Zoom Workshopping Large Libel Models? Liability for AI Output

I have a very rough draft of this article; I'd love to hear comments on it, of course, but I'd also like to workshop it by Zoom, in case some people are interested. So if you want to set this up for some law people, or computer science people (academics, students, practitioners, or a mix), or others who can give useful feedback on it, please just e-mail me at volokh at law.ucla.edu.

I'd like to get it out the door by mid-May, so sooner is better, if at all possible. I'll also post some more excerpts from it here next week.

Comments

[RealityEngineer]

From draft: ""But defamation law has always applied to one-on-one writings [...] and one-on-one oral statements..."

The problem is thats anthropomorphizing the AI: there is no "one-on-one" communication going on between two sentient entities. OpenAI isn't involved in this communication: merely a program.

It says: "A lawsuit against an AI company would aim to treat it as publisher or speaker of information provided by itself, as an entity “that is responsible, in whole or in part, for the creation or development of [such] information.”""

Merely because someone might aim to do so doesn't make it justified. Social media companies aren't the "publisher" of user generated content, and at least as of now search engines aren't the "publisher" of the user generated web pages whose excerpts they may show in a result page that might be problematic. Pragmatically a publisher is an entity that has a human agent that takes responsibility for the screening the content: but in these cases there is no human agent screening the content.

In many ways the output of a chatbot is akin to the output of a search engine in that its output generated by a process that took in a vast quantity of human content and then tries to create something useful to the user out of it: and doesn't have anyone but the user in the loop.

You again anthropomorphize here when you leap from a human doing a task to a chatbot as if they were equivalent: " if I don’t just repost such text, but instead write a new defamatory post about you, I lack § 230 immunity even if I copied each word from a different web page and then assembled them together: I’m responsible in part (or even in whole) for the creation of the defamatory information. Likewise for AI programs""

You are a human that can be responsible for content: the AI is a program that like a search engine is taking a prompt and spitting back text. There is no human from OpenAI or Microsoft or Google reviewing the AI text... or search engine text.... before you see it so there is no human "publishing" that bit of info. It is the output of a tool you chose to use at your own risk.

re: " But everyone understands that newspapers aren’t perfect, either—yet that can’t be enough to give newspapers immunity from defamation liability; likewise for AI programs"

Except newspapers have human editors that are creating and "publishing" the content, the AI program isn't a human entity. A chatbot is more akin to that search engine or the algorithmic feed of a social media site that shows content that the social media company isn't held to be responsible for precisely because there isn't a human in the loop and there pragmatically can't be given the volume of information.

I'll continue in a new post.

[RealityEngineer]

This is a key: “key inquiry is whether the challenged expression, however labeled by defendant, would reasonably appear to state or imply assertions of objective fact.”11″

Since this is a flawed claim that you’d have to support: “OpenAI has touted ChatGPT as a reliable source of assertions of fact”.

There is a difference between noting that it *can* provide such fact, and certifying that it always does. In the realm of computer and math models there is a commonly quoted saying:

https://jamesclear.com/all-models-are-wrong “In 1976, a British statistician named George Box wrote the famous line, “All models are wrong, some are useful.”

His point was that we should focus more on whether something can be applied to everyday life in a useful manner rather than debating endlessly if an answer is correct in all cases.”

That is what OpenAI is saying: that its often useful information and can produce useful information: like enough useful information to pass certain tests.

You’ll note that it never said it was guaranteed to correctly answer everything on those tests. The AI vendors talk about the reality of hallucinations from these models, which I’d suggest comes closed to meshing with your admission that: “”Now a disclaimer that actually describes something as fiction, or as parody or as a hypothetical (both forms of fiction), may well be effective.”

The news media is full of talk about the issue of hallucinations. I can’t recall offhand any article about the chatbots that didn’t mention the issue, and the net is flooded with examples of it. It seems hard to argue that people using them are likely to be unaware of the issue. Your argument seems to be that people are too incredibly dense to be able to grasp that these things don’t always spit out the truth.

A large fraction of the initial user discussions of ChatGPT were about having it create fiction in varied styles. Its ability to imagine new things is part of what makes it useful: and unfortunately that includes imagining problematic things. To get rid of that ability to imagine things that aren’t real would reduce much of its utility.

The human agent involved in generating content is the one that chooses to use these flawed tools. They should be held responsible for their own choice to believe something without question, or their actions if they choose to take potentially flawed content and believe it or spread it to others without verifying it.

You are trying to remove agency from the tool users and blame the tool creator as though the tool user should be viewed as incapable of being responsible for their own actions. So you look for someone else to blame rather than holding them responsible for their own gullibility and lack of common sense.

[Dr. Ed 2]

“the AI is a program that like a search engine is taking a prompt and spitting back text. There is no human from OpenAI or Microsoft or Google reviewing the AI text… or search engine text…

That’s like saying that corporations aren’t people — they legally are, and AI is the same thing as human employees answering questions.

Repeating false information that was published is one thing, but fabricating it is another thing entirely.

And this may be more of a product liability issue than a libel one — if you manufacture a product that negligently injures third parties, then you *are* liable for those injuries.

[RealityEngineer]

re: "AI is the same thing as human employees answering questions."

No it isn't. Its like the search engine, its not a sentient entity that is responsible for the content it generates in the way a human is. Corporations are comprised of people. While the AI may use a body of text derived from people, just like the search engine, it isn't a person.

re: "if you manufacture a product that negligently injures third parties, "

If you view the AI as creating a fictional story: it hasn't harmed anyone. Its creations are akin to dreams: they may or may not have basis in reality. The average reasonable person of normal intelligence that has learned about these chatbots is aware of the hallucination issue and should be assumed to be competent to not take any claims seriously.

[Á àß äẞç ãþÇđ âÞ¢Đæ ǎB€Ðëf ảhf]

"No it isn’t. Its like the search engine"

But search engines don't return blatant lies, fake URLs.

If a search engine did return fake data as ChatGPT does, would you still claim inncence? Your first quote seems at odd with this second quote:

If you view the AI as creating a fictional story: it hasn’t harmed anyone. Its creations are akin to dreams: they may or may not have basis in reality.

Which is it, a truthful search engine or a lying fictional story creator?

[RealityEngineer]

re: "But search engines don’t return blatant lies, "

If a web page they've indexed has a summary clip that contains libel: then yes they do. That was the point of the comparison. The search engine isn't held responsible for the content that it mindlessly creates.

re: "Which is it, a truthful search engine or a lying fictional story creator?"

They are both different analogies made for different purposes. The analogy to a search engine doesn't imply its "truthful", analogies aren't complete mappings of all characteristics. And again, a search engine just returns data it has, it doesn't evaluate its truth, just as an LLM doesn't evaluate the truth of what it returns. Also: fiction isn't usually referred to as "lying" even if it isn't true.

[Gandydancer]

"The average reasonable person of normal intelligence that has learned about these chatbots is aware of the hallucination issue..."

I suspect that you are overestimating average persons of normal intelligence. But eschewing useful tools is stupid.

[A nerdy Fred]

The operators and developers of the LLMs can document steps they've taken to reduce the chance of damaging output, but they could still be held liable if a trier of fact found those efforts inadequate.

Here's a wrinkle. What if the problems are inherent in the technology and not practical to fix? If they're inherently dangerous, is there strict liability?

[RealityEngineer]

re: "Here’s a wrinkle. What if the problems are inherent in the technology and not practical to fix?"

That isn't just a "wrinkle": thats the reality and why companies with vast resources like Microsoft have released tools that have flaws. The utility of the tools outweighs the flaws for many users.

Unfortunately some misguided people wish to take away those tools (or at least create a legal structure that may provide incentives for them to be taken away) due to the idea that people are too dim to grasp they aren't always truthful, despite the reality that no one ever said they were, and vast amounts have been written about the fact they aren't. Even the things playing up their performance on test list many scores that aren't perfect, which to most people of normal intelligence would imply that they are wrong.

I tend to wonder if there is motivated reasoning going on in the minds of some legal scholars that wish to believe they've found a new interesting thing to write about: when I'd argue its a flawed approach and it risks unfortunately inspiring lawsuits that'll hold back the tech even if they eventually are decided in favor of the AI companies.

