat session giving enough context for the LLM to put it together when you ask for it.
The other one is just being confused when Bing, who has some sort of deal with the publishers, includes the article in the prompt when you casually use the system to ask for its first ten lines, and get it. It isn’t baked into the model, the attention aspect is just really leaned into.
Mostly it’s just people not understanding how LLMs work, and assuming everything is coming from the model when they don’t realize it’s all coming from information in the prompt, and the model is just really good at getting things from the prompt.
There's no such thing as "entrapment via prompt injection" since it still requires the LLM to have the underlying article in its training set.
And that is precisely the problem: that the underlying article is in the training set and in the body of reproducible output without permission. This requires an act of copying, and was done without permission, so it's a copyright violation unless a fair use exception applies. (Courts haven't yet ruled on that; it's likely that the NYT case is the case that will ultimately lead to a SCOTUS ruling resolving the issue one way or the other.)
This is incorrect. The model does not need to be trained on an article for the LLM to regurgitate it if the article is fed into it (even in small pieces) in the prompt.
Again, quit misunderstanding the tech or we can’t argue about it in an intelligent way.
OpenAI states directly in their response the NYT lawsuit that an article needs to be part of its training set in order to be regurgitated, and indeed they note that "memorization" can be a big problem if an article shows up multiple times in its training set. They further admit that their LLM is (or was) capable of replicating content from its training set wholesale and that specific safeguards were put in place after complaints to limit that.
OpenAI accuses the NYT of using "manipulated prompts" to trigger regurgitation of NYT artices.
But so what? For legal purposes, and more relevantly for copyright purposes, it's irrelevant that the LLM needed large prompts to "regurgitate" the entire article.
What matters is that, given a prompt that was less than the full article the LLM was able to replicate the entire article. In other words, it distributed a copy of an NYT article without the right to do so. And that is literally what they are being sued for.
Whether the use of the article in the training set, or the reproduction of the article in the LLM responses, constitute fair use is still to be decided (and those answers will determine who wins this case.) But it's the legal facts that will determine the winner, not the technical implementation, because those details are irrelevant for purposes of copyright law.
> OpenAI states directly in their response the NYT lawsuit that an article needs to be part of its training set in order to be regurgitated, and indeed they note that "memorization" can be a big problem if an article shows up multiple times in its training set. They further admit that their LLM is (or was) capable of replicating content from its training set wholesale and that specific safeguards were put in place after complaints to limit that.
> Memorization is a rare failure of the learning process that we are continually making progress on, but it’s more common when particular content appears more than once in training data, like if pieces of it appear on lots of different public websites. So we have measures in place to limit inadvertent memorization and prevent regurgitation in model outputs.
...
> We had explained to The New York Times that, like any single source, their content didn't meaningfully contribute to the training of our existing models and also wouldn't be sufficiently impactful for future training. Their lawsuit on December 27—which we learned about by reading The New York Times—came as a surprise and disappointment to us.
> Along the way, they had mentioned seeing some regurgitation of their content but repeatedly refused to share any examples, despite our commitment to investigate and fix any issues. We’ve demonstrated how seriously we treat this as a priority, such as in July when we took down a ChatGPT feature immediately after we learned it could reproduce real-time content in unintended ways.
> Interestingly, the regurgitations The New York Times induced appear to be from years-old articles that have proliferated on multiple third-party websites. It seems they intentionally manipulated prompts, often including lengthy excerpts of articles, in order to get our model to regurgitate. Even when using such prompts, our models don’t typically behave the way The New York Times insinuates, which suggests they either instructed the model to regurgitate or cherry-picked their examples from many attempts.
End quote.
> But so what? For legal purposes, and more relevantly for copyright purposes, it's irrelevant that the LLM needed large prompts to "regurgitate" the entire article.
They used multiple queries to build up a fairly large prompt, it is not not irrelevant at all.
> What matters is that, given a prompt that was less than the full article the LLM was able to replicate the entire article. In other words, it distributed a copy of an NYT article without the right to do so. And that is literally what they are being sued for.
