A few weeks ago we wrote about how the lawyers for OpenAI and Microsoft in the massive multi-district litigation (MDL) against the companies over how they trained their AI had called for the exclusion of an expert witness report after it was revealed that lawyers for the plaintiffs had apparently surreptitiously paid for the report, and then tried to sneak it into evidence past the deadline for expert reports via a different expert who admitted he wasn’t familiar with the details of how the report was put together.
There are so many filings in this case (currently over 2,100) that it’s difficult to keep up with them all, but around the same time as they called out that expert’s report, they also tried to disqualify a number of other experts put forth by the plaintiffs. Trying to disqualify experts isn’t that big a deal, it happens all the time. But buried in this particular motion was an incredible claim regarding Professor Oren Etzioni, a long-term AI optimist turned skeptic, who has been warning that AI will supercharge the spread of disinformation online.
But in 2019 Dr. Etzioni, a University of Washington professor and founding chief executive of the Allen Institute for A.I., became one of the first researchers to warn that a new breed of A.I. would accelerate the spread of disinformation online. And by the middle of last year, he said, he was distressed that A.I.-generated deepfakes would swing a major election. He founded a nonprofit, TrueMedia.org in January, hoping to fight that threat.
Indeed, in a recent article he warned against people relying on AI-generated output without carefully and thoroughly checking the results:
A word of caution that I can’t repeat often enough: always verify anything important yourself. AI models hallucinate. They fabricate quotes, invent statistics, and present fiction with the serene confidence of a tenured professor. The verification step is essential. If you skip it, you are not reading with AI. You are gambling with AI.
Well… apparently Etzioni himself gambled and lost.
From OpenAI and Microsoft’s motion to exclude the plaintiffs’ experts, filed last month:
News Plaintiffs’ expert Etzioni’s opinions as to the supposed importance of news content in LLM training should be excluded because they are based on now-withdrawn AI-generated charts… To support his opinions, Etzioni offered a series of figures that included hallucinated figures based on the cited academic papers….
[….]
None of the numbers or measurements above were in the cited papers. Etzioni had used Anthropic’s Claude to generate the figures, which generated fake numbers. Etzioni admitted he failed to verify whether the figures were supported…. Etzioni conceded the charts did not meet his “standards for academic rigor,” … and following his deposition withdrew them. He also removed another AI-generated chart from his Rebuttal Report. Etzioni provided no other evidence to compensate, and his opinions concerning the value of news should be excluded.
Ooooops.
To have one expert report entered under questionable circumstances may be an error. But to have two smacks of desperation, and this is true even as these experts are coming from two different sets of plaintiffs in this case (and even after the publishing plaintiffs decided to drop that original report we spoke about a few weeks ago). This is a big case that could shape how AI gets built for years. When the plaintiffs keep leaning on experts this shaky, it’s hard not to read it as a sign that the underlying case is weaker than they’d like, and that they’re hoping to pressure the AI companies into folding anyway. It’s tough enough to fight a legitimate case. When the other side is so desperate to find experts to prop up their flimsy case that they end up in situations like this, it becomes even more difficult.
Yes, there have been plenty of examples making the rounds of expert witnesses submitting hallucinated reports in multiple court cases. But this is a case literally challenging the ability of AI companies to read the internet, and to have one of the experts arguing against the AI companies get caught with hallucinated findings seems extra embarrassing.
Indeed, you could argue there’s a deep irony with Etzioni’s reliance on fabricated data, beyond just his own warnings about it, given that if the side who retained him wins this case, it will make it that much more difficult to fight back against hallucinations and misinformation. AI tools that can read and retrieve content from the wider web, including news sites, can check and cite their sources instead of making things up. But the plaintiffs who brought Etzioni into this case are effectively trying to block that, which would likely leave us with less reliable tools that are even more of a “gamble” to rely on.