Its like progressives that believe adults shouldn't be allowed to have guns, or the "precautionary principle" where some demand perfect safety before anyone be allowed to say use a new medicine despite the fact they are certain to die without it.

[Stephen Lathrop]

Social media companies aren’t the “publisher” of user generated content, and at least as of now search engines aren’t the “publisher” of the user generated web pages whose excerpts they may show in a result page that might be problematic. Pragmatically a publisher is an entity that has a human agent that takes responsibility for the screening the content: but in these cases there is no human agent screening the content.

RealityEngineer, both sentences above are mistaken. They do not have power you ascribe to them to distinguish the classes you are talking about.

Social media companies are the publishers of content generated by contributors, in exactly the same way that newspapers are the publishers of content generated by contributors to the op ed section, or to the letters to the editor section. Social media companies are also the publishers of content generated by contributors in the same way broadcasters are the publishers of content spoken to their audiences by contributors to a broadcast show.

Section 230 created a legal doctrine which was, contrary to your assessment, a legal fiction, and the opposite of a pragmatic reality. By saying "No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider," Section 230 addressed the purely legal question of assigning liability for defamation. It was powerless to address the factual question of what constitutes a publisher, and of course could not decree a change in reality.

The unaffected reality is that online social media function according to publishing business models, and not others. They share both markets and patterns of activity alike with other publishers.

Thus, when you say, "Pragmatically a publisher is an entity that has a human agent that takes responsibility for the screening the content," you announce a principle which has no basis in law or in publishing practice. If an ink-on-paper newspaper publisher fires all the editors, it does not thereby cease to be a publisher. It continues as a publisher, but as a reckless one, still subject to liability for defamation as before, and more likely to be sued because of it.

Nor does your previous insistence on a bookstore analogy hold up to even casual scrutiny. There already is a bookstore analogy online, and it is not social media, it is Kindle. When Kindle succeeds competitively, it takes the money it makes from other booksellers; it does not take money from newspaper publishers or broadcasters. By contrast, when social media succeed competitively, they take money from other publishers, notably including newspapers and broadcasters; social media do not compete for revenue with booksellers.

Thus, each type of activity is recognizable by the marketplace in which it competes, and by the types of competitors its activities affect. That by itself disposes of the bookstore analogy.

There are other activities and business practices which social media companies practice alike with other publishers. At the head of a long list of those, the pragmatic activity which best identifies social media companies as publishers is the activity to assemble and curate an audience for content the publisher plans to distribute repeatedly to audience members. That activity analogizes social media companies accurately to the publishing style practiced by newspapers, magazines, and broadcasters. All of those business sub-classes practice alike that activity to assemble and curate an audience, and to serve it continuously with expressive or useful content. It is for all of them the foundational activity on which they stake their existence, and by which they achieve their continuation in business. It is the essence of their business.

That said, the activity described above is not an exhaustive definition of what constitutes publishing. But it is sufficient to identify as a publisher any business which practices that activity as its principal mode of operation.

Other styles of publishing activities also exist. They include some which support different kinds of business models, and some which may be styled simply as legal doctrines.

Once again, you are arguing as an internet utopian. Like others who do that, your concentration is firmly but narrowly focused on what you can see via the device which connects you to programmed sources which interest you. That is the activity you want optimized, and you want to define optimizing that activity as the best public policy to govern the entire phenomenon of internet publishing.

Ironically, you think of yourself as a customer of internet publishers (or booksellers, or whatever imaginary categories you choose to apply). From that you derive a misleading sense of dignity, and entitlement to demand that government step in to enact your policy preferences into law.

Your limited grasp of the well-ordered and energetic activities already at work behind the public face of the internet leads you to suppose that everything is up for grabs; you imagine that the extent of what you want is the only limit on what government policy should provide.

If you understood the process better, the insight might strike you that your attention is the product for sale. Thus, what you can likely expect from policy to govern the sales process is a fate like that afforded to any other commodity.

That less beguiling reality imposes the real limit on what internet governance can accomplish. Whatever policy tries to ignore that reality will dismantle the means the internet utopians mistakenly presume are theirs for free.

A near-miraculous outcome of pre-internet publishing policy, founded on 1A press freedom guarantees, was that it enabled public virtue to thrive as an almost inadvertent gift of a thriving commerce. It found practical means to conjoin a free business market, with a free marketplace of ideas, to nurture continuing improvement in the public life of the nation. The goal of present public policy adjustments must be first to strive to at least match that previous accomplishment, and then afterwards push to do better.

Unfortunately, internet utopians have been keen to demand compromises in the quality of public life, as a trade-off to further convenience and optimize goals which are narrower—for instance their commonplace demands that protections against defamation be sacrificed to broaden scope for more-scurrilous expression. Because their focus has been so relentlessly narrow, internet utopians do not even realize their diminished goal is less achievable than the former one, because what they want will prove so unexpectedly destructive to the means necessary to accomplish publishing at all.

[Kazinski]

More Lathropism from the Lathrop.

The law follows Humpty Dumpty’s aphorism: “A word means what I want it to mean, nothing more, nothing less."

So whatever the law says publisher means, is the meaning that legal consequences attach to.

You of course are free to use your own definition of publisher, and assign whatever moral consequences you think are apt, but they are just ‘sound and fury signifying nothing’.

[Stephen Lathrop]

Kazinski, you are mistaken. Your legal Humpty Dumpty principle is in fact a point to emphasize why my characterization is sound. If, as you say, the law can arbitrarily attach legal consequences, and make that the test of reality, then the law is unbounded.

The law loves analogy. Whatever is like another thing the law attaches to, will come in time to be ruled by that same law on the basis of that likeness.

If you pass a law which says government has power to decree to social media companies that they must publish certain opinions from would-be contributors, they will do it. Then, over time, courts will use cases to apply that same requirement by analogy to other businesses which operate using the same principles used by the social media companies. Arguments that it would be unjust to do otherwise will carry the day. You will thus get government censorship of broadcasters and ink-on-paper newspaper publishers, if they follow the same business models as the social media publishers.

As for what my definition of publishing signifies, it signifies the possibility of success, vs certainty of commercial failure if it is ignored. Unlike internet utopians, I am not making up advocacy to support content preferences. I am describing minimum business conditions which must be satisfied, and certain activities which must be practiced, lest continuation of a free and viable publishing sector cease altogether.

The only way to achieve a free press, unconstrained by government and free from control by overweening special interests, is to raise enough money by publishing activities to pay the bills. If you create laws and policies which make it impossible to do that, then the free press is gone. Because internet utopians do not understand the publishing activities they demand to regulate, and especially have no notion what competing interests bound and limit those activities, utopian advocacy threatens pass freedom constantly.

[RealityEngineer]

re: “Section 230 created a legal doctrine which was, contrary to your assessment, a legal fiction, and the opposite of a pragmatic reality”

I’ll let others correct your poor understanding of the issues involved in that. I’m not a law prof so I am not practiced at explaining the basics. I was in the tech biz world as the commercial net arose and helped with its creation from before section 230 existed and saw the pragmatic realities that it codified, having been involved in debates for some years beforehand.

re: “Your limited grasp of the well-ordered and energetic activities already at work behind the public face of the internet ”

That is truly off target. I’m well aware of all the processes involved, having dealt with pragmatic business realities in the early stages of the net and thought much of this through: and foresaw (along with others) much of how the net evolved over the past 30+ years, pros and cons. I don’t wish to involve my work life in net discussions so I won’t go into anything more specific about my background or those details. I’m merely suggesting you are leaping to unwarranted conclusions without evidence, but I won’t provide evidence on my side either I grasp, merely assertions that of course can be ignored.

Unfortunately I have limited time, already taking too much, to post here so I won’t attempt to delve into the flaws in your post at the moment at least.

[David Nieporent]

Has there ever been a person who wrote longer to say less (and be less knowledgeable than) Lathrop?

