Washington Post on ChatGPT Making Up Defamatory Accusations

The article (Pranshu Verma & Will Oremus) (paywalled), which was prompted by Jonathan Turley's USA Today piece, just went up. The general analysis should be familiar to our readers (see the Large Libel Models thread), but here is a comment from an OpenAI spokesperson:

When users sign up for ChatGPT, we strive to be as transparent as possible that it may not always generate accurate answers. Improving factual accuracy is a significant focus for us, and we are making progress.

And one from Microsoft (which operates Bing, which uses GPT technology):

We have developed a safety system including content filtering, operational monitoring, and abuse detection to provide a safe search experience for our users… [U]sers are also provided with explicit notice that they are interacting with an AI system.

I think it's good to provide disclaimers, but I should note that they're not likely to be much of a defense for the companies in defamation lawsuits; more on that here.

Comments

[Brett Bellmore]

"We have developed a safety system including content filtering, operational monitoring, and abuse detection to provide a safe search experience for our users…"

It's also worth noting that what they mean by a "safe search experience" maybe has less to do with whether the results are responsive, accurate, and comprehensive, and more to do with whether WrongThink makes it through to the user.

[Mr. Bumble]

Given WAPO's history of getting things wrong what difference would ChatGPT make?

[Social Justice is neither]

Them plus the NYT not bothering to care or check if they've got facts correct and their history of defamatory statements being the base data. If journos can't be held to base standards why worry about those downstream in the information chain

[Brett Bellmore]

Look at the WaPo's list of "Trump lies", some time. A lot of it never was a lie in the first place.

For instance, Trump promised to renegotiate NAFTA. Then he... Renegotiated NAFTA! So, naively, you'd think that promise couldn't be in the WaPo's database of lies.

Nope, it is, because they don't like the way he renegotiated it.

[Tall Paul]

Perhaps the WP editors don't know it, but there's a big difference between being "inaccurate" and just making things up.

[Brett Bellmore]

Yeah, but it's not the big difference you think.

"Just making things up" is exactly what ChatGPT does, all the time. However, because it's doing so based on a rigorous and highly detailed statistical model derived from a ginormous database skimmed off the internet, more often than not the things it just makes up happen to be true. Even though it's ALL 'just made up'.

Things that are inaccurate, OTOH, are always untrue, because that's what "inaccurate" MEANS.

[RealityEngineer]

A law review article from a UCLA assistant prof that I’ve posted before and hadn’t seen its arguments addressed would suggest differently:

https://www.bu.edu/bulawreview/files/2020/09/SELBST.pdf “NEGLIGENCE AND AI’S HUMAN USERS …Decision-assistance tools are frequently used in contexts in which negligence law or negligence analogues operate, including medicine, financial advice, data security, and driving (in partially autonomous vehicles)…

If a doctor relies on a tool to help her decide to inject a drug or release a patient, we still analyze the case in malpractice despite a tool being involved; we expect the doctor to understand her tools enough to satisfy her duty of care while using them. The same goes for any other user in a context where negligence applies: if a driver cannot operate a car, we do not assume that the manufacturer is to blame….

This Article starts from the premises that AI today is primarily a tool and that, ideally, negligence law would continue to hold AI’s users to a duty of reasonable care even while using the new tool….

AI neither aims for nor can achieve perfect accuracy.299 As a result, the presence of errors does not imply a defective product required for a finding of products liability…

Moreover, in a normative sense, do we really want to simply tell the users and purchasers of complex machinery that they bear no liability for carelessness in its use?…

Where society decides that AI is too beneficial to set aside, we will likely need a new regulatory paradigm to compensate the victims of AI’s use, and it should be one divorced from the need to find fault.”

