Large Libel Models: The CheckBKG Analogy

To better understand the debate about possible defamation liability for OpenAI, based on its Large Libel Models tendency to sometimes communicate entirely made-up quotes about people—supposedly (but not actually) drawn from leading media outlets— let's consider this hypothetical:

Say a company called OpenRecords creates and operates a program called CheckBKG, which does background checks on people. You go to CheckBKG.com, enter a name, and the program reviews a wide range of publicly available court records and provides a list of the criminal and civil cases in which the person has been found liable, including quotes from relevant court records. But unfortunately, some of the time the program errs, reporting information from an entirely wrong person's record, or even misquoting a record. CheckBKG acknowledges that the information may be erroneous, but also touts how good a job CheckBKG generally does compared to ordinary humans.

Someone goes to CheckBKG.com and searches for someone else's name (let's say the name Jack Schmack, to make it a bit unusual). Out comes a statement that Schmack has been convicted of child molestation and found liable in a civil case for sexual harassment, with quotes purportedly from the indictment and the trial court's findings of fact. The statement accurately notes Schmack's employer and place of residence, so readers will think this is about the right Schmack.

But it turns out that the statements about the court cases are wrong: The court records actually refer to someone entirely different (indeed, not someone named Schmack), or the software missummarized the court records and wrongly reported an acquittal as a conviction and a dismissal of the civil lawsuit as a finding of liability. The quotes are also entirely made up by CheckBKG. It also turns out that Schmack has informed OpenRecords that its software is communicating false results about him, but OpenRecords hasn't taken steps to stop CheckBKG from doing so.

It seems to me that Schmack would be able to sue OpenRecords for defamation (let's set aside whether there are any specialized statutory schemes governing background checks, since I just want to explore the common-law defamation tort here):

1. OpenRecords is "publishing" false and reputation-damaging information about Schmack, as defamation law understands the term "publishing"—communication to even one person other than Schmack is sufficient for defamation liability, though here it seems likely that OpenRecords would communicate it to other people over time as well.
2. That this publication is happening through a program doesn't keep it from being defamatory, just as physical injuries caused by a computer program can be actionable. Of course, the program itself can't be liable, just as a book can't be liable—but the program's developer and operator (OpenRecords) can be liable, just like an author or publisher can be liable.
3. OpenRecords isn't protected by 230, since it's being faulted for errors that its software introduces into the data. (The claim isn't that the underlying conviction information in court records is wrong, but that OpenRecords is misreporting that information.)
4. OpenRecords' noting that the information may be erroneous doesn't keep its statements from being defamatory. A speaker's noting that the allegation he's conveying is a rumor (which signals a risk of error) or that the allegation he's conveying is contradicted by the person being accused (which likewise signal a risks of error) doesn't keep the statements from being defamatory; likewise here.
5. OpenRecords now knows that its software is outputting false statements about Schmack, so if it doesn't take steps to prevent that or at least to diminish the risk (assuming some such steps are technologically feasible), it can't defend itself on the grounds that this is just an innocent error.
6. Indeed, I'd say that, OpenRecords might be liable on a negligence theory even before being alerted to the specific false statement about Schmack (if Schmack isn't a public official or public figure), if Schmack can show that it carelessly implemented algorithms that created an unreasonable risk of error—for instance, created algorithms that would routinely make up fake quotes, in a situation where a reasonably effective and less harmful alternative was available.

If I'm right on these points, then it seems to me that OpenAI is likewise potentially liable for false and reputation-damaging communications produced by ChatGPT-4 (and Google is as to Bard). True, CheckBKG is narrower in scope than OpenAI, but I don't think that matters to the general analysis (though it might influence the application of the negligence test, see below). Both are tools aimed at providing useful information—CheckBKG isn't, for instance, specifically designed to produce defamation. Both may, however, lead to liability for their creators when they provide false and reputation-damaging information.

I say "potentially" because of course this turns on various facts, including whether there are reasonable ways of blocking known defamatory falsehoods from ChatGPT-4's output (once OpenAI is informed that those defamatory falsehoods are being generated), and whether there are reasonable alternative designs that would, for instance, prevent ChatGPT-4's output from containing fake defamatory quotes. But I think that the overall shape of the legal analysis would be much the same.

Comments

[QuantumBoxCat]

Say a company called OpenRecords creates and operates a program called CheckBKG, which does background checks on people.

It also turns out that Schmack has informed OpenRecords that its software is communicating false results about him, but OpenRecords hasn’t taken steps to stop CheckBKG from doing so.

Sounds like CheckBKG meets the statutory definition of a “consumer reporting agency” under the Fair Credit Reporting Act. As such, OpenRecords has certain statutory obligations that it must follow, and if it fails to do so, Schmack may be able to pursue litigation under the FCRA. It seems worth noting that, to my understanding, the FCRA specifically bars defamation, invasion of privacy, and negligence claims that concern the reporting of information from being brought against any consumer reporting agency, any user of a consumer report, or any furnisher of reported information, except for false information furnished with malice or willfully intended to injure the consumer.