That is incorrect, they only need to provide enough context to reproduce the article, providing excerpts of the article in their own prompting would immediately raise red flags in a legal context. NyTimes could be found at fault here, because if they are found to have purposely primed the prompt, it won't surprise me if NyTimes winds up paying OpenAI over this.
> But it's the legal facts that will determine the winner, not the technical implementation, because those details are irrelevant for purposes of copyright law.
In this case, there is no difference, this was resolved long ago in lawsuits against xerox (is xerox liable if you use it to copy a copyrighted book? nope). Just because you get ChatGPT to reproduce an article because you yourself provided excerpts of the article to it doesn't mean it infringed on the content. If anything you did.
NyTimes will have to provide their methodology as evidence in a civil trial, and the fact they haven't released it so far, they are definitely not confident of their position and are hoping for a settlement.
Just because you get ChatGPT to reproduce an article because you yourself provided excerpts of the article to it doesn't mean it infringed on the content. If anything you did.
This is a fundamentally incorrect statement of the facts as they exist in the real world. OpenAI admits that the NYT article was in their training set, they are taking issue with the means the NYT used to prove that their article was part of the training set and reproducible as a response to a prompt. I repeat: the NYT prompts did not "put" the article into the LLM; it was already there waiting to be extracted.
Furthermore, the NYT methodology is irrelevant to the copyright case. The technology underlying the LLM is irrelevant to the copyright case. All of these claims might earn brownie points on the internet, but they are completely irrelevant to the court case in the real world.
What matters is that the NYT article was copied without permission, and whether OpenAI can claim a fair use exception to do so.
the fact they haven't released it so far, they are definitely not confident of their position and are hoping for a settlement.
This is not how the legal process works in general, and it's not how copyright cases work in particular. (Things like "methodology" get disclosed during the discovery stage, not the complaint stage. This is how trials have worked in the U.S. for over two centuries...)
Yes, the NYT is hoping for a settlement. They want OpenAI to license their articles. They were upfront about that from the beginning. It was OpenAI's refusal to license the NYT articles that prompted the lawsuit.
The only thing I am sure of about copyright law, is that I don't fully understand it.
Search engines, for example, necessarily read as much of the internet as possible, in order to create an index, which is a derivative of the original content; when giving you results, many show snippets from those search results, so that users have an understanding of why the link is relevant to their query, which is definitely a copyright violation because Google lost (or settled?) on that exact issue with regard to news sites (and image results?).
Search engines are one category of AI. PageRank was a significant improvement over simple indexing, and personalised results are a significant improvement over pure PageRank (and Google's "moat" after Page's patent expired).
So the impression I have is that merely downloading a public page and putting into an algorithm isn't a violation by itself, while repeating even short clips of text verbatim totally is.
But things aren't static, they are rewritten to respond to the possibilities of the world. Is this going be accepted or not, going forwards, given what AI can now do?
I don't know.
I can see the two camps, but I can't foresee which will be the victor; and it won't necessarily be the same in all countries, given the balance of interests between creators and users isn't assured to be the same everywhere.
(And even that's assuming no good models trained on public domain or suitably licensed documents, even though there are already some such models).
As much as I wish it were otherwise, it's far from clear that what OpenAI did constituted copyright infringement under US copyright law. I suspect not. But the courts will clear this up.
54 comments
[ 3.0 ms ] story [ 124 ms ] thread> The SEC enforces laws that forbid people from misleading investors, regardless of whether fundraisers seek capital in public or private markets. The SEC often closes investigations without making formal accusations of wrongdoing.
> Some of the people familiar with the investigation described it as a predictable response to the former OpenAI board’s claim in its November statement. One of the people said that the SEC hasn’t pointed to any specific statement or communication by Altman that it has deemed misleading.
> The SEC’s civil investigation has been percolating in the background as OpenAI officials pitched investors as part of its recently closed tender offer, which valued the AI juggernaut behind viral chatbot ChatGPT at more than $80 billion.
> OpenAI is governed by a nonprofit. Investors in its for-profit arm include employees, venture capitalists, and Microsoft MSFT 0.06%increase; green up pointing triangle, which has committed $13 billion to the company in exchange for what is essentially a 49% stake in the earnings of its for-profit arm.