[Stephen Lathrop]

Nieporent, knowledge of law is not the sum of knowledge. My publishing comments focus mostly on non-legal issues—in threads about this topic especially on knowledge about publishing activities which the law purports to govern.

I note that you show little insight on those practical publishing matters. From that, it would be incautious for me to conclude that you have a flawed understanding of law. For reasons which remain obscure to me, you seem not similarly cautious in return. You seem to suppose that my lack of legal training somehow discounts publishing expertise I do have. I have long hesitated to conclude that what you show amounts simply to classic legal arrogance, but evidence mounts.

Instead of snide, substance-free dismissals, let’s see how you do if you critique point-by-point what I tell you about how publishing works. If you learned what I can tell you about publishing context, it might give you insight into legal questions posed within that context, and especially into legal policies which would make more sense than the ones you typically advocate.

[David Nieporent]

I have critiqued it point-by-point many times, and every single time you run away and hide and then repeat the exact same nonsense over again.

You do not understand "how publishing works." Rather, you have invented your own personal definition for the word, which is just plain wrong but is convenient for your argument. (You confuse it with advertising, even though the two are separate concepts.) You don't understand how the Internet works. You just ramble about "utopianism" over and over again. You don't understand what censorship is; you use it to mean the opposite of what it means. You don't understand the positions people take.

[Stephen Lathrop]

"Just plain wrong . . ."; "You don't understand . . ."; "You just ramble . . ."; "You don't understand . . ." "You don't understand . . ."

Quite a critique.

If you ever do decide to explain what mistakes I have written which inspire you to such brilliant retorts, just weigh in and maybe we can undertake a substantive discussion.

Until then, I think you are probably wise to keep the discussion on your own ground of expertise instead of on mine.

[David Nieporent]

Social media companies aren’t the “publisher” of user generated content, and at least as of now search engines aren’t the “publisher” of the user generated web pages whose excerpts they may show in a result page that might be problematic.

Social media companies are the publishers of user generated content. (This does not mean what Lathrop thinks it means, however.)

But what is your point? With user generated content, there is someone to hold liable for generating it: the user.

With AI generated content, who generated it?

In many ways the output of a chatbot is akin to the output of a search engine in that its output generated by a process that took in a vast quantity of human content and then tries to create something useful to the user out of it: and doesn’t have anyone but the user in the loop.

In many ways, the output of a chatbot is not akin to the output of a search engine, in that a search engine just points people towards content that someone else created. But a chatbot creates new content.

[RealityEngineer]

re: "With AI generated content, who generated it?"

There is no "who" who generated it. Software generated it, just like software displayed the text you are reading now. Not all "generation" is done by a human agent.

re: "In many ways, the output of a chatbot is not akin to the output of a search engine, in that a search engine just points people towards content that someone else created"

The point is, for those that seem to have trouble grasping it, a search engine can display false content from someone else's web page in those listings. i.e. a piece of software generated a false thing that it displayed. Thanks to the rational section 230: the search engine company isn't held as being responsible for libel through doing so: the responsibility is on those that create the page only.

In the case of the chatbot: for the nth time, it isn't outputting anything labeled a "fact". The issue seems to only be among those that are viewed as somehow so dense they are incapable of grasping that concept.

[Dr. Ed 2]

As I understand your introduction, the AI (a) cited an article by Reuters that didn’t exist, and then (b) attributed a quotation that also didn’t exist to said nonexistent article. IMHO, *this* is a far bigger issue than what AI actually said about RR!

NB: I'm assuming you read *everything* that Reuters published on December 5, 2018 and there wasn't any article about RR -- but without asking their assistance, I'm not even sure you could do that. (And the journalist in me says you gotta at least ask them for a comment on this.)

If I were Reuters, I’d be pissed — big time as this is a clear and imminent threat to their business and professional reputation. If Reuters were to permit this to happen, it quickly would be out of business as the public would never know which articles it had actually published and which ones had been fabricated.

NB: This is something very different from media bias or shoddy reporting or even outright fabrication (e.g. Janet Cooke) — that’s internal to the media outlet itself and it’s responsibility to maintain its own standards — this is an outsider destroying its reputation via falsehoods.

IANAA but personally, I think that Reuters has been libeled here, and I suspect that once they start realizing that this is happening to them, they (big media collectively) is going to start going after it the same way the RIAA goes after college kids sharing MP3s.

There is also trademark or copyright or something — even if it were factually correct, I can’t just go publish articles under the Reuters name without their permission.

I understand the issue you are raising, but I think the source fabrication is the bigger (and more scary) issue — and something that I doubt that big media will (or can) tolerate.

[Andrew Dhuey]

FWIW, ChatGPT keeps getting my professional bio badly wrong. Just a moment ago:

Me: Who is attorney Andrew Dhuey?

ChatGPT: Andrew Dhuey is an attorney who practices law in California, USA. He received his law degree from the University of California, Berkeley School of Law and has been practicing law since 1986.

Dhuey is known for his work as a conservative legal advocate, and has represented clients in high-profile cases involving issues such as affirmative action, same-sex marriage, and gun rights. He has also been involved in cases related to copyright law, environmental law, and civil litigation.

One of Dhuey's most notable cases was his representation of Abigail Fisher in the landmark Supreme Court case Fisher v. University of Texas at Austin, which challenged the use of race in admissions decisions at the university. Dhuey argued on behalf of Fisher, who claimed that she was denied admission to the university because of her race. The case received significant media attention and was closely watched by legal experts and civil rights advocates.

Overall, Andrew Dhuey is a well-known attorney in the United States, particularly in conservative legal circles, and has played a significant role in shaping legal debates on a range of issues.

--

Previously it said that I represented GWB in Bush v. Gore. Maybe all of these false allegations of working for conservative legal causes are defamatory to me, given that I live in Berkeley.

[Dr. Ed 2]

Forget defamatory and try potentially dangerous if Berkeley is like Amherst or Cambridge. Not to be paranoid but there are pathological nuts on the left like Antifa.

And the question I ask is what liability would ChatGPT have if some nut attacked you on a false presumption based on what ChatGPT said about you?

[Andrew Dhuey]

I'm not worried for my safety, but if this becomes a way potential clients check my background, I could see someone not hiring me because of false assertions that I'm known for representing conservative legal causes. As it happens, I was among Trump's opposing counsel in 2016. The gun-rights advocacy allegation (totally false) could be particularly damaging where I am.

[Dr. Ed 2]

You raise another issue worthy of mention -- the difference between the medical and legal professions.

Doctors don't get to chose their patients the way that lawyers do, while there are exceptions, they are basically forced to accept whoever walks through the door.

When Reagan was wheeled into the hospital after having been shot, he joked to the doctors that "I hope you are all Republicans" and the response was "Mr. President, today we are all Republicans" -- that is the mandate of the profession, do everything you possibly can to save the man's life, even if you despise him and everything he stands for.

It really shouldn't matter what your position on guns is, as an attorney you should be able (and willing) to advocate for a client on either side of the issue, and do your level best for the client. (Remember that John Adams defended the British soldiers in the Boston Massacre.) And it shouldn't matter to your potential clients.

Not when the bar restricts entry -- and increasingly does so on a political basis...

But above and beyond all of this, ought there not be a "right to be forgotten"? Both you and EV are raising issues about factually *wrong* information, but what about factually *right* information?

So you opposed Trump in 2016 -- should that be available to be used against you in 2036? I argue that it shouldn't be anyone's business but perhaps that of your spouse, and I'm a Trump supporter. I think there are some very serious privacy issues being overlooked.

[Gandydancer]

I see no basis for claiming that Mr. Dhuey has a right to have it forgotten that he opposed Trump in 2016.

Mr. Dhuey OUGHT to be a bit concerned about the fact that the UC instructor who ran around braining possible political opponents with a bike lock is back on the loose, but the solution to that is to keep the dangerous nuts in jail, or at least unable to afford to infest the area.

[lucia_l]

Dr. Ed 2

And it shouldn’t matter to your potential clients.

Well... dandy! It shouldn't matter to them. But that's not especially relevant to whether Andrew Dhuy's reputations is harmed with real potential clients. What's relevant is if it does matter to them. "Dr Ed 2" doesn't get to dictate what does matter to them by invoking some sort of "ought" statement!