[John F. Carr]

See TransUnion v. Ramirez. The credit report said “input name matches name on the OFAC database.” Your customer has the same name as a terrorist or other evil person. The subject of the report was not the person on the OFAC list. The statement was literally true but still actionable, weasel words notwithstanding. Under the particular legal scheme set up by the FCRA.

[JJH3]

Prof. Volokh’s hypothetical changes the facts from the real world in a way that changes the result. OpenAI is not a credit reporting agency, and its terms and conditions explicitly prohibit any use cases with a high risk of economic harm, including “Automated determinations of eligibility for credit, employment, educational institutions, or public assistance services.” Far from promoting its service as offering “background checks on people,” OpenAI prohibits the very actions that Prof. Volokh hypothesizes.

Regarding point #4, it seems unlikely that courts would disregard OpenAI’s clear statements, which users must agree to before using the service, that the model may generate false information and should not be considered a factual source. As for point #6 (the negligence theory), there isn’t a “reasonably effective and less harmful alternative” that OpenAI is neglecting. The only option would be training the AI to avoid mentioning individuals at all, which would limit its usefulness and still not eliminate the risk that Prof. Volokh is concerned about. Judges are unlikely to adopt an impossible-to-satisfy standard of care that could hinder AI development in a particular jurisdiction.

I would note the theory being discussed has limited real-world applications. Only users who input specific prompts would see the supposedly defamatory statement (and OpenAI’s terms require users who republish model results to include disclaimers, so it’s unlikely OpenAI could be held liable for some user’s decision to republish to a wider audience). My guess is that most product manufacturers face a lot more product liability lawsuits for toasters that catch on fire or cars where the brakes failed than OpenAI will see libel lawsuits.

[David Nieporent]

Congratulations on repeating a bunch of talking points about the law that were already refuted!

[Stephen Lathrop]

QuantumBoxCat, what constitutes malice? Is it knowledge that the report is false? If the credit reporting agency relies on a trusted source which makes a mistake, I suppose you are saying the credit reporter is off the hook. But if the mistake is then pointed out, can the credit reporter just say, “Sucks to be you?,” lower your credit rating, and continue with misreports?

[David Nieporent]

As Prof. Volokh alluded to, consumer credit reporting agencies are subject to special rules, so they’re not a great way to look at this. But setting those special rules aside and discussing regular defamation law, malice is not required. First, it’s a matter of private concern, not public concern, and second, a credit agency is not a media entity. See Dun & Bradstreet v. Greenmoss Builders — a case involving… a credit agency. (Though not a consumer one.)

[Brett Bellmore]

“CheckBKG acknowledges that the information may be erroneous, but also touts how good a job CheckBKG generally does compared to ordinary humans.”

This raises a common issue with computerizing a task formerly or also done by humans: The computer may actually be doing a better job than the humans, on average, while still making mistakes. Just fewer mistakes.

BUT, the mistakes typically will be different mistakes than a human would make!

Humans are not required to be perfect, and probably machines shouldn’t be, either, and we don’t want to discourage rolling out systems that actually make fewer mistakes. But people are going to be rather unsympathetic when the self-driving car mistakes a mural on a brick wall for an open road, and drives straight into it, even if that car is actually racking up a better record than a human would, on average, and saving people’s lives in situations where a human driver wouldn’t have.

[Life of Brian]

But people are going to be rather unsympathetic when the self-driving car mistakes a mural on a brick wall for an open road, and drives straight into it, even if that car is actually racking up a better record than a human would, on average, and saving people’s lives in situations where a human driver wouldn’t have.

Right. For years self-driving proponents have been arguing aggregate statistics and comparing self-driving fails to drunk/distracted/tired drivers. But that was supposed to be one of the major selling points of the technology — that they’re not subject to distractions/fatigue/impairment like humans are. And we can do tangible things to help reduce the number of drunk/distracted/tired drivers on the road, while spectacularly ridiculous AI fails (e.g, repeatedly trying to drive “under” semi trailers or suddenly veering toward a cyclist in a bike lane as though for bonus points) “just happen” and thus all we can really do is clean up the messes when they do.

[Life of Brian]

Also: drunk/distracted/tired drivers tend to exhibit patterns of behavior (weaving, erratic starts/stops, etc.) over a period of time, which alert defensive drivers have a chance to see and choose to steer clear from, as well as minimizing overall risk by avoiding major routes on weekend nights etc. It’s a whole different world if every vehicle I pass 24×7 has the potential to just whip over and plow into me out of the blue and for no knowable reason.

[Brett Bellmore]

And my point is, we SHOULD be looking at the aggregate statistics, because that’s what you want to focus on if you want to save lives!

Focusing obsessively on the rare weird fails is irrational. People are irrational, so they’ll do it anyway, but we SHOULDN’T obsess about rare failure modes. We should adopt the automated systems as soon as they’re better than the human ones, and then treat each rare failure mode as a learning opportunity to IMPROVE them.

