"AI Hallucination Cases," from Courts All Over the World

From Damien Charlotin, 87 cases so far, mostly from the U.S. but also from Brazil, Canada, Israel, Italy, the Netherlands, South Africa, Spain, and the UK. I expect that there are many more out there that didn't make the list (especially since many state trial court decisions don't end up in computer-searchable databases, and I expect the same is true for other countries' courts).

Note that the pace has been increasing: There are more than 22 listed (all but five from U.S. courts) over the last 30 days alone.

Comments

[Roger S]

AI has moved into law. Get used to it. There is no going back.

[Stupid Government Tricks]

Is that supposed to mean something? Get used to .... what? Mistakes? Lies? Hallucinations?

[Roger S]

Every case has a winning side and a losing side. The losing side usually makes bad arguments. I don't see why AI bad arguments are any worse than any other bad arguments.

[David Nieporent]

Yes, you've demonstrated repeatedly that you don't see anything wrong with lying.

[Supremacy Claus]

Everything the lawyer says is a lie.

[KenveeB]

There's a difference between a bad argument and making up law to fit whatever you want to say.

[Riva]

The solution is obvious. Make some AI tools S.Ct. justices where they'll be allowed to make things up.

[Andrew Dhuey]

@grok, please confirm that all of the citations in this brief are legit.

-- I’ve reviewed the provided legal citations to confirm their legitimacy by checking their existence, accuracy, and publication in the respective reporters. I've confirmed that all of these cases exist and are accurately cited.

Let me put it this way, Mr. Dhuey. Grok is the most reliable artificial intelligence ever made. Grok has never made a mistake or distorted information.

[Darth Buckeye]

Just like the HAL 9000.

[John F. Carr]

Hallucinator 3: Rise of the Machines

[Darth Buckeye]

I asked Grok on X to give me case citations for cases with a precise set of facts and legal issues. It gave me a list of 5 cases with brief descriptions of each. The 1st cite brought up a case that appeared to be on point (though I did not read it closely). The 2nd case had a name and a citation to a reporter where there was a case with that caption that started 1 page off from Grok's citation. However that case was on a totally different topic.

The other 3 citations were totally made up. Non-existent. Wow. Use AI at your own risk.

[Stephen Lathrop]

Most texts consists of denotative meanings inflected by a context of creation. You can do a simple experiment for yourself, to discover what that means.

Here is the experiment. Choose either the first or second sentence above, then look up and concatenate whatever dictionary definitions you judge most useful to deliver the meaning intended by the sentence. Read the concatenated result. See if on that basis you can even guess the meaning you readily grasped on the basis of the contextual understanding you brought to the words in the first place.

A few experiments of that sort ought to clarify the importance of insight into context of creation for understanding texts.

Problems for AI arise in at least two ways:

1. The AI training materials do not encompass the correct context of creation. That will be particularly likely in instances of antique historical texts—a near certainty for many of them.

For instance, AI training materials are unlikely to contain the context-building texts of many unpublished archival materials from repositories where such texts are stored. They certainly will not contain the texts of contextual materials which have been lost to history, or which still lie undiscovered in attics, cellars, and closets of old buildings.

2. The AI training materials encompass many culturally determined contexts for the text in question, but the singular correct context gets overwhelmed by culturally more-common erroneous contexts. That too is a problem which becomes more likely the further back in history the search for insight delves.

Consider: each succeeding decade since the original creation of a text layers on a somewhat modified context for understanding it. That always consists of modifications impossible for the creators of the text to have anticipated. Thus, all but invariably, any antique historical text arrives in the present as a mysterious survival, deprived entirely of its original context of creation, which inherently deceptive succeeding contexts have replaced piecemeal.

Whether those problems are susceptible to correction, or whether AI methods will have to be reworked from scratch to fix such problems, thus becomes a salient question, urgently in need of an answer.

Historical scholarship has developed laborious methods to somewhat circumvent those problems. Evidently, those methods have not yet been incorporated into AI algorithms. It seems unlikely that they can be. The scholarly methods seem to depend too heavily on a summarizing power in human intelligence which AI methods have yet even to approximate.