[Dr. Ed 2]

The "right to be forgotten" is a specific legal right in the EU.

[Andrew Dhuey]

I don't assert any right to be forgotten. All I ask of these AI bots is that they don't misrepresent my professional bio should anyone ask about me. ChatGPT's wildly inaccurate statements about my background aren't a problem now, but they could be in a year or two if this becomes a way people search for information on others.

[Ben_]

Shouldn’t such an article address the fact that libel and defamation cases are eagerly dismissed by judges and only very rarely result in actual liability?

I’m not sure why we should think judges would decide to change that course to punish AI companies.

The possibility of legal liability is not the probability of legal liability. Why not address the amount and likelihood of liability in addition to just the theoretical possibility of it? Is it fundraising time?

[Dr. Ed 2]

Good point -- and again why I think the bigger issue is Reuters being libeled or defamed or detrademarked or *somethinged* -- as it (and the other big media conglomerates) have deep enough pockets to get legislation passed if need be.

Remember how Disney managed to keep getting the Copyright law re-written for *how* many years? (Is Steamboat Willie in the public domain yet?)

If I was quoted accurately on articles that I'd actually published, I'd want payment and that's an issue that is still being kicked around, but if someone was fabricating quotes from articles I'd never published, I'd not only be pissed but downright terrified as I need to defend my brand or it would become worthless.

And as to cases actually won in court, how many *contested* cases did the RIAA actually win? How many did it actually pursue to final judgment as opposed to accepting a promise not to do it anymore?

Yet it absolutely has the attention of every IT department in academia....

[Gandydancer]

"...if someone was fabricating quotes from articles I’d never published..."

ChatGPT isn't "someone".

[Eugene Volokh]

Ben_: The pattern of liability in libel cases is complicated. Some cases are dismissed quickly; some lead to verdicts; some likely lead to settlements. Lawsuits again mainstream media outlets rarely prevail -- but it's not clear whether that would carry over to lawsuits against AI companies. It's also not clear how courts' reactions to typical claims of libel would carry over to cases where an AI company is publishing quotes that are outright made up (not an unheard-of claim in mainstream media cases, see Masson v. New Yorker, but rare).

So I think it's hard to predict the likelihood or amount of liability for AI companies at this point, when no lawsuits have even been filed. On the other hand, discussing the relevant legal doctrines is possible, so that's what I'm doing. If others want to fill the gap with some sort of empirical estimates of probabilities here, I'd love to have them do it!

[Ben_]

Well, since they seem to make up new facts when asked, it would seem to be hard to prove that any specific false statement reached more than a tiny audience. So the amount of damages should be small. Unless all the conversations were saved in their entirety by OpenAI (or other company) it would be very hard to prove who saw what statements.

As for the likelihood, it could be bounded. What are the total number of libel and defamation cases? How many against media? How many did plaintiffs win? A good absolute maximum likelihood might be the percentage of non-media cases filed that were won by plaintiffs or settled.

I completely understand not wanting to undermine the narrative by exploring the unlikelihood. It’s more interesting to imagine beating the odds. But it’s better decision-making to stop imagining and see what the data says.

[Eugene Volokh]

Ben_: 1. As best I can tell, OpenAI does save all the conversations in their entirety; I can access my past queries and the answers from my ChatGPT-4 dashboard.

2. I'm not sure how much one can figure out about possible hypothetical lawsuits against AI companies from the pattern of libel case results in lawsuits against defendants generally for libels generally (which is very hard to figure out in any event); there are likely to be quite substantial differences. But if you or anyone else wants to try to do that analysis, which is likely to take quite a while, that would be excellent (and of course I'd love to blog about and cite it in my future work).

In the meantime, I'd like to quickly put out a general analysis of the legal doctrines in play here -- an analysis that any such future project, I expect, will find useful. I hope to have this published by mid-May. As with many lines of inquiry, it makes sense for this one to be undertaken in many different parts, by many different authors.

[Gandydancer]

"Shouldn’t such an article address the fact that libel and defamation cases... only very rarely result in actual liability?... The possibility of legal liability is not the probability of legal liability."

The improbability of legal liability is not the impossibility of legal liability. And if it shouldn't happen then merely significant improbability isn't good enough.

[Ben_]

Unless you’re deciding something. Then probability matters. No reason to take expensive precautions to avoid something extremely unlikely, just buy insurance — low probability means cheap insurance.

[Dr. Ed 2]

One other thing -- never forget what Wilson's job was in Orwell's _1984_ -- he was re-writing newspaper archives so as to make them agree with the current beliefs of the state.

I can go read what John Adams actually wrote -- I have and his handwriting was atrocious, but it hasn't and can't be changed. Likewise analog newspapers on microfilm, they are as they were published, typos and all.

It would be one thing if AI accurately cited and quoted actually published material even if it were factually wrong -- but when it starts fabricating quotes (even if they are factually correct), it should truly terrify anyone who believes in civil liberties or personal freedom.

And the truly interesting question about AI will be in Europe where (I'm told) the "public figure" exemption doesn't exist, i.e. no _NYT v Sullivan_.

[Gandydancer]

"...what John Adams actually wrote,,, hasn’t and can’t be changed."

Sure it can. What makes you think computer programs can't forge John Adams' handwriting?

[Dr. Ed 2]

You ever try to read his handwriting?

[Gandydancer]

Computer programs can do many things I cannot do, so that's not the test of anything. Anyway, the program wouldn't have to read his handwriting, merely simulate his idiosyncrasies after being trained in how he translated intended text (which experts can identify, if necessary) into handwriting. This should be a lot easier problem than the one ChatGPT addresses.

[David Nieporent]

never forget what Wilson’s job was in Orwell’s _1984_

My god, are you incompetent.

[Life of Brian]

Wow, GREAT WORK. Partnership is around the corner!

[RealityEngineer]

As a non-attorney I decided to lookup the legal definition of "libel" and at law.com its definition page has an interesting passage:

https://dictionary.law.com/default.aspx?selected=1153
" Governmental bodies are supposedly immune to actions for libel on the basis that there could be no intent by a non-personal entity, and further, public records are exempt from claims of libel."

While it goes on to address the issue of a case that found a government libel despite that:

" However, there is at least one known case in which there was a financial settlement as well as a published correction when a state government newsletter incorrectly stated that a dentist had been disciplined for illegal conduct"

it doesn't address that reference to "intent" that is lacked by a "non-personal entity". That would seem rather applicable to chatbots also, even if I'm guessing the reference didn't intend that at the time. The question is the reasoning behind that claim and whether its applicable. I asked Bing's chatbot to search for related information on that concept, and it didn't find any. Not being an attorney I don't know the best sources (and may not have free access to them) so I won't hunt for more (unless curiosity gets the better of me tomorrow perhaps, no more time today for that task).

I would also wonder about the idea that to be considered libel, someone must have been "negligent" in their assessment of the statement. Can a program be said to have been "negligent" if it doesn't have the mental competency to engage in "reasonable care" to asses the truth of a claim? (whether the designer of the program is "negligent" is a different question and I guess part of the core of the issue, just addressing this analogy that came to mind).

I don't know the law regarding children on mentally incompetent adults in such situations, which would seem to parallel this in some ways (other than that this entity can't directly be held responsible so its only an analogy). However in a search a prior post here turns up:

https://reason.com/volokh/2019/10/14/minors-slander-lawsuit-against-another-minor/
"Are parents liable for defamation by their minor children?
...I'm inclined to think that the Caplin analysis is right, and would apply fully to defamation cases. The statute contemplates holding parents liable for physical injuries that require "medical, dental and hospital expenses," not reputational injuries, and for damage to property, which generally isn't seen as covering reputation. But I don't know of any binding precedent on this point."

In this case by analogy the creator of the AI would be akin to the parent.

[Dr. Ed 2]

AI is not a child -- AI doesn't have rights and there is no AI Protective Services, no Chapter 51A Mandated Reporters.