[Life of Brian]

Focusing obsessively on the rare weird fails is irrational. People are irrational, so they’ll do it anyway, but we SHOULDN’T obsess about rare failure modes. We should adopt the automated systems as soon as they’re better than the human ones, and then treat each rare failure mode as a learning opportunity to IMPROVE them.

I get the sentiment and the math is the math, but I think there’s also a less-obvious cost to aggregate mental health due to the feeling of loss of control that needs to be stacked up against “lives saved” in a vacuum.

Let’s say we could deploy Boston Dynamics dog-bots with guns on their backs and running some sort of Minority Report like mindreading tech that would chase down and kill 95% of murderers before they could act. BUT, the algorithms very occasionally misfire and the dog-bots just shoot completely random people walking down the street. I don’t know that the average person would (or should!) feel safer as a result, no matter how much trumpeting of the raw math and potential for improving the algorithms so maybe they’ll randomly kill FEWER people.

Per my second comment above, current accident statistics are somewhat clustered in time/geography/etc. and often happen at the tail end of protracted and visible bad driving, and people often make choices accordingly to try to minimize their own personal risk exposure. A completely random distribution takes that off the table.

[Brett Bellmore]

“but I think there’s also a less-obvious cost to aggregate mental health due to the feeling of loss of control that needs to be stacked up against “lives saved” in a vacuum.”

Ah, but rare events are, by definition, “rare”. That helps a bit, especially if the media aren’t deliberately giving the impression they’re common.

[bernard11]

Per my second comment above, current accident statistics are somewhat clustered in time/geography/etc. and often happen at the tail end of protracted and visible bad driving, and people often make choices accordingly to try to minimize their own personal risk exposure. A completely random distribution takes that off the table.

This is an interesting point. On the whole I lean towards Brett’s position here but this is worth thinking about.

The implication seems to be that drivers assume certain risks, and the more risk-averse get into fewer accidents, so it’s not quite fair to compare aggregates.

But it’s worth knowing how much benefit there is to some risk-averse practices. It’s also worth knowing what abandoning those practices as EV’s become common is worth. Maybe someone doesn’t drive late at night on the weekend, but would like to do that sometimes. Is the benefit then worth the possibly elevated risk of daytime driving?

My guess – all it is – is that the EV’s come out ahead.

[Stephen Lathrop]

The baffling question is what net effect to expect if a theoretically perfect accident avoider is compelled to accept EV-based liability standards. If that happened, the perfect driver would increase his personal probability to cause an accident, because EV driving is less than perfect.

But because of his perfection, absent EVs, essentially every accident which involved the perfect driver would be the result of bad driving by worse drivers, who would be conventionally at fault. Nearly-perfected EV vehicles might thus reduce net risk to the perfect driver, by improving the driving results of other less proficient drivers. Or, they might not.

I suspect the chance that any of that will come to a test in real life must await an AI solution to how to manage EV driving in a pedestrian-rich urban situation. That might have to be a solution based on a network of mutually interacting AI devices carried by all vehicles and all pedestrians. Otherwise, how will AI driving accommodate pedestrian unpredictability, except by stopping every time it cannot tell what a pedestrian ahead is about to do—which would presumably deliver perpetual gridlock.

[Brett Bellmore]

“That might have to be a solution based on a network of mutually interacting AI devices carried by all vehicles and all pedestrians.”

Not a good idea, because you’d be doing it in a larger environment where some of the people are actively attempting to break the system, whether for pranks, or out of darker motives. So, for instance, you’re in a line of cars at a stop light. The light turns green, usually the first car moves, then the second, and so forth; Only the first car is directly acting on the light.

But if the cars are communicating, somebody can inject a spoofed signal, and you get rear-ended because the car behind you thought you’d started moving! You can see other scenarios.

Usually we deal with the ‘pedestrian rich’ situation by separating the pedestrians from the cars. The situations where it’s unavoidable, we normally keep speeds so low that stopping distances are very short, and the consequences of collisions reduced.

[Stephen Lathrop]

Bellmore, seems like if you train the cars to act without input from the pedestrians, that does nothing to reduce chances for ill motivated people to break the system. More the opposite. Any pedestrian who wants to cross any street can just step toward a curb and traffic on that street grinds to a halt, because what else can the car do? If the pedestrian does that not because he wants to cross the street, but only because he Iikes to look at stopped cars, or ogle the attractive women driving them, same result, but more continuous.

[Brett Bellmore]

And what happens today, in a world of humans driving cars, if you walk into traffic? Oh, yeah: Traffic grinds to a halt…

Just standing there, though? No reason the car can’t treat the standing pedestrian the same way a human would: Keep an eye on them, but don’t act unless they actually start across the road.

[Dr. Ed 2]

It would not be surprising for jack to commit suicide and what is the liability for that?

