Seemingly Nonexistent Citation in Anthropic Expert's Declaration [UPDATE: Apparently Caused by Lawyer's Misuse of Claude to Format Citations]

UPDATE 5/15/2025 (post moved up): Anthropic's lawyers filed a declaration stating that the error was not the expert's, but stemmed from the (unwise) use of Claude AI to format citations.

The Declaration filed by a "Data Scientist at Anthropic" in Concord Music Group, Inc. v. Anthropic PBC includes this citation:

But the cited article doesn't seem to exist at that citation or at that URL, and Google found no other references to any article by that title.

Bloomberg (Annelise Levy) has a story about this, under the title "Anthropic Expert Accused of Citing Fake Article in AI Lawsuit" (Chat GPT Is Eating the World links to that). Magistrate Judge Susan van Keulen ordered the parties, apparently (according to Bloomberg) referring to this problem, to explain matters:

[N]o later than May 15, 2025, Defendant shall file a Statement addressing the issue raised by Plaintiffs' counsel at the outset of the hearing ….

I'll report what the Statement asserts once it is filed. Thanks to Prof. Edward Lee for the pointer.

UPDATE 5/15/25, 4:17 pm: Here's the explanation, from one of Anthropic's lawyers (emphasis added):

Our investigation of the matter confirms that this was an honest citation mistake and not a fabrication of authority. The first citation in footnote 3 of Dkts. 340-3 (sealed) and 341-2 (public) includes an erroneous author and title, while providing a correct link to, and correctly identifying the publication, volume, page numbers, and year of publication of, the article referenced by Ms. Chen as part of the basis for her statement in paragraph 9. We apologize for the inaccuracy and any confusion this error caused.

The American Statistician article reviewed and relied upon by Ms. Chen [the Anthropic expert], and accessible at the first link provided in footnote 3 of Dkts. 340-3 and 341-2, is titled Binomial Confidence Intervals for Rare Events: Importance of Defining Margin of Error Relative to Magnitude of Proportion, by Owen McGrath and Kevin Burke. A Latham & Watkins associate located that article as potential additional support for Ms. Chen's testimony using a Google search. The article exists and supports Ms. Chen's testimony in her declaration and at the May 13, 2025 hearing, which she proffered based on her pre-existing knowledge regarding the appropriate relative margin of error for rare events. A copy of the complete article is attached as Exhibit A.

Specifically, "in the context of small or rare-event success probabilities," the authors "suggest restricting the range of values to εR ∈ [0.1, 0.5]"—meaning, a relative margin of error between 10% to 50%—"as higher values lead to imprecision and poor interval coverage, whereas lower values lead to sample sizes that are likely to be impractically large for many studies." See Exhibit A, at 446. This recommendation is entirely consistent with Ms. Chen's testimony, which proposes using a 25% relative margin of error based on her expertise.

After the Latham & Watkins team identified the source as potential additional support for Ms. Chen's testimony, I asked Claude.ai to provide a properly formatted legal citation for that source using the link to the correct article. Unfortunately, although providing the correct publication title, publication year, and link to the provided source, the returned citation included an inaccurate title and incorrect authors. Our manual citation check did not catch that error. Our citation check also missed additional wording errors introduced in the citations during the formatting process using Claude.ai. These wording errors are: (1) that the correct title of the source in footnote 2 of Ms. Chen's declaration is Computing Necessary Sample Size, not, as listed in footnote 2, Sample Size Estimation, and (2) the author/preparer of the third source cited in footnote 3 is "Windward Environmental LLC", not "Lower Windward Environmental LLC." Again, we apologize for these citation errors.

Ms. Chen, as well as counsel, reviewed the complete text of Ms. Chen's testimony and also reviewed each of the cited references prior to submitting Ms. Chen's declaration to the Court. In reviewing her declaration both prior to submission and in preparation for the hearing on May 13, 2025, Ms. Chen reviewed the actual article available at the first link in footnote 3 of her declaration and attached hereto as Exhibit A, and the article supports the proposition expressed in her declaration with respect to the appropriate margin of error.

During the production and cite-checking process for Ms. Chen's declaration, the Latham & Watkins team reviewing and editing the declaration checked that the substance of the cited document supported the proposition in the declaration, and also corrected the volume and page numbers in the citation, but did not notice the incorrect title and authors, despite clicking on the link provided in the footnote and reviewing the article. The Latham & Watkins team also did not notice the additional wording errors in footnotes 2 and 3 of Ms. Chen's declaration, as described above in paragraph 6.

This was an embarrassing and unintentional mistake. The article in question genuinely exists, was reviewed by Ms. Chen and supports her opinion on the proper margin of error to use for sampling. The insinuation that Ms. Chen's opinion was influenced by false or fabricated information is thus incorrect. As is the insinuation that Ms. Chen lacks support for her opinion. Moreover, the link provided both to this Court and to Plaintiffs was accurate and, when pasted into a browser, calls up the correct article upon which Ms. Chen had relied. Had Plaintiffs' counsel raised the citation issue when they first discovered it, we could and would have confirmed that the article cited was the one upon which Ms. Chen relied and corrected the citation mistake.