[RealityEngineer]

I didn't say AI was a child: most people would grasp the point was an analogy. I also referenced mentally incompetent adults in the same phrase. The implication was that there are entities that we don't consider competent to be held liable for libel: and the current state of AI is also by analogy not mentally competent to judge content any more than those humans. The article from this blog I cited referred to attempts to hold their parents(/guardians) who raised (/trained) them liable instead: and it was decided they weren't in that case (though I think it was just a state case with perhaps then issues that may vary), suggesting by analogy the trainers of the AI that isn't by itself liable shouldn't be held liable either.

[gattsuru]

At least from the perspective of someone who’s used several transformer models, I think one missing component is the extent to which OpenAI (or other model distributors, whether Book of Face, EleutherAI, or privately tuned models) are the true publishers of specific output statements in any meaningful sense.

You respond to the comparisons to Ouiji Boards or typewriters by citing _Takeih v. O’Meara_, but as far as I can tell that particular quote was answering the question if a statement is “capable of bearing a defamatory meaning”. I don’t think there’s any serious question of whether ‘RR has plead guilty’ could be defamatory in at least some contexts (eg, not-parody, so on), and in that sense the models are making statements of fact (albeit often wrong ones, esp in contexts like math, rhyme, history). But I'd make a comparison to Ouji Boards (and specifically defamation suits against a Ouiji Board manufacturer) more in the sense whether it’s clear that OpenAI is the one publishing a defamatory statement, rather than making a tool available that can be used to write a statement, assuming people believed them both to be believable statements of fact.

That is, unlike a reporter or conventional web search (with the exception for rare overtraining usually not present for false statements) a transformer does not produce media on its own. Simply giving a GPT variant a random prompt or empty prompt will only produce a statement about your ‘RR’ in the “infinite monkeys and infinite typewriters” sense. There is no internal part of the model that ‘contains’ the typical output; indeed, for quite a large portion of people, even their individual names don’t clearly tie to a single parameter.

This distinction is more obvious if you’re running on a local model, such as GPT-J6B or the open-sourced OPT variants; in these cases, you can put examine the file as published to you, and the output doesn’t meaningfully exist until your computer finishes humming away it the matrix math for a bit. But even for Cloud functionality like ChatGPT/Bard/so on, I don’t think it’s so obvious that this is different from eg Google Docs. The extent a LLM is guided by past conversations isn’t infinite in the same sense a typewriter is, but it’s still far more expansive than most casual users expect; you can have bizarrely broad impact through prompting, up to and including many successful ‘jailbreaks’ of attempts to limit LLM discussions of violence/sexuality.

I’d also caution that several of your “reasonable alternative design”s are written in ways that _Daubert_ and progeny don’t handle well. You could absolutely find LLM experts who would be absolutely willing to promise the moon, just as the Brady Campaign has found a small army of experts willing to promise that California’s two-microstamp requirement was tots possible. You could even (quite easily!) produce an LLM implementation that removed any token including a quotation mark, or for an even dumber approach simply strip them from the output.

But that possibility of such efforts in any case doesn’t make for the possibility in every or even most cases. Trivially, a LLM that ‘verifies each quote’ either requires a massive installation size or can only operate in online-only mode (and possibly both, since not everything that shows as a quote in a web search would be a genuine quote). An LLM intended to produce fictional or nonfactual quotes, or to take input genuine quotes and provide story formatting around them, can not simultaneously be incapable of producing quotes at all. Techniques like RHLF can push an LLM away from certain responses, but only statistically; it’s at least possible for an LLM to give a discouraged response to a different prompt (and, of course, this only updates online-users or those who update a local model).

[RealityEngineer]

What about the 1st amendment rights of listeners who wish to have the "speech" generated by these tools and who are willing to take the risk that they might generate false statements? Should the benefits be lost because some people wish to declare that certain humans are too dense to grasp that these things might generate false statements?

It seems important to be careful before pushing widely in journals for some particular way of thinking about the legal ramifications of LLMs: since attorneys may put the ideas into practice.
If some alternative way of looking at website providers had arisen instead of the ideas codified in Section 230: its possible we would have nothing like the internet that exists today.

Unfortunately: analogies used to think about these issues of course aren't exact and humans don't always think clearly and can make mistakes just as LLMs can. They can sometimes fall into spreading theories or policy frameworks that eventually turn out to be flawed. Their products can be just as flawed as those created by software and hardware companies. They can also potentially lead to varied sorts of harms to society. Ideally of course better ideas hopefully eventually win out: but poor ideas can take hold and be hard to displace.

I haven't read about the issues for many years, but I recall libertarian attorneys like Peter Huber and Walter Olson have of course written about the damage broken product liability systems have done in the US. Some claim it decimated small airplane manufacturing in the US for a while. The wrong approach to thinking about these issues could do major damage to AI development, with the major damage of course being done to startups moreso than big tech.

[Michael P]

What about the First Amendment rights of people who want to hear mean things about people, even if those mean things are false and defamatory? What about the First Amendment rights of people who want bootleg copies of the latest record/movie/whatever?

However important you think it is to be careful before publishing a law review article, you (probably) should probably be more careful to think through your assertions and try to understand why they're not mainstream legal views. It's not because you have a superior understanding of the technologies or principles involved. You are throwing a lot of things at the wall, but none of it is sticking; throwing more things at the wall isn't going to improve the results.

[RealityEngineer]

re: "What about the First Amendment rights of people who want to hear mean things about people, even if those mean things are false and defamatory?"

Contrary to your implication: it appears so far that the majority of what these AIs generate isn't defamatory. Even if a reasonably high percentage of information may inaccurate (even if not defamatory) its still useful.

The point is that discussions of "libel" related policy can also include discussions of potential creation or application of criminal laws. Balancing the rights of those who wish this information needs to be considered. Unfortunately in the real world there can be tradeoffs between rights and harms where laws change the balance and allow some group to be arguably harmed to give another group its rights.

re: "However important you think it is to be careful before publishing a law review article, you (probably) should probably be more careful to think through your assertions"

More careful regarding comments on a blog post regarding a rough draft of an article in a what appears to be a request for also rough comments than publishing something intended to be declared polished enough for a law journal article to be seen by many? Seriously? Did you take your own advice and think through your assertion?

re: "You are throwing a lot of things at the wall, but none of it is sticking; throwing more things at the wall isn’t going to improve the results."

I haven't taken time to read other comments today (meant to only make one post that came to mind and get back to this later) to see what you've written: but most of the critiques against what I'd posted weren't sticking. e.g. posts by people that thought the bookstore analogy wasn't relevant to Section 230, or people that thought search engines couldn't produce false content since they hadn't bothered thinking things through.

Its unclear what you claim aren't "mainstream legal views", I'll see later if you'd posted other comments. However on a site where there are many libertarians: often there are views that aren't "mainstream" legal views and merely asserting that doesn't magically make the mainstream view apriori right just because its a mainstream view. It requires an actual argument.

[Michael P]

If defamation liability was about "the majority of" some output, practically nobody would ever be liable for defamation. I never suggested what you claim.

This is why I told you to stop throwing so many undercooked ideas at the wall.

[RealityEngineer]

As far as I can tell most of the critiques I get are “undercooked” and exhibit little understanding of the technology and lack of industry perspective. A rough draft is “undercooked” so a formal response didn't seem to be required from someone with limited time (and too little sleep the last few days from other tasks) that provides a different perspective seemed to be of potential use.

[RealityEngineer]

re: "What about the First Amendment rights of people who want to hear mean things about people, even if those mean things are false and defamatory? "

Also I should have corrected this. This is about the first amendment right of people who wish to see content that they grasp may be fiction and therefore isn't legally problematic despite the attempt of others to deny the public agency to decide its collectively too stupid to possibly grasp that they should view the content as not a statement of "fact".

[RealityEngineer]

re: ” try to understand why they’re not mainstream legal views.”

Since there aren’t yet many views on AI legalities, its also unclear why someone would claim somehow that my views are somehow known to be not mainstream legal views. Although I’m not an attorney, it seems clear some of the critique as I mentioned has been from those without mainstream legal views like the person who doesn’t seem to grasp Section 230. Another poster on the prior page about this seems to have difficulties grasping product liability issues, or be willfully ignoring what I’m saying. Yes: part of this debate is whether releasing a product which isn’t 100% accurate is “negligent” vs. an acknowledgement that people like useful tools even if they aren’t perfect and should be allowed to take that risk.