[Frank Drackman]

Who shot Jack?? Jack, so it’s Jack’s fault

[M L]

Somewhat related,

Clearview AI scraped 30 billion images from Facebook and gave them to cops: it puts everyone into a ‘perpetual police line-up’

https://www.businessinsider.com/clearview-scraped-30-billion-images-facebook-police-facial-recogntion-database-2023-4

[DaveM]

Could these legal risks be mitigated by placing the results of the background check under human review? I’m thinking of a workflow like this:

1. Customer requests a background search on Schmack.
2. CheckBKG does its thing, but reports it to a human reviewer who works for OpenRecords.
3. Reviewer makes final determination that the results are legit
4. Customer gets their answer.

[OtisAH]

You should give Turley a call. He’s always by his phone. He claims ChatGPT is accusing him of sexual assault.

[A nerdy Fred]

Professor Volokh has already reached out via email, according to this piece: https://www.usatoday.com/story/opinion/columnist/2023/04/03/chatgpt-misinformation-bias-flaws-ai-chatbot/11571830002/

[John F. Carr]

“The court records actually refer to someone entirely different”

In driver’s license record keeping there is an expectation that the combination of name and birth date is unique. People do get hit with others’ misdeeds as a result. The old solution was unique ID numbers. The new solution is biometrics. Either way, mistakes can still be made.

[Frank Drackman]

So you’re saying Ted Kennedy DIDN’T leave a young woman to asphyxiate (not drowned, there’s a difference) in July 1969??

[apedad]

Just like news media publish erratum (not mandatory AFAIK), for sites like “CheckBKG.com” that publish publicly available court records, they should have a function to make corrections or refute info, i.e., have an established process where people can rebut the published info (this should be mandatory!).

Additionally, once the info is corrected, it should be mandatory for the sites to publish the erratum.

[Tom Tildrum]

Imagine that OpenAI lied about inventing ChatGPT, and instead it really just has 100 people in a room scouring the internet, having been told “do your best to come up with answers.” If one of those people fabricated a newspaper article accusing someone of a crime, I see no reason why OpenAI would not be liable for defamation. Why would that change if it’s actually a computer program that they wrote, instead of 100 people?

[RealityEngineer]

If Microsoft Excel or Google Sheets made an error in a calculation that caused harm: the issue would presumably be “design negligence”. This seems this is a question of “design negligence” in what sounds like a simple database program (with much more complicated analysis in the question of claimed design negligence for an AI).

The vendor of the software publishes the software, not the output of the software. It seems this may be a case of “if the only tool you have is a hammer, everything looks like a nail”. A usual conceptual tool used related to legalities of content is to find its “publisher”: so the assumption is there must be one. Every example of a piece of text is looked at as if there must be one: but perhaps there isn’t, or at least not until the human tool user communicates the created content to some other human.

re: “That this publication is happening through a program doesn’t keep it from being defamatory”

Its unclear the output of a tool is “published”. A law dictionary site:

https://thelawdictionary.org/publisher/
“PUBLISHER Definition & Legal Meaning
One whose business is the manufacture, promulgation, and sale of books, pamphlets, magazines, newspapers, or other literary productions.”

The program isn’t a “one”, it isn’t an entity. It is a tool that is displaying the result of a computer process. The intent of Section 230 was to make clear that the outputs of these social media programs doesn’t make the vendor a “publisher”. It seems the intent of the term “publisher” is to apply to a human entity that reviews content and takes responsibility for it. Merely because content exists doesn’t mean there is a publisher of that content.

Other definitions of publisher talk about communicating the content to at least one other person: but that act of communication seems to imply a human as the source of that communication. The act of publication seems to require communication from one human to at least one other human: in which case the human that causes that to happen would seem to be the “publisher”.

The copyright office says:

https://www.copyright.gov/comp3/chap1900/ch1900-publication.pdf
“What Constitutes Publication?

Although it is not expressly stated in the statutory definition, the legislative history indicates that publication occurs only (i) when copies or phonorecords are distributed by or with the authority of the copyright owner, or (ii) when an offer to distribute copies or phonorecords to a group of persons for further distribution, public performance, or public display is made by or with the authority of the copyright owner”

Granted thats only its use in one realm. Given recent decisions by the copyright office, the output of even an AI tool isn’t copyrightable at the moment so there is no copyright owner: but many seem to hope to persuade them the tool user is the copyright owner, in which case publication only occurs when the content goes beyond the user.

Even if they aren’t the copyright holder, it would seem the only relevant human able to be the source communicating to other humans would be the human that caused the content to come into existence through the use of a tool.

[Noscitur a sociis]

“Publication” is a statutorily-defined term of art for purposes of the Copyright Act. That definition is not the same as the “publication” required as an element of a defamation claim. Nor, of course, need a defamer be a “publisher”, in the sense of someone involved in the publishing industry.

[RealityEngineer]

I noted “Granted thats only its use in one realm”. The problem is that most definitions don’t seem to be well specified. However they tend to share the assumption you make when you then write “Nor, of course, need a defamer be a “publisher”, in the sense of someone involved in the publishing industry.”