As would this law review article whose ideas regarding agency and the problems with analogies seem like ones most posters here aren’t familiar with:

https://open.mitchellhamline.edu/cgi/viewcontent.cgi?article=1223&context=mhlr “AI Entities as AI Agents: Artificial Intelligence Liability and the AI Respondeat Superior Analogy … In a relationship between an agent and its principal, the former is authorized to act on behalf of the latter for various purposes.6 This Article claims the human principal is responsible for the damages caused by its AI agent given the respondeat superior doctrine manifested via the nature of their relationship. …. Nonagency legal analogies are reduced to the basic problem of judgment proof agency, where an AI entity cannot be held liable and so a human guardian, keeper, custodian, or owner must be found liable instead in order to provide a remedy.12 ” …. Recovery is only possible via an AI agent’s human principals because AI agents are effectively judgment proof. This is because these principals are the only entity the regulator can properly incentivize to prevent damages and to invest in achieving an optimal level of activity.1″

Perhaps the concepts are too subtle and nuanced for some here: but for those from the realm of AI used to dealing with logic, there are fuzzy terms (like “publisher” or “creator”) that imply agency that are being used for things without agency like tools. There is confused imprecision throughout the blog posts and the comments regarding how these tools are magically different from examples with MS Word or other tools, search engines, etc.

When pointed out the responses tend to be mere assertions, handwaving without justification trying to imply the author somehow is apriori correct and won’t deign to bother explaining since it should supposedly be obvious to some who knows the law. Rather instead its likely the issue is they are leaping to conclusions too quickly and are suffering from the “illusion of knowledge” where they are assuming things that they haven’t questioned that are where the flaws are. Implicit leaps of logic and bait-and-switch between tools and their vendors, etc.

I’m guessing its because people take certain things for granted when dealing with humans and don’t think through carefully how the arguments differ. Subtle differences matter in a case like this, anthropomorphism seems to be implicit in some of the comments and not recognized.

Most of the replies I’ve gotten from other posters on this site fail to grasp that a full logical argument needs to be made without holes, rather than their constant unwarranted assertions. To hold these companies liable: the full chain of logic needs to be established and so the burden is on those who claim to be doing so. Hand waving at human related analogies questionably applied isn’t enough. There are fundamental holes in the arguments that people seem to not be spotting, perhaps studying those articles I linked may help.

The whole set of posts keeps making a number of assumptions that hand wave over differences between tools and humans in their attempt to apply case law analogies meant for humans. Several leaps of logic and hand waving need to be closed.

[cmcc_aus]

I'm an old software guy, who often worked with Beta versions. In order to take advantage and test out the newest features available, the caveat was always that something else (not limited to the newly-added features) could go wrong. "Watch out, and please report any unexpected results." But I would be at fault, if I ran the beta version in a production process, and did not verify all results.
Here, with the chatbot, if I, the user, 'publish' the false information, I cannot blame it on the tool. As you say about medical doctors using tools or following procedures, ultimately the fault is not with the tool, but with the practitioner.
But that does not mean we cannot also hold the tool, and its manufacturer, responsible, especially for 'reckless' aspects of the design.

[RealityEngineer]

Yup: the "design negligence" argument is the applicable one: which they note is also questionable, as have other posts on this site. Companies with huge resources have created programs with similar flaws since its inherent in the tech, and not easy to rectify. Its like cars, which risk hurting bystanders in accidents, since they don't spend the vast $ per car it would take to make them safer but not viable, or hold them off the market for decades for new safety tech.

[RealityEngineer]

oops, I meant some posts of comments by me and others have critiqued the design negligence claims . The posts by the blogger (or others commenting) haven’t make a strong case, instead indicating a lack of background knowledge of the tech thats needed to credibly even attempt such a thing.

[A nerdy Fred]

Have I misunderstood Professor Volokh? I was under the impression that if I am using ChatGPT and carrying out my duty of care and not passing along defamatory reports, even in that case defamation has happened just in the delivery of the false output to me.

[lucia_l]

A nerdy Fred,
Yes. That's my understanding too.
Moreover, even if you don't pass along the information, if you act on the information (e.g. not hiring a nanny you were investigating) without telling anyone what you read, the nanny who might have been hired might be harmed. You will not have defamed her--the defamation is entirely on the part of ChatGPT.

[RealityEngineer]

No, you understood his point: the point is that its flawed.