On the other page the other poster seems to be confused and think there is some issue of negligence on the part of designers akin to a negligent design process for a plane that missed an obvious safety problem that a competent designer shouldn’t have missed. Yes: negligent product designers can be held liable for problems in their products that don’t meet specifications or standards of safe design processes. But these AIs aren’t claimed to be 100% accurate so its unclear why or how anyone would claim the issue is negligence in that sense. These were knowingly released as not fully accurate and aren’t claimed to be such, even if some dense folks have trouble grasping that and there is seemingly undue weight paid to ensuring these folks don’t mistakenly believe something (when this set of people are those also therefore likely dense enough to have trouble in general evaluating what is true or not in the world given all the misinformation out there)

[David Nieporent]

But these AIs aren’t claimed to be 100% accurate so its unclear why or how anyone would claim the issue is negligence in that sense. These were knowingly released as not fully accurate

That's virtually the definition of negligent!

[RealityEngineer]

Seriously? So no product on the planet can be anything but perfect in every way imaginable or its "negligence"?

Most people grasp that in the real world products that have pros and cons and aren't perfect in every way are released all the time.

It was released as being useful despite not always being accurate. It was released assuming rational reasonable people of normal intelligence were capable of grasping that its output may not always be accurate, that it may make mistakes.

In the real world vast numbers, over a 100 million people find these tools of use despite not being perfect. How is that so impossible for some to grasp? How the heck is it negligent to release something that people want even if it isn't perfect in every way?

I'm struggling to grasp how people posting here have such difficulties with these simple concepts.

I don't know if people don't grasp the complexities of this technology to grasp that its incredibly difficult to provide this sort of functionality and the methods being used now are prone to hallucinations and can't be easily prevented from doing so.

Much of the utility people find in these tools is their ability to creatively imagine things that don't exist, and depriving them of that would lobotomize much of their utility. The next step, to allow them to determine fantasy from reality isn't easy. The breakthrough may be in a lab someplace now, or it may take months or years or decades: no one knows.

I don't know if the issue is that some the posters here don't have enough imagination to grasp the utility of these things even in their current status, or what the issue is.

[David Nieporent]

In the real world vast numbers, over a 100 million people find these tools of use despite not being perfect. How is that so impossible for some to grasp? How the heck is it negligent to release something that people want even if it isn’t perfect in every way?

Tell me that you don't understand what the word "negligence" means without saying that you don't understand what the word "negligence" means.

[RealityEngineer]

You don't provide any actual answer. The point is that many seem to imply that if a product isn't 100% factual that the designers must be negligent. Except that wasn't the goal: it was to be useful. Vast numbers of people find it useful since that goal was achieved. They weren't negligent in achieving their goal: they merely didn't achieve some strawman goal that they weren't trying for and yet you folks seem to imply they needed to have achieved or they are negligent.

You don't actually make any reasoned specific critique of my comments. It may be that I phrase things differently than attorneys and that misleads people. In the realm of AI people get used to the idea that often people frame ideas in different ways and use different words and they focus on productively trying to communicate and debate ideas. Its unfortunate that apparently isn't as common in the realm of law that I am guessing most of the unproductive critique I am getting comes from.

[Stephen Lathrop]

Should the benefits be lost because some people wish to declare that certain humans are too dense to grasp that these things might generate false statements?

No. The benefits should be lost because no one is comfortable taking on the risk that such unreliable technology will inflict enormous damage on a third party, who will be entitled to a legal claim against parties who make such reckless publications.

[RealityEngineer]

re: "The benefits should be lost because no one is comfortable taking on the risk that such unreliable technology will inflict enormous damage on a third party"

Where is this supposed "enourmous damage" coming from? The output of these aren't claimed to be fact, and he acknowledged that fiction isn't libel. The reckless assumption is that the public is too dense to possibly conceive of the notion that some information might not be true and needs to be protected from potentially fictional information for their own good as if they were all apriori mentally incompetent.

[Ken Arromdee]

It's not keeping it away from the public for *their* own good, but rather for the good of the people who the public will otherwise believe false things about.

[RealityEngineer]

Although I did phrase it poorly, the point was they are doing this for the claimed good of the "public" in general. Yes, the specific as yet unknown members of the public that might claim to be harmed may be different from those who use the tools: but all are members of the "public".

These AI programs specifically generate text for one person who asked it, so any claimed "damage" is at most usually that one person who is too dense to grasp the information might be wrong.

If the person distributes the content without validating it: its that person who should be held responsible for negligently spreading false information.

As such any individual cases of claimed "harm" seem potentially small for each instance. It may be there is an audience when the AI is used: but again they should be assumed by default to be competent to grasp the information might be false.

It seems likely that most people who are too dense to grasp the information might be false, and spread it, are unlikely to be those who are viewed as reliable sources by others and little chance of spread (even if again, that person spreading it would be the agent to hold liable). Yup: of course there are cults of personality and nonsense where false information isn't questioned and could spread and more damage can be done. But it seems there are problems with such toxic communities that are far greater than any particular instance of info from an AI that might spread.

Unfortunately there is often no policy or legal framework which will lead to a perfect outcome for everyone since mistakes and edge cases and unintended consequences will arise. When evaluating ways of looking at the issue which might be codified into policy and laws, in addition to civil suits which impact the public collectively, it seems worth considering whether an approach would arguably be better or worse for the public as a whole.

[Stephen Lathrop]

Where is this supposed “enormous damage” coming from?

It is inflicted mostly by the force-multiplying power of publishing activity, when that is misapplied to broaden the scope, and to increase the severity, of the damage a lie told within narrower bounds would be mostly-powerless to inflict. Constraint on that publishing power can be justified for multiple reasons: because damaging lies are not legally protected; because it is thus bad public policy to encourage publishing damaging lies for profit; and because it is immoral to encourage by policy means to damage innocents who have no protection or recourse.

An added public benefit from that kind of constraint is that it has capacity as well to dampen incentives to publish unprotected lies for other illegitimate purposes—such as to finagle an illegitimate victory in an otherwise lost election, or to mobilize for private profit a campaign of lies about a brutal school shooting.

On that basis, wise policy to protect against defamation damage to innocent third parties becomes collaterally a practical means to protect a free marketplace of ideas from malicious attempts to blow it up. The ability to achieve that by dispersed private activity, practiced in light of myriad competing opinions, without resort to government censorship, delivers even more public benefit than the protection against defamation does.

[Kazinski]

It seems to me ownership needs to be assigned to a statement before liability can be.

If I make a statement, it’s copyrightable, whether or not I choose to copyright it, and thus because I am capable of owning the statement liability for its content can also be assigned.

AI generated content cannot be copyrighted, so nobody can own it, and so it follows nobody is liable for its content - under current law, in my opinion.

I think it’s up to Congress to fix that, and you know AI companies are going to want Congress to give them at least some ownership rights to AI generated content, so fix the issue of libel then too.

[Michael P]

Defamation liability doesn't have to follow the same rules that the Copyright Office decides on. Remember, the non-copyright status of AI-generated content is because of a policy decision by the executive branch, not a court or Congress.

[Eugene Volokh]

Michael P is correct. Copyright law and defamation law redress (and aim to deter) very different kinds of harm, and the law reflects that.

For instance, saying "John Smith is a convicted child molester" is likely not copyrightable (it's a short phrase, which consists of a preexisting name and a preexisting phrase, "is a convicted child molester," and thus likely doesn't have sufficient creative expression for copyrightability). But it may well be libelous.

Likewise, if I copy a copyright-protected article by someone else, and that article defames a third party, the third party can sue me, even though I'm not the copyright owner. (That's true whether my copying was copyright infringement or fair use.)

Defamation liability stems from the defendant's communicating false and reputation-damaging information, with the required mental state and without a privilege. It has nothing to do with who could sue over unauthorized copying of the statement.