You use the word “someone”. Almost all of the discussion implicitly, even if not explicitly, seems to revolve around the intent that a human who has agency takes responsibility for some particular content.

I posted on a prior page from this weekend a law review article discussing the issue of agency. Most of the discussion on this topic seems to be implicitly trying to pretend an AI program is somehow an entity with agency rather than just an inanimate tool like any other. It can’t commit libel since it can’t be “negligent”: only a human can.

The human agent involved is the user of the tool who has responsibility for the results. At most they might try to claim that there is “design negligence” which is also a complicated issue and highly questionable in a case like this when companies with vast resources come up with tools with the same type of flaws. That suggests that it is a complication of the nature of this tool rather than a design flaw. Its a limitation users need to accept when they choose to use these tools.

[David Nieporent]

Most of the discussion on this topic seems to be implicitly trying to pretend an AI program is somehow an entity with agency rather than just an inanimate tool like any other.

Nope. Nobody is saying that; you’re just a liar. The AI company is the entity with agency.

[Stephen Lathrop]

Nieporent, I thought he was saying the party managing publication decisions involving AI output was the entity with agency—which would deliver impunity to AI companies, unless they published output themselves. And then he noticed that defamation law would make that model all-but-useless, except to support AI research which never resulted in disclosing AI output to any third party. So his proposed solution is to let AI rip, under universal and arbitrary impunity for AI, with society at large being forced to adjust, come what may.

He is an all-but-utopian optimist about what AI will do someday, and cares not a bit for any damage AI does before someday arrives. His program for that is: force the public to accept whatever damage ensues, and hope for heaven later. He has given no thought to the possibility that too much prompt social damage—or damage to the public life of the nation—might postpone someday permanently. His advocacy seems unwise.

[RealityEngineer]

If you print out something created with Microsoft Word, are they the publisher? That’d be news to lots of people. Unfortunately much of the discussion on this serves to obfuscate the reality that this is a computer too. Its avoiding making the case of why this tool is magically different, and arguing what class of tools is different and why.

[Michael P]

“The vendor of the software publishes the software, not the output of the software.”

They’re providing the output through their web page. They are publishing it.

Please stop throwing uncooked copypasta at the wall in hopes that some of it will stick.

[RealityEngineer]

You make an assertion that isn’t well reasoned and doesn’t address the points. You say “they’re providing the output through their web page”. So do social media sites and search engines: but they aren’t viewed as the publishers. You make no argument as to how this is different. I suspect the issues are too nuanced and subtle perhaps for some. Check the prior page where I posted some law review articles, perhaps its necessary for legal types to have it explained using legal jargon they are used to.

If a user uses the tool of a social media site to post content: the user is in essence the publisher. They are the only human involved in that action. When the user users the tool described above: they are the only human involved in that action: arguably they are the publisher.

I’d suggest perhaps some legal types might benefit from taking logic classes in math and philosophy and some computer science classes where they need to deal with needing to fully deal with all the logic involved in solving a problem or there is a bug.

[David Nieporent]

You make no argument as to how this is different.

Social media sites and search engines don’t create content. AI companies do. QED.

[A nerdy Fred]

A severely stretched but understandable case could be made that the output of an LLM is the result of the prompt from the user and there is no other agency involved.

One problem with that idea out of many goes back to Professor Volokh’s original article on the subject. What he asked for was factual information with citations. What he got instead, he was not responsible for.

[RealityEngineer]

re: “there is no other agency involved”.

An AI doesn’t have agency, despite the confused anthropomorphism involved in attempting to imply it does. As I posted before, in case people need to see it from a law review:

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″

Another law review article, this time from a UCLA assistant prof:

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.”

[David Nieporent]

An AI doesn’t have agency, despite the confused anthropomorphism involved in attempting to imply it does.

Actually, you’re the only one claiming it does.

[RealityEngineer]

You explicitly stated: “Social media sites and search engines don’t create content. AI companies do. QED.”

The “create content” comment was used in a way that implied agency on the part of the AI rather than its use purely as a tool like Microsoft Word. You folks seem to imply differences between how Microsoft Word is treated where the user is the content creator, or an illustration program where an artist creates content, and a switch to where somehow the AI is the content creator.

You folks make arguments that imply certain things that are required to make your point, but then try to pretend the implication isn’t there. The fuzzy logic here isn’t too useful when dealing with an issue that is going to turn on clearly thinking through subtle distinctions, which apparently isn’t the forte of some of the posters here.

[RealityEngineer]

So photoshop is creating content and Adobe is the publisher of all the images? You also confuse the software with the entity that publishes the software. You folks seem to have trouble with the devil being in the details that you just handwave over and don’t even seem to grasp you are doing it.

[David Nieporent]

So photoshop is creating content and Adobe is the publisher of all the images?

No. That’s stupid. Nobody reading in good faith could think that.

You also confuse the software with the entity that publishes the software.