To provide an example to illustrate this issue: the reason Section 230 arose is partly because people had a fuzzy conception of the notion of “publisher”: and unfortunately still do. The intent was to not hold human agencies responsible for things they weren’t actually reviewing and couldn’t therefore be held responsibility for. Just as a bookstore owner couldn’t review all the books in their shop and take responsibility for the content.

Section 230 was an acknowledgement that the only human agent involved in the process of “publishing” content was the user and therefore the social media companies and search engines weren’t responsible since there was no human agency from them involved in the process.

Unfortunately the issue of agency again has arisen to confuse people because AI programs are more complicated and use the term “AI”.

The argument being made to hold AI responsible for “defamation” is pretending the AI has agency to be a party to a claimed act of “defamation”. But an AI has no agency, just as a search engine didn’t have agency. There is also no human from the AI vendor in the room with agency to be held responsible, just as there wasn’t in the case of humans using search engines or posting to social media.

The act of a user believing a false statement is claimed to be the harm involved: but the only human agent involved in that act is the user. If there was harm done: there can be a claim it was due to the program spitting out false information: but that claim would then be that the program shouldn’t have done so and therefore there is “design negligence”.

There is sort of a bait-and-switch shell game going on (unintentionally) trying to pretend an inanimate object tool had agency to try to instead apply defamation law: ignoring that there is an implicit “agency” aspect to that. But then of course the AI isn’t a human that can be sued, or people point out the AI isn’t an agent: and they just plug in the AI vendor in its place: but thats bait-and-switch. Its like pretending a gun manufacturer is responsible for the user of its products: the only agent involved in the use of the tool is the user of the tool, not the creator of it.

The problem is that there is no human agent that actually generates this new false content: so they are hunting around for one to blame it on and settling on one not in the room since they have deep pockets or something. The answer is to acknowledge they aren’t in the room to have been able to review or take responsibility for that content. So that legal framework isn’t the relevant one: the relevant one would be claims of design negligence: which are very questionable for different reasons.

[David Nieporent]

To provide an example to illustrate this issue: the reason Section 230 arose is partly because people had a fuzzy conception of the notion of “publisher”: and unfortunately still do. The intent was to not hold human agencies responsible for things they weren’t actually reviewing and couldn’t therefore be held responsibility for. Just as a bookstore owner couldn’t review all the books in their shop and take responsibility for the content.

You can say it 100 more times, troll bot, and it still will be a lie. The intent was not to hold human agencies responsible for things they were actually reviewing.

[RealityEngineer]

re: "You can say it 100 more times, troll bot, and it still will be a lie."

er, no it isn't. There were multiple reasons for why it was created. Regardless, the EFF, an entity with actual expertise on the topic explains:

https://www.eff.org/issues/cda230
"Section 230 embodies that principle that we should all be responsible for our own actions and statements online, but generally not those of others. The law prevents most civil suits against users or services that are based on what others say."

Regardless of whether some folks wanted it to allow moderation to happen, the core concept is, as another entity focused on tech policy explains:

https://itif.org/publications/2021/02/22/overview-section-230-what-it-why-it-was-created-and-what-it-has-achieved/
"This distinction between publishers, which are liable for the statements they circulate, and distributors—such as a bookstore or a newsstand—which are not, emerged from the Supreme Court case Smith v. California (1959). That case dealt with a Los Angeles city ordinance that criminalized the possession of obscene books. After bookseller Eleazar Smith was convicted of being in violation, he appealed on the grounds that the ordinance was unconstitutional. The Supreme Court agreed, stating that if ordinances such as Los Angeles’ were permitted, “Every bookseller would be placed under an obligation to make himself aware of the contents of every book in his shop. It would be altogether unreasonable to demand so near an approach to omniscience....

If, as the Supreme Court ruled in Smith v. California, it’s unreasonable to expect a bookseller to know the contents of a few hundred books, it’s even more unreasonable to expect the operators of a social media platform to know the contents of thousands, millions, or even billions of posts, even when using thousands of human moderators and advanced algorithms."