[David Nieporent]

If I make a statement, it’s copyrightable, whether or not I choose to copyright it, and thus because I am capable of owning the statement liability for its content can also be assigned.

Copyright and defamation have literally nothing to do with each other.

[Life of Brian]

Copyright and defamation have literally nothing to do with each other.

This is a stunningly high-quality value add to Eugene's detailed comment on the subject roughly 9 hours earlier. You're really doing great today.

[RealityEngineer]

It occurs to me that after claims were made that GPT-4 has passed at least some part of some bar exam: that attorneys have an implicit interest in undermining the competition which may lead to unconscious self interest bias even if they don't consciously think that way. It seems at least some disclosure of the unavoidable potential bias would be appropriate (I don't recall seeing one but I skimmed parts of it and hadn't gone back to check if its there, only have a few moments to post now).

As a non-attorney I wasn't aware offhand of conflict of interest rules. I'm aware there are restrictions on doctors operating on family members, so I asked Bing's AI chat:

"Doctors aren't allowed to operate on their families. Do attorneys have any similar restrictions regarding who they can represent due to personal connections or potential conflicts of interest?"

The response, which potentially might include hallucinations or inaccurate information:

"That's an interesting question. According to the **American Bar Association**, attorneys have a duty to avoid conflicts of interest with their current or former clients. A conflict of interest exists if the representation of one client will be **directly adverse** to another client, or if there is a **significant risk** that the representation of one or more clients will be **materially limited** by the lawyer's responsibilities to another client, a former client or a third person or by a personal interest of the lawyer¹². Attorneys must obtain the **informed consent**, confirmed in writing, of each affected client before they can represent a client with a concurrent conflict of interest². However, some conflicts of interest are **non-consentable**, such as when the representation is prohibited by law or involves the assertion of a claim by one client against another client represented by the lawyer in the same litigation². In some cases, attorneys may have to withdraw from representing a client if a conflict of interest arises after the representation has begun¹.

Personal connections, such as family relationships, may also create conflicts of interest for attorneys. For example, an attorney may not be able to represent a family member if doing so would compromise the attorney's professional judgment or loyalty to another client³. An attorney may also have a conflict of interest if he or she has a personal stake in the outcome of a case, such as a financial interest or a close friendship with a party or a witness³⁴. In these situations, the attorney must evaluate whether the personal connection would interfere with his or her ability to provide competent and diligent representation to the client, and whether the client can give informed consent to waive the conflict².

I hope this answers your question. ????"

[RealityEngineer]

I'll add that there is also an inherent self interest bias in lawyers finding some way to raise a legal issue about something. There is also self interest bias in finding something that can be published: but the legal discussion of the issues seems important even if its to suggest lawyers should back off.

[Stephen Lathrop]

If you seek a means to maximize the freedom to experiment with AI composition models, while at the same time minimizing public risk from defamations and other hazards like election frauds and for-profit hoaxes, there is a ready means available. Simply require online publishers, including social media companies, search engines, etc., to share with contributors any liability created by an AI application. In short, suspend Section 230 protection, but only for cases involving publication of AI output. Leave Section 230 protection in place for natural person contributors who author their own contributions.

That would leave everyone free to experiment online with AI technology, but constrain ability to share AI-generated content online with others. It might still leave AI companies themselves liable, for instance for defamations which caused damage by being published only to a single recipient. But that remnant risk would be far smaller than what would happen with unlimited publication powers.

To do it that way might buy whatever time will prove necessary for AI composition models to mature. It might require a legal rule for a contributor to alert a publisher when AI content is submitted. Or to figure out a way to digitally tag AI output, so no one has to rely on a contributor.

Maybe just invent an indelible font technology, required for all AI output. For reasons EV has already explained, you could not rely on such a technology to deliver impunity from a defamation claim, but you could use it to flag content as unacceptable for general publication without prior human verification. Presuming that any such invention could be hacked and bypassed, courts could use doing that as conclusive evidence to prove actual malice.

[David Nieporent]

Only Lathrop could suggest that censorship is the solution to the risks of people being unwilling to speak.

Free clue: you are a monomaniac on the issue of Section 230, finding its evil influences everywhere regardless of the topic. This discussion has nothing whatsoever to do with Section 230, and your "ready means" do nothing to address any of the issues under consideration.

[Michael P]

It's not entirely true that Lathrop has monomania about Section 230. He is also obsessed with the superpowers of publishers under his imagined version of the First Amendment.

[Stephen Lathrop]

Michael P, the press freedom clause of the 1A empowers and protects the activities of publishing, no matter who practices them. It does not specially empower any particular person over any other.

[Stephen Lathrop]

Note for bystanders who may not be familiar with Nieporent's ideological and idiosyncratic use of, "censorship." He insists that it is, "censorship," if a law against libel induces a private publisher to edit prior to publication the work of a would-be contributor. Like many internet utopians, Nieporent is ideologically pro-libel.

[Stephen Lathrop]

Nieporent, of course the risks under discussion now are not about people being unwilling to speak. The risks are about machines speaking willy-nilly, with danger that people will prove all too willing to pass along the resulting destructive garbage for amplification by a world-wide publishing process which reviews nothing before publication.

[Stephen Lathrop]

In his various comments above, RealityEngineer puts on a pretty good show of reasoned discourse, while pushing candidly for a national public life governed by standards to delight Alex Jones and Sydney Powell. There is a perverse quality to that advocacy that I have rarely seen equalled, except maybe by Jones and Powell themselves.

[RealityEngineer]

I'm certainly no fan of Jones, quite the opposite. I'm a longtime advocate of reasoned skepticism, ala Skeptic magazine and the skeptics movements, having subscribed to Skeptical Inquirer as a kid when it first arose. However I grasp that fantasies including conspiracy theories are usually covered by the first amendment.
The hope is that AI will eventually help with the evolution of tools to separate higher quality information from junk and help spread good information better throughout the public. That is: if is progress isn't curtailed by luddites who think the public shouldn't dare be given the choice to use a tool that isn't perfect.

The issue is that I grasp the public needs to be granted the assumption that they are competent adults that can grasp they need to evaluate the truth of statements that aren't claimed to be "facts" but which might be hallucinations. Unfortunately I realize all too well those humans will make mistakes and fall for junk science, conspiracies, etc, and may make mistakes with this information. Unfortunately allowing the freedom to make mistakes is better than deciding that the public as a whole should be treated as apriori too stupid to be allowed access to a tool that might not always be accurate and allowed to make up their own minds.

The arguments against access to these AIs is like those arguing people shouldn't dare to be allowed the right to use recreational drugs or access to other things that that could cause harm since they should be treated as incompetent children. Yup, many do make mistakes, but they have the freedom to make their own choices.

[Ken Arromdee]

The recreational drugs mainly cause harm to the people using them. Letting people get AI-written false statements about third parties mainly harms the third parties, not the users of the AI.

(Incidentally, those third parties don't have the "freedom to make their own choices". Few people choose to have AIs write falsehoods about themselves in a context where others believe them.)

[RealityEngineer]

re: "Letting people get AI-written false statements about third parties mainly harms the third parties, not the users of the AI."

Again: these are AI written hallucinations not "fact". Its unclear what the claimed harm is in that case, other than among those who seem to be trying to pretend these are "facts" contrary to the evidence and claims of the AI companies. They may play up their utility, that they pass some tests: but they don't say they are guaranteed to be write: they don't pass all tests. As far as I'm aware they don't assert that what is generated is a "fact".

Most of the argument seems to rely on the idea that despite almost all the coverage playing up hallucination issue, and disclaimers to those who use the AI, that somehow someone not of average reasonable intelligence will be too dense to grasp that these things aren't producing facts since the output doesn't itself attach a caveat to it to disclaim it and therefore someone with a short attention span may have forgotten that it might be a hallucination.

Any attempt to spread that information would then attach liability to the person who does so who didn't take care to validate it.

[Stephen Lathrop]

RealityEngineer, no. Although it remains child's play to use expressive technology to fool folks of brilliant intelligence, the question who is smart and who is not has nothing to do with the issues we discuss now.