No. That’s stupid. Nobody reading in good faith could think that.

[RealityEngineer]

quoting my statement that “You also confuse the software with the entity that publishes the software.”

you claim re: “No. That’s stupid. Nobody reading in good faith could think that.”

Yet a reply above explicitly states: “They’re providing the output through their web page. They are publishing it.”

The things you claim are “stupid” are implicit assumptions being made in posts here. Yes: they are questionable: which is why I challenged them. Yet people provide no alternative as to why these AI tools are magically different from from photoshop or word and somehow their vendor has become the “publisher”. Its like dealing with the South Park underpants gnome meme where people skirt around the fact that they are making incomplete arguments and handwaving at things.

[Stephen Lathrop]

You say “they’re providing the output through their web page”. So do social media sites and search engines: but they aren’t viewed as the publishers.

RealityEngineer, more accurately, social media sites and search engines enjoy benefits of a legal fiction, which pretends that social media sites and search engines need not be treated as the publishers they are, for liability purposes only. You demand addition to that pretense, so the same legal fiction can be extended to encompass AI output however published, and no matter who publishes it.

The challenge your advocacy has so far evaded is to say why doing as you demand will not sooner blow up the pretenses created by Section 230, than enable the near-utopian benefits you predict for AI. You do not seem to notice that evolution of public response to Section 230 has during recent years become continuously grumpier—despite near-fanatical agreement among internet fans that Section 230 needs to be saved—but with mutually exclusive, self-serving proposals for means to save it frustrating every discussion. Internet fans have yet to put 2 and 2 together, and notice that internet occurrences they loathe are inevitable consequences of Section 230, which they reckon an unalloyed boon.

As a matter of practical reality—as opposed to mistaken public opinion (advocacy at cross purposes shows near-universal acceptance that critical mistakes have happened, even though no agreement on which mistakes is in sight)—that means Section 230 is already under plenty of stress. I doubt it will accommodate the giant extra load of stress that AI seems capable to inflict if it too gets sheltered under Section 230’s umbrella of pretense.

I am someone who has long advocated unconditional repeal of Section 230—both because I think it has too much damaged the public life of the nation, and also because it works against optimal use of internet-delivered publishing efficiencies. Perhaps in that style of advocacy I should cynically encourage AI-related demands for the same legal-fiction treatment social media gets, as the straw which would break the Section 230 camel’s back.

Problem is, what happens if Section 230 doesn’t collapse right away? Then everything just gets exponentially worse for as long as it takes for internet utopians to learn by experience the lessons they refuse to be taught quicker by observation and reflection. At present, utopian hopes seem nearly undiminished. Thus a long interval of continued pro-pretense advocacy apparently lies ahead.

Not very far down that road, I expect the nation to encounter existentially threatening public chaos. Politics is the means American constitutionalism prescribes to manage public affairs. Without firm consensus on standards for public discourse, politics loses capacity to match its challenges.

The legal fiction insisted on by Section 230 has diverted attention from actual occurrences, and actual vices, which drive politics now. A public increasingly uninformed (or misinformed) about everything loses touch with standards for public discourse. It becomes progressively less competent to manage politics.

In some cases the remedy for bad speech is rightly reckoned to be more speech. It is never true that the remedy for bad confusion is more confusion. A legal fiction deployed to license an AI technology prone to publish falsehoods in unlimited profusion is the opposite of today’s critical political need.

Your advocacy is unwise.

[Michael P]

The fact that you don’t understand the laws in question does not mean my comment “isn’t well reasoned and doesn’t address the points”.

I am an engineer, I apparently just continue to overestimate your ability to educate yourself.

[RealityEngineer]

re: “The fact that you don’t understand the laws in question”

i.e. you can’t make any actual argument and merely confuse blustering assertion with doing so. You’ve not given any evidence you have some actual understanding of the laws in question or can construct a relevant argument. Most engineers learn that hand waving isn’t useful.

[Michael P]

I identified the sufficient basis for what OpenAI or Google are doing to qualify as “publishing” defamatory material, rebutting your assertion. The fact that you don’t realize why that is sufficient basis is a reflection of your ignorance, which others have been kind enough to address in more words, rather than of my argument.

The only thing you have engineered is your own rectocranial inversion.

[RealityEngineer]

re: “I identified the sufficient basis for what OpenAI or Google are doing to qualify as “publishing” defamatory material,”

You never remotely did any such thing. Merely asserting something doesn’t make it so. You provide no argument as to how these tools are different from social media sites which are not the publishers of users content, nor software like photoshop where users would be rather upset if Adobe claimed to be the publisher, and MS Word where Microsoft would be upset if they were held to be the publisher of problematic content its users create.

Those who deal with complex software systems and the world of AI need to learn to deal with subtle logic points and carefully made arguments, and so do those who hope to apply legal ideas to such systems. All I keep getting are hand waving comments that seem oblivious to the idea that they haven’t actually made a full argument thats consistent.