We have implemented procedures, including multiple levels of additional review, to work to ensure that this does not occur again and have preserved, at the Court's direction, all information related to Ms. Chen's declaration. I understand that Anthropic has also preserved all information related to Ms. Chen's declaration as well….

Comments

[ah....Clem]

Not too long ago, the dicta was "Don't cite Wikipedia as a source".

Now it seems to be "Don't use AI to generate impressive-sounding but non-existent sources. Or at least be careful to check all of them before submitting to make sure they are real."

[MikeFr]

You would think an Anthropic expert would be aware of AI hallucinations.

I use AI to check contracts, but I always find the cited clause in the original contract. AI in this context is particularly prone to hallucinations. When a clause is similar to a standard clause used in many contracts but different in important ways AI will often hallucinate the standard clause.

[Dr. Ed 2]

Are human lawyers any better?

Humans will see what they think they see and (often) not what the actual text of the document is.

[mike petrik]

Yes, they are, at least for now. No doubt.

[Stephen Lathrop]

I have tried a few times to use AI to evaluate hypothetical historical citations. I judge the failure rate to approach 100% so far.

Problem is, a best-available historical method is often to make historical survivals critique each other. And thus to infer a context useful to understand something which happened in a long-ago past where the original context of creation for the survivals in question has not survived.

That will usually be the case if the history involved is not a recent one. Or if it involves a less-studied record—which is to say a large portion of all the studies likely to deliver new or revised historical insight.

Statistical methods which drive generative AI seem especially unsuited for that kind of analysis. Compared to the entire field of available inferences, the correct answers sought seem almost always to be statistically less likely than a historical counterfactual which some permutation of the training data can support. Distinguishing a less-likely outcome which did happen, from a more-likely outcome which obviously did not happen, seems for the present too challenging for generative AI.

I expect more experience will disclose that to be a commonplace failure mode for generative AI. It will likely plague many different kinds of inquiry, whenever a specific less-likely-but-correct answer is awash in a sea of otherwise commonplace-but-irrelevant statistical possibilities presented by training data.

Whether that is a problem hard-wired into current AI methods, and thus unsolvable without starting anew, seems a question urgently in need of an answer.

[Dr. Ed 2]

It sounds like AI can't distinguish between quality of sources -- that's something that a lot of undergraduates have problems with.

Just Sayin....

[mike petrik]

Undergraduates are not practicing law, or at least shouldn't be. And in my day undergraduates could make such distinctions. I guess humans are getting stupider.

[JonFrum]

It's long been the practice of graduate students - and professors - to cite papers without ever reading them. If you know that there's an expert in the sub-field you're writing about, you make sure to cite his/her work on the subject 'to show you know the literature. It's a game that's played, and as long as you don't get the gist of the cited article wrong, no one cares - although you should read everything yourself. But that's not law, for God's sake. If you get the evolution of flower petal length wrong, no one suffers. You'd think lawyers would be consciencious enough to check work that's been done for them.

[The Turtle Dove]

Amusingly, both cited authors are philosophers who specialize in epistemology. #irony

Mr. D.

[SMP0328]

Our manual citation check did not catch that error.

Someone got yelled at.

[Bored Lawyer]

AI Hallucinations -- sounds like drug

[Longtobefree]

Band name?

[Dr. Ed 2]

The one thing that AI could do -- and would be lovely at doing -- is recording URLs and the document and (a) keeping track of the URLs, (b) be able to link them to quotes from them, and (c) ping them to determine if the links are still valid AND the resource hasn't changed. If nothing else, a bit count match could do that because it is unlikely that a revision would result in exactly the same number of bits -- although that is a gamble.

[Soronel Haetir]

The problem with "has not changed" is that there are plenty of changes that are meaningless to us humans, but that would make it appear a different document to the computer. Things like changing the page styling, navigation boilerplate, even reorganizing the table of contents.

At least with Wikipedia, if the page has changed, you can see what it said at some earlier time (as long as you know when that time was).

[Andrew Dhuey]

Eugene, I do not know much about how The Google works in the context of lawyers who get into trouble via AI citations in courts of law, but I commend your continuing efforts to highlight what can and does go wrong when lawyers turn to AI.

I hope your posts on this blog have a deterrent effect for lazy, short-cutting lawyers who are tempted to skip past what I like to call the law part of law via AI.

[wvattorney13]

Note to lawyers: it doesn't seem that this attorney will get into any trouble with his law license and that is because at the first opportunity he admitted that he screwed up and relied on AI when he should have manually checked his citations. That is a mess up, not an ethics violation.

In so many of these similar stories, the attorney does not disclose and vehemently denies wrongdoing despite overwhelming evidence. That is where you get in trouble. Always be candid with the court.