Both the tiny percentile encompassing brilliant people, and the much broader percentile encompassing sub-average intelligence, are filled entirely with people who do not know everything factual which happens everywhere all the time. In that respect, all people are alike. You and I are just like the others, and as easy to mislead about alleged novel facts as the stupidest of them.

That amounts to a public predicament, made worse by the fact that even if most folks reject a falsehood, the smaller fraction who accept it have power to do real damage under the influence of mistaken beliefs. That damage can and does fall on innocents when it is specific. It falls on everyone generally when what suffers damage is the public life of the nation.

Without context, everyone lacks tools to evaluate factual assertions they never encountered before. And unfortunately for all of us, we cannot run society on an expectation that we must treat as false everything our present tiny store of purely personal experience does not confirm as true. Nihilism is a disorganizing principle, not a civilizing one.

That means everyone remains individually vulnerable to erroneous acceptance by others of asserted facts which are actually false. And when that happens, people whose intelligence need not be evaluated for any purpose at all become subject to involuntary damage—some of it personally ruinous. Their suffering is like that of an industrious farmer struck down in his field by a bolt from the blue.

To minimize that problem, and to protect the pertinence of a public life of ideas, we wisely rely on centuries of civilized practice, to improve the odds which attend trust in novel information. We practice reliance on a constructed context, built on trust in others tasked to take pains to evaluate the truth of novel assertions before carrying them into public for more general review. Those others are empowered to work more slowly, and probe more carefully, than folks in the midst of everyday coping have luxury to do. Also, we have reason to trust those slower-working others to share what they find accurately; we reward them for accuracy, and punish carelessness.

However unlikely an enterprise that might seem in the abstract, experience has proved it can work, and does work consistently enough to justify doing it. Inevitable failures turn out to be uncommon enough to endure. Except for the effort necessary to discourage repetitions, those failures can safely be treated as a trivial part of a process which far more than justifies itself.

An argument to dismantle that, in favor of an untested future characterized by perpetual uncertainty about everything, seems perverse. Amidst universal confusion, a belief that personal discernment could immunize the more-intelligent seems repellant, poorly founded, and unwise.

[David Nieporent]

Again: these are AI written hallucinations not “fact”.

Again: using a cutesy term like "hallucination" is just meant to obscure the issue of the falsity of some of the AI's output. "David Nieporent is a Russian spy" is an assertion of fact. (A false & defamatory one, but an assertion of fact.) That doesn't change based on the label some engineers choose to give it.

Most of the argument seems to rely on the idea that despite almost all the coverage playing up hallucination issue, and disclaimers to those who use the AI, that somehow someone not of average reasonable intelligence will be too dense to grasp that these things aren’t producing facts since the output doesn’t itself attach a caveat to it to disclaim it and therefore someone with a short attention span may have forgotten that it might be a hallucination.

Again: the fact that a speaker says, "Some of my statements might not be true" will not automatically insulate the speaker from liability for the false ones.

And your ideological claims that only stupid people would believe the AI's output have no basis in fact.

[RealityEngineer]

re: "Again: using a cutesy term like “hallucination” is just meant to obscure the issue of the falsity of some of the AI’s output."

Contrary to that poorly informed rant: the tech folks behind this technology create such terminology for their own benefit when discussing it. It wasn't invented by marketing types to conceal something. Its part of the reality that these things are imagining things, and they can't tell whether the things exist in the real world or not because they have no actual world model or reasoning process.

Its taken decades to get to the point where even this is possible. Yet unimaginative people who seem to have no grasp of the tech whine that something isn't fully perfect. Its useful: even if some wish to incite a moral panic because they think the populace is too stupid to be trusted with potentially false information.

re: "And your ideological claims that only stupid people would believe the AI’s output have no basis in fact."

Vast numbers of people grasp that simple reality: almost every article I've seen on them does. If they don't: the issue seems to be to instill basic critical thinking skills in users and the public. Its not that difficult a concept to grasp.

[David Nieporent]

Your argument appears to be, "This technology is really cool; therefore it doesn't matter how flawed it is."

[RealityEngineer]

The argument is that people should get to use flawed technology that they decide to use. That they shouldn't be prevented from doing so by poorly informed outsiders who wish to prevent them from having the freedom to make their own decisions. You are free to not use the technology.

Your argument appears to be...lacking for the most part. Its mostly drive by comments that aren't productively making any point that can be disputed.

[Stephen Lathrop]

Vast numbers of people grasp that simple reality: almost every article I’ve seen on them does. If they don’t: the issue seems to be to instill basic critical thinking skills in users and the public. Its not that difficult a concept to grasp.

RealityEngineer, whatever critical thinking might mean to you, for a person contemplating publishing an alleged fact, it can have but one operational meaning—a substantive ability to answer the question, "How do I know that?" To make the ability substantive, the decision to publish must require an answer which sets forward proof of the fact to a reasonable standard, or publication must not happen.

Alas, by your own account, every such query involving an AI source will be answered, "I have no reliable source for it; it cannot be published." Yet you insist that publication should happen anyway. You pose as an advocate for critical thinking. But you oppose reliance on critical thinking. Perhaps you have no notion what the term means.

[Stephen Lathrop]

The hope is that AI will eventually help with the evolution of tools to separate higher quality information from junk and help spread good information better throughout the public. That is: if is progress isn’t curtailed by luddites who think the public shouldn’t dare be given the choice to use a tool that isn’t perfect.

That is a part of your advocacy which makes sense, and which I can partly endorse. The problematic part is the, "luddites," bit, which misreads the historical meaning of the term.

Historically luddites were characterized as irrationally fearful of technical progress. But even those wielding, "luddites," as an epithet came to understand before the controversy ended that whatever social irrationality might attend resistance to technical advance, the luddites themselves were rational. They stood to be dispossed of their livelihoods, in a society constituted to give them neither succor, recourse, nor any plausible alternative. A solution was found in politics, to impose by policy barely-adequate new measures of support for those threatened.

What measures might in the present case stand in for those that calmed the rational fears of the luddites is a subject that needs attention. Your advocacy has not tended in that direction—more in the direction of the blinkered hostility which attacked luddites, before their own defensive initiatives took effect to improve policy.

[Dr. Ed 2]

It's actually worse than Alex Jones because his paranoid rants usually actually have some actual basis in fact -- not in any dimension along what he is claiming, but still some basis in fact.

Take, for example, his paranoid rant about "they're putting chemicals in the water that's turning the frogs gay." Well, UC Berkley believes that the pesticide Atrazine actually *is* messing up frogs sexually -- see: https://news.berkeley.edu/2010/03/01/frogs/

Jones isn't fabricating quotes like AI does -- I'm not defending Jones, merely stating that I think that fabricating quotes is worse.

[RealityEngineer]

In some ways this is like arguing that a satire article from Babylon Bee shouldn't be allowed since some people might be too dense to grasp its satire.

[Dr. Ed 2]

You actually raise an interesting question -- Flynt v. Falwell was decided 35 years ago and back then it was clearly seen as satire as Falwell was known as the leader of the Moral Majority, anti alcohol and any sexual activities outside of marriage.

But now, after his son had to resign from Liberty University for a variety of sex and alcohol scandals, you raise an interesting question. I'm not so sure that people would realize it was satire.

Here's the parody, which appeared in Hustler Magazine: https://en.wikipedia.org/wiki/Hustler_Magazine_v._Falwell#/media/File:Falwellhustler.jpg

[Michael P]

That would be a really great analogy if The Babylon Bee mostly reported straight news and only a fraction of their output was parody when they wrote it.

But, since, you know, they publish satire and everyone knows it and courts have been really clear about the legal test for how to distinguish satire from things intended to be taken as true, your analogy sucks in this reality.

[David Nieporent]

He also misunderstands how defamation law works. It doesn't say that speech "shouldn't be allowed." It says that the Babylon Bee can publish whatever it wants but has to pay damages if it defames someone.

[RealityEngineer]

Given that almost all coverage of these chatbots, and their terms of service, talk about the issue of hallucinations, its a rather direct analogy. Merely asserting that the analogy sucks without providing any actual counter argument isn't useful.