[RealityEngineer]

A law dictionary explains:

https://dictionary.law.com/default.aspx?selected=458
"defamation

n. the act of making untrue statements about another which damages his/her reputation."

Granted not all use that phrase "an act": but its taken for granted. The "act" implies an "actor" taking that act which is an actual human entity that has agency that can "act".

Again: the actual arguments on this topic need to be careful about what entities are involved taking agency and what are inanimate tools. An inanimate program can't be "negligent"

One site refers to:

https://www.law.cornell.edu/wex/negligence#:~:text=Definition,victims%20of%20one's%20previous%20conduct).
"Negligence
Definition
A failure to behave with the level of care that someone of ordinary prudence would have exercised under the same circumstances. The behavior usually consists of actions, but can also consist of omissions when there is some duty to act (e.g., a duty to help
victims of one's previous conduct). "

A program isn't a "someone" and it can't "behave" using the implication of a human behaving.

That other law dictionary starts it:

"negligence
n. failure to exercise the care toward others which a reasonable or prudent person would do in the circumstances, or taking action which such a reasonable person would not."

A program isn't a person who is involved in "circumstances" where they could be acting, and can't be "reasonable" since it doesn't reason in the human sense.

Maybe one day an AI will fall within the analogies people carelessly apply, but at the moment you folks need to engage in more careful reasoning. Its tiring dealing with people that seem to be completely unaware that they aren't making complete arguments and are handwaving. You need to question all your assumptions when you try to make analogies since most seem to not notice glitches in them. Like trying to pretend that "oh, the AI can't be acting so it must be the software vendor acting": er , no, in that case the implication is that its a software design negligence issue, which is a different argument which has also not been well made on this site so far.

[RealityEngineer]

Or another source, Cornell:

https://www.law.cornell.edu/wex/defamation “defamation …Elements

To prove prima facie defamation, a plaintiff must show four things: 1) a false statement purporting to be fact; 2) publication or communication of that statement to a third person; 3) fault amounting to at least negligence; and 4) damages, or some harm caused to the reputation of the person or entity who is the subject of the statement.”

A program can’t be “negligent” (at most it can be arguably negligence of design by its authors, which again is a different and weak case in this case).

The statement “publication or communication of that statement to a third person”

The phrase “third person” implies 2 other people.

Its implicitly describing 3 humans capable of acting as human entities with agency and awareness of the world. 2 who are communicating about a third person who is the subject of the statement. There is the user and a program: thats 1 person not 2. Hand waving away the need for 2 people to be involved in the communication needs to be justified vs. the mindless repetition “but thats how defamation works!” or whatever other replies I get. There are assumptions embedded in most of the comments here that aren’t being questioned, as if an AI program just can be substituted in for a person. Despite the use of the term AI: these aren’t entities like humans.

[David Nieporent]

Mindless repetition by a troll.

[RealityEngineer]

As the law review articles I posted should suggest: much of this whole issue turns on the question of agency and responsibility. Tools aren’t responsible for things, they lack agency. There are legal frameworks being applied that make the presumption of agency implicitly: and yet they are being applied to inanimate tools (then a shell game is played, presumably unintentionally, to swap out the tool for its maker when necessary, without justifying that swap).

I made an actual argument citing sources. It appears you can’t make any actual coherent argument and resort to juvenile mindless disparagement, which suggests you are acting like a troll. The burden of proof is on those that claim the legal system can punish these entities: and therefore you folks need to make full arguments and not handwave or engage in juvenile trolling.

[David Nieporent]

Tools aren’t responsible for things, they lack agency.

Troll: nobody is trying to hold a tool responsible. OpenAI is a corporation, not a tool. We are talking about holding the company offering the service liable, not the service itself.

[RealityEngineer]

The problem is in that train of logic. There is an attempt to pretend the tool can be held responsible based on case law regarding humans: but then "oops, it can't be held responsible, the vendor is".