[RealityEngineer]

For whatever its worth, here is Bing AI’s response to:
“pretend you are a law professor. write a counter argument to this post.”

here:
“This is a response based on web search results. A possible counter argument to this post is:

The author of this post assumes that OpenAI and Google are analogous to OpenRecords, a company that operates a background check program that sometimes produces false and defamatory statements about people. However, this analogy is flawed for several reasons.

First, OpenAI and Google do not provide factual information about specific individuals, but rather general conversational or creative content that may or may not involve real people. Unlike CheckBKG, which claims to review court records and provide quotes from relevant sources, ChatGPT-4 and Bard do not purport to be authoritative or reliable sources of information about anyone. They are merely tools for generating text based on user input and large datasets of web data.

Second, OpenAI and Google do not have the same level of control or responsibility over the output of their AI models as OpenRecords does over CheckBKG. OpenRecords is directly involved in creating and operating CheckBKG, and has access to the underlying data and algorithms that produce the output. OpenAI and Google, on the other hand, do not control how users interact with their models, what inputs they provide, or what outputs they receive. Moreover, they do not have access to the internal workings of their models, which are based on complex neural networks that are not easily interpretable or modifiable.

Third, OpenAI and Google have taken steps to mitigate the potential harms of their AI models, while OpenRecords has not. OpenAI has provided a content filter that aims to detect and flag unsafe or sensitive text generated by ChatGPT-4, and has recommended that users do not return such text. Google has also implemented a content filter for Bard, and has limited its access to a select group of researchers and developers. OpenRecords, by contrast, has ignored the complaints of Schmack and other victims of CheckBKG’s errors, and has continued to operate its program without any safeguards or warnings.

Therefore, OpenAI and Google are not liable for defamation for the false and reputation-damaging statements produced by their AI models, as they are not publishers of such statements, nor are they negligent or reckless in creating or operating their models. The author’s hypothetical scenario does not apply to their situation.
“

[Michael P]

“OpenAI and Google do not provide factual information about specific individuals”

If they don’t, they certainly pretend they do.

Kindly stop flooding us with such ridiculously wrong theories.

[Brett Bellmore]

No, if you dig into the disclaimers deep enough they do tell you. But they’re kind of equivocal about it.

[Michael P]

Such disclaimers don’t carry much legal weight at all, when they’re pushing the models very hard as being useful in the real world rather than solely for composing fiction.

[RealityEngineer]

There is a common phrase that most in the computer science world are familiar with “all models are wrong. Some are useful”. It seems you are struggling with that simple concept as the one that these companies are pushing. They don’t state that these get 100% on each and every test: so anyone who thinks they aren’t wrong has rather poor critical reasoning skills and is making a completely unwarranted assumption (and literally unwarranted since the terms of service provide no warranty for OpenAI to provide factual information).

[Michael P]

The existence of an aphorism does not remove liability. “Caveat emptor” does not excuse fraud, and EV has already explained why slapping a disclaimer on a product or service advertised for certain purposes does not excuse defamation.

[RealityEngineer]

Arguments were made regarding human generated content: this wasn’t generated by any human except the user of the tool. Yes, if the user distributes the problematic content then a defamation argument has been made. If they don’t: there hasn’t been a full argument made, merely handwaving where these programs are somehow implied to be apriori agents capable of being held responsible in the way humans are. Perhaps you folks can get an AI to explain the problems to you or perhaps reading the law review articles several times might lead to you to start to grasp relevant concepts and distinctions if you need them couched in the jargon lawyers are used to.

[RealityEngineer]

OpenAI’s terms of service and the popup you close to get access to the chatbot explains they don’t provide factual information. I just opened the Bard page and it explicitly states ” I have limitations and won’t always get it right, but your feedback will help me improve.”

It others who wish to pretend that they are claiming fully 100% factual information to make their case, or perhaps who don’t bother actually learning about the topic you are posting about.

[David Nieporent]

Unfortunately, reading isn’t your specialty any more than law is. “I have limitations and won’t always get it right” is not the same thing as “explaining that they don’t provide factual information.” In fact, it’s the opposite. “Won’t always get it right” implies that it usually will get it right.

And the AI companies do not have to “claim fully 100% factual information” to be liable.

[RealityEngineer]

re: “And the AI companies do not have to “claim fully 100% factual information” to be liable.”

Except you gloss over several steps to make that assertion, useless handwaving. Again, there are myriad real world products that aren’t 100% safe to 3rd parties. The users of cars kill large numbers of bystanders in ways that the car companies could have prevented with either different designs that would make the cars cost $1 billion each, or require tech that doesn’t yet exist and would keep them off the market.

The fact that these may display content that is false isn’t apriori something that makes them liable. There are many steps involved in making that claim: and the mere mindless assertion of it doesn’t magically make you right. This series of posts is all about making that case: and you seem to assume you can just assert it.

[David Nieporent]

Again, there are myriad real world products that aren’t 100% safe to 3rd parties.