> At that time of the leadership turmoil, OpenAI executives started getting questions from regulators and law-enforcement entities such as the Manhattan U.S. Attorney’s Office about the board’s accusation of Altman’s lack of candor, The Wall Street Journal reported in November.
That's a quote from the article. WSJ puts ticker data after company references. Green up pointing triangle is the accessible tag for the image and was simply copy and pasted from the source content.
"Everything is Securities Fraud" - OpenAI edition, coming Feb 29 to Matt Levine's Money Stuff near you
If you have a purely creative, engineering or technical background you will end up under, or dependent on financial engineers soon enough.
To get out of that trap, either develop your own business and finance skills. Or work for people who dont have a pure business or finance background. A good mix is best.
And you can then live a life where you dont have to pay attention to content creators on wall st and build things you actually care about.
Agree on developing some financial basics so, this only helps!
https://archive.is/igIqF
"I sometimes write “everything is securities fraud” as a shorthand for a particular weird argument about how modern US legal dynamics transmute every bad action taken by a public company into securities fraud.."
Then Altman cooked up an excuse to attempt to remove one of the independent directors and lied to other board members about it and that convinced Ilya Sutskever to vote to remove Altman from the board to restore its independence. Firing Altman was just fallout from that.
Edit: It occurs to me that Ilya could have resolved the situation by resigning from the board, but I suppose that was a non-obvious solution.
My pet conspiracy theory on this is that members of the board were blackmailed by someone who wanted Sam gone, didn't want to admit whatever the embarrassing secret was, were therefore unable to give a convincing explanation to any of their chosen replacement CEOs.
But even I don't take this seriously, it's merely the only story I've got that doesn't seem to contradict any of the observed actions.
To accidentally hire one such CEO may be regarded as a misfortune; to accidentally hire three in a week looks like carelessness.
If this crap flies I'm going to deploy the first non-Euclidean capital structure.
https://cc.bingj.com/cache.aspx?d=4416390193669&w=T-7EMYqCcE...
Unless being a CEO in the US gives you a free get-out-of-jail card.
The other one is just being confused when Bing, who has some sort of deal with the publishers, includes the article in the prompt when you casually use the system to ask for its first ten lines, and get it. It isn’t baked into the model, the attention aspect is just really leaned into.
Mostly it’s just people not understanding how LLMs work, and assuming everything is coming from the model when they don’t realize it’s all coming from information in the prompt, and the model is just really good at getting things from the prompt.
Again, courts will sort those claims out quickly.
And that is precisely the problem: that the underlying article is in the training set and in the body of reproducible output without permission. This requires an act of copying, and was done without permission, so it's a copyright violation unless a fair use exception applies. (Courts haven't yet ruled on that; it's likely that the NYT case is the case that will ultimately lead to a SCOTUS ruling resolving the issue one way or the other.)
Again, quit misunderstanding the tech or we can’t argue about it in an intelligent way.
OpenAI accuses the NYT of using "manipulated prompts" to trigger regurgitation of NYT artices.
But so what? For legal purposes, and more relevantly for copyright purposes, it's irrelevant that the LLM needed large prompts to "regurgitate" the entire article.
What matters is that, given a prompt that was less than the full article the LLM was able to replicate the entire article. In other words, it distributed a copy of an NYT article without the right to do so. And that is literally what they are being sued for.
Whether the use of the article in the training set, or the reproduction of the article in the LLM responses, constitute fair use is still to be decided (and those answers will determine who wins this case.) But it's the legal facts that will determine the winner, not the technical implementation, because those details are irrelevant for purposes of copyright law.
Nope, that's not what they claim, and you are intentionally misrepresenting them. See https://openai.com/blog/openai-and-journalism:
> Memorization is a rare failure of the learning process that we are continually making progress on, but it’s more common when particular content appears more than once in training data, like if pieces of it appear on lots of different public websites. So we have measures in place to limit inadvertent memorization and prevent regurgitation in model outputs.
...
> We had explained to The New York Times that, like any single source, their content didn't meaningfully contribute to the training of our existing models and also wouldn't be sufficiently impactful for future training. Their lawsuit on December 27—which we learned about by reading The New York Times—came as a surprise and disappointment to us.