Its that first step where there isn't recognition its a tool without agency. Its is a shell game pretending the tool and vendor can be used interchangeably in the reasoning process, which isn't logically justified. Its not a rigorous argument. Maybe reading the law review articles would help if it requires the right jargon to get lawyers who are struggling to grasp the problems with applying laws to tools without agency.

The relevant question would be design negligence, and again those that seem to know little about the tech have made flawed arguments. It seems like people need to learn more about these topics before trying to discuss them.

[RealityEngineer]

Yann LeCun, one of the pioneers of this type of AI, head of AI at Meta explains in a tweet:

https://twitter.com/ylecun/status/1643944294380609538 “Repeat after me: 1. Current Auto-Regressive LLMs are *very* useful as writing aids (yes, even for medical reports). 2. They are not reliable as factual information sources. 3. Writing assistance is like driving assistance: your hands must remain on the keyboard/wheel at all times”

and a clueless reply similar to those who just assume this is easy:

“Could you just, you know, fix it”

He responded:

” Yes, but that requires hard-core research, not just more data, more GPUs, and more RLHF duck tape.”

The design alternatives I’ve seen proposed on this site are “duct tape” and bailing wire, not seriously useful approaches.

There may be ways for 3rd party products to help validate against reality using other methods: and there is market inspiration for many people to work on the problem.

The incentive for people to find solutions will grow if *users* are held accountable for falsely believing content that they spread to others, rather than trying to pretend they can't be held responsible (perhaps in search of deep pockets).

This is a tool, just like alcohol is a tool for recreation, and cars are a tool for transportation. Humans are held responsible for harming 3rd parties in drunk driving accidents rather than the makers of the alcohol or the vehicle. Yes: there are design negligence cases, but the case for it in this instance is weak.

[RealityEngineer]

Humans imagine things that don’t exist as part of planning and creating new things: and part of the utility of these tools is that they imagine things also. Those things may be useful at times, not at other times, but they are never guaranteed to be fact. Unfortunately that inherent nature of them makes it difficult to ground things in reality. If it were trivial to automatically validate possibilities vs reality: the fact that humans get things wrong would have led to a tool to validate things by now.

The inherent nature of alcohol is that it impairs humans cognition: and thats one reason some people use it. Is it “design negligence” that it does so when thats why people use it? Its very nature has led people to grasp that the users of it are the entities responsible (with some minor exceptions where other humans are directly involved in their consumption and can influence potential aftermath).

One problem with the suggested band-aid, duct tape and bailing wire “design alternatives” is that it may give people the false impression the tools can be trusted when it won’t solve the problem. Its like the reality that many trust Wikipedia because its deceptively trustworthy on many topics but can still get things wrong and people’s guard has been let down. The whole point should be for people to grasp these tools can’t be trusted as factual and should validate things. People in general need to learn that information on the net needs to be validated.

The issue seems to be like having a program that generates Babylon Bee articles (they actually have an AI generator of AOC articles now): and wanting to view the public as apriori too dense to be allowed to see satire since they might believe it. So the “design alternative” is a hotline where they need to hear every misguided thing people took as true and add it into a neverending database that won’t prevent the next satire article by being misunderstood by someone not mentally competent. Perhaps it only generates Bablyon Bee articles sometimes and Not the Bee articles other times: but by default it should be assumed to be the Bee unless verified.

A Wharton prof had a useful article noting for lay people the inherent problem with these tools:

https://oneusefulthing.substack.com/p/how-to-get-an-ai-to-lie-to-you-in “How to Get an AI to Lie to You in Three Simple Steps I keep getting fooled by AI, and it seems like others are, too.”

If you use the proposed “design alternatives” in that paper to play whack-a-mole with individual examples: it’ll be a never ending process due to the nature of the tools where you can prompt them differently to get them to imagine things that don’t exist. By their very nature the current tools don’t have any grounding in reality in what is real and what isn’t. If you set the context talking about person X not by name but by a cultural reference “that orange dude we elected” (and there are myriad ways of using cultural references, perhaps in multiple steps narrowing in the context so no keyword will match all of them) you can’t rely on name searches. It can phrase things differently so you can’t rely on the claimed false statements using the same words.