And the manufacturers of those products are routinely held liable to those 3rd parties when those 3rd parties are injured.

[RealityEngineer]

Are you seriously unaware that there are vast numbers of car accidents each year where bystanders are harmed and yet the manufacturers of automobiles are not held liable? Or the manufacturer of the alcohol someone drank isn’t held liable? An actual case needs to be made, not merely the presumption that somehow its guaranteed to be the case that since you dislike these systems that you seemingly have no understanding of the technology of, that somehow you magically are certain there is a case for design negligence that is too obvious you needn’t bother explaining it?

[Robert]

When an AI algorithm produces a worshipful poem about Joe Biden but refuses to produce one for Donald Trump for so-called ethical reasons, as ChapGPT did, it is not because the algorithm is acting objectively. It is because the creators of the algorithm trained it using supervised learning; in other words, data that was labeled with their subjective judgments. It is therefore conceivab le that an AI algorithm can be libelous with the source of the offense being the algorithm’s creators.

[Rev. Arthur L. Kirkland]

Speaking of former Pres. Trump, has anything interesting occurred with respect to him recently?

The Volokh Conspiracy seems to have found nothing interesting in that regard lately. Also, not much to be found here about what is being called the Defamation Case Of The Century, which produced some fascinating and important developments a few days ago, but the Volokh Conspirators seem not to have much to say about it.

Can anyone or anything explain this sudden-onset timidity among a bunch of customarily garrulous right-wing culture warriors?

Carry on, clingers.

[A nerdy Fred]

How does the analysis change if we add to the hypothetical that CheckBKG’s owner spent years trying to stop it from producing false negative information, and that at least one expert in the field has said publicly that the problem with CheckBKG’s technology is unfixable?

[Michael P]

In that case, a lawyer or judge would probably work in a reference to The Clash’s song “Train in Vain (Stand by Me)”, perhaps in reference to refusing to stand by their model’s output no matter how long they train it. If anything, persistence in the face of failure — and expert predictions of assured failure — should increase their liability.

[Eugene Volokh]

Well, that might be a defense to a negligent design claim, on the theory that there is no reasonable alternative design that’s safer but still sufficiently effective. I discuss this in some detail here. But I’m not sure that this would be so on the facts, especially as to the manufactured quotes, for reasons I discuss there.

[RealityEngineer]

In the case of the hypothetical company design negligence would seem the issue to address, but that doesn’t mean its applicable to the AI companies despite your superficial handwaving at “alternatives” that didn’t fully address the problems. I posted a link to an article detailing many of the ways these things can hallucinate that aren’t as easy to spot with simplistic approaches since the basic issue is they don’t have a model of the world and reason about it. They inherently aren’t built to be “truth machines” but instead “creative machines which may be useful” ala “all models are wrong but some are useful”. Just as cars are useful despite 3rd parties being victims of accidents, and yet they are allowed to exist without achieving some utopian complete safety to bystanders standard.

Any legal model succeeding in confusing people into holding AI vendors liable leading to their withdrawal is likely to be like the war on drugs that leads to people using less safe alternatives. Home brew LLMs, open source ones operating in other countries or trained by the Chinese (who will likely find some way to create or steal a useful model) that people get to remotely if the US ones are pulled. If they try to block access to them (as they are threatening to do with TikTok), Americans are familiar with VPNs and word will spread of them if that is the only way to use these tools, just as word of these tools spread quickly. There are browsers with free VPNs in them. Perhaps Chinese LLMs would intentionally throw in misinformation to undermine the US or to steer people to their products or explain things in ways that support them. If they are useful for what most people care about, they’ll still use them just as people complain in theory about social media and Twitter but still use them.

In terms of AI “design negligence”: these are companies with vast resources that have have each independently come up with systems with similar flaws, as have startups hoping to find an edge to compete with them. That should suggest considering some humility before assuming they ignored obvious alternative designs. Google saw what many argue was a $100 billion stock drop after a demo which had errors in it, yet they have released something also with flaws. The idea there are easy design alternatives falls into the category of “extraordinary claims require extraordinary proof” when someone not fulltime involved in this field thinks they can easily outdo them.

Your proposed design alternatives were flawed in ways suggesting a lack of basic understanding of these systems, which suggests learning more before proposing them. As Yan LeCunn, AI pioneer and head of AI at Meta notes, there are inherent problems in the LLM technology that lead them to hallucinate, and I posted a paper detailing all the myriad ways. Your band-aid and whack-a-mole/bailing wire aren’t a useful strategy to make a dent in the inherent flaws in these systems, even if they might catch some things.

It seems more useful to the world for it to grasp that they will make mistakes and hence hold human users responsible for treating the information they create as factual. That should give incentive to users to be more cautious: which as a side effect is useful since they should learn to better evaluate information they receive from other humans given all the misinformation out there. (from quack medicine to junk conspiracy theories).

The free market sees the need for ways to validate this information, and having these products out there leads to competition to address their problems.