> Along the way, they had mentioned seeing some regurgitation of their content but repeatedly refused to share any examples, despite our commitment to investigate and fix any issues. We’ve demonstrated how seriously we treat this as a priority, such as in July when we took down a ChatGPT feature immediately after we learned it could reproduce real-time content in unintended ways.
> Interestingly, the regurgitations The New York Times induced appear to be from years-old articles that have proliferated on multiple third-party websites. It seems they intentionally manipulated prompts, often including lengthy excerpts of articles, in order to get our model to regurgitate. Even when using such prompts, our models don’t typically behave the way The New York Times insinuates, which suggests they either instructed the model to regurgitate or cherry-picked their examples from many attempts.
End quote.
> But so what? For legal purposes, and more relevantly for copyright purposes, it's irrelevant that the LLM needed large prompts to "regurgitate" the entire article.
They used multiple queries to build up a fairly large prompt, it is not not irrelevant at all.
> What matters is that, given a prompt that was less than the full article the LLM was able to replicate the entire article. In other words, it distributed a copy of an NYT article without the right to do so. And that is literally what they are being sued for.
That is incorrect, they only need to provide enough context to reproduce the article, providing excerpts of the article in their own prompting would immediately raise red flags in a legal context. NyTimes could be found at fault here, because if they are found to have purposely primed the prompt, it won't surprise me if NyTimes winds up paying OpenAI over this.
> But it's the legal facts that will determine the winner, not the technical implementation, because those details are irrelevant for purposes of copyright law.
In this case, there is no difference, this was resolved long ago in lawsuits against xerox (is xerox liable if you use it to copy a copyrighted book? nope). Just because you get ChatGPT to reproduce an article because you yourself provided excerpts of the article to it doesn't mean it infringed on the content. If anything you did.
NyTimes will have to provide their methodology as evidence in a civil trial, and the fact they haven't released it so far, they are definitely not confident of their position and are hoping for a settlement.
This is a fundamentally incorrect statement of the facts as they exist in the real world. OpenAI admits that the NYT article was in their training set, they are taking issue with the means the NYT used to prove that their article was part of the training set and reproducible as a response to a prompt. I repeat: the NYT prompts did not "put" the article into the LLM; it was already there waiting to be extracted.
Furthermore, the NYT methodology is irrelevant to the copyright case. The technology underlying the LLM is irrelevant to the copyright case. All of these claims might earn brownie points on the internet, but they are completely irrelevant to the court case in the real world.
What matters is that the NYT article was copied without permission, and whether OpenAI can claim a fair use exception to do so.
the fact they haven't released it so far, they are definitely not confident of their position and are hoping for a settlement.
This is not how the legal process works in general, and it's not how copyright cases work in particular. (Things like "methodology" get disclosed during the discovery stage, not the complaint stage. This is how trials have worked in the U.S. for over two centuries...)
Yes, the NYT is hoping for a settlement. They want OpenAI to license their articles. They were upfront about that from the beginning. It was OpenAI's refusal to license the NYT articles that prompted the lawsuit.
Search engines, for example, necessarily read as much of the internet as possible, in order to create an index, which is a derivative of the original content; when giving you results, many show snippets from those search results, so that users have an understanding of why the link is relevant to their query, which is definitely a copyright violation because Google lost (or settled?) on that exact issue with regard to news sites (and image results?).
Search engines are one category of AI. PageRank was a significant improvement over simple indexing, and personalised results are a significant improvement over pure PageRank (and Google's "moat" after Page's patent expired).
So the impression I have is that merely downloading a public page and putting into an algorithm isn't a violation by itself, while repeating even short clips of text verbatim totally is.
But things aren't static, they are rewritten to respond to the possibilities of the world. Is this going be accepted or not, going forwards, given what AI can now do?
I don't know.
I can see the two camps, but I can't foresee which will be the victor; and it won't necessarily be the same in all countries, given the balance of interests between creators and users isn't assured to be the same everywhere.
(And even that's assuming no good models trained on public domain or suitably licensed documents, even though there are already some such models).