A flawed partial approach to getting rid of some of the flawed statements may lead people to let their guard down. Its like people having the mistaken impression the government protects them from all medicines that don’t work: so they are too trusting of quack products like homeopathic treatments.

The effort to squash recreational drugs due to that impairment issue has to many arguably done more harm than good. It seems like one potential opportunity cost is: if it were legal perhaps people could invent safer drugs: for instance ones where you can take a different pill that reverses the impairment. There isn’t going to be R&D from legal companies on that now (and the fact that it doesn’t exist with alcohol yet doesn’t mean it won’t: but more importantly perhaps there is an easier approach with different chemicals than that one).

The problem of validating information against reality is one that tech folks are working on: but its not an easy fix. A vast number of people want these tools: so they aren’t going away. The issue is how to reconcile that reality vs. perhaps creating a legal framework that does more harm than good. For instance by either directly undermining startups who can’t afford the legal expenses, or through requiring laws to be passed that protect AI: but may be prone to regulatory capture that again protects big players and squashes the innovative startups that might improve things.

It may slow the process of creating better things to fix these problems. Worst case if it raised liability costs and squashed the US industry (as happened arguably with small planes before): people will use lower quality AI from overseas (via VPN: word of those would spread virally just as word of ChatGPT did) or via home versions that are less capable and more likely to generate flawed information. The misguided attempt to address harms may generate more harm than good. It may generate more harm than good by slowing down innovation and preventing the positive things AI can do, even other sorts of AI that are indirectly aided by this through the rise of better hardware for all AI, etc.

[RealityEngineer]

re: "If you use the proposed “design alternatives” in that paper "

Oops, that was poorly phrased and can't edit, the referent was the paper being discussed on this site, not the blog post by the Wharton prof.

[RealityEngineer]

In case anyone is still reading, AI pioneer Yann LeCun retweeted this, noting its not surprising:

https://twitter.com/AlainGoudey/status/1644680576102719489 ” Le ban de ChatGPT en Italie fait exploser les recherches Google sur les #VPN! ”

which translates to:

“ChatGPT ban in Italy explodes Google searches on #VPN !”

Efforts to shut down access to AI if litigation squashes legal use of it in the US (or in some other country) aren’t going to go well: and again if the US is sidelined it seems likely it’ll lead to people using lower quality products with a higher percentage of misinformation.

Again its like the drug war and alcohol prohibition before that. Again the parallel to those is that user should be held responsible for any harms, which will help teach other users to be careful.

[RealityEngineer]

Economics prof Bryan Caplan points out the issue of the litigation industry doing what governments wish to do but can't. Libertarians need to be concerned about civil law squashing innovations like AI just as much as the government squashing them:

https://betonit.substack.com/p/lawsuits-are-the-deep-state
"Lawsuits Are the Hitman of the State

...

The government starts with the blatantly illegal goal of banning “bigots from expressing their opinions in a way that abuses or offends their co-workers.” Then instead of respecting those limits, the government’s judicial branch gets creative: “Murder’s illegal? Fine, we’ll hire hitmen instead.” By affirming liability, it dangles piles of cash in front of potential plaintiffs to terrorize employers into banning what the government, legally, must allow."

Unfortunately much of the public doesn't reason very clearly and if the legal world is pointed off in a direction that'll squash AI they may be able to lead juries to along, and academics lead judges to also. Its rather useful for the world if early work regarding legal issues related to AI isn't rushed into print to be at the leading edge, but instead takes time to be well considered.

[Brett Bellmore]

For instance, August 30th,2018

"AUG 30 2018
“Just this week, we made another truly historic announcement. We are replacing NAFTA with a beautiful, brand-new U.S.-Mexico trade deal.”

Trump announced an informal trade agreement with Mexico to replace NAFTA. But he needs Congress to approve any change to NAFTA, and that hasn't happened. The U.S. is still negotiating with Canada. So it's inaccurate to claim Trump "signed a trade agreement with Mexico.""

Of course, he didn't say Congress had ratified it. He said he'd signed it.