The 8 Most Common Types of AI Hallucinations in Case Law
In recent presentations on AI and the law, one of the most frequent questions I’ve received is whether AI hallucinations, particularly in case law, are simply early-stage teething problems or a permanent risk we'll need to manage.
For me, the answer depends largely on the specific type of hallucination we're dealing with.
In my latest article, I explore the eight distinct types of hallucinations I've identified through reviewing numerous AI-related cases. These range from obvious mistakes like fabricated case names, to subtler, more concerning issues such as accurately cited cases with incorrectly stated facts, misrepresented ratios, or even AI citing AI.
While some hallucinations are relatively straightforward to catch, others carry the risk of embedding incorrect legal precedents into our legal regime, especially as AI-generated content starts filtering into judgments, headnotes, and academic commentary.
It seems to me that while certain hallucinations might become easier to detect over time, others will require significant human expertise and supervision.
But who will pay for that and how do we ensure consistency when even human interpretations of judgments frequently differ?
Read the full article here.
There was a case I argued where the law report emphasised the wrong thing and glossed I what I considered to be the most important parts of the ratio. That report was picked up and amplified on ‘aggregator’ sites and now provide the definitive summary of the case. That was with entirely human authors of the law reports, working off the judgment alone (which left a lot out) rather than the case argued by the barristers. So of the issues you identify above, I think AI *interpretation* is likely to be the trickiest.
I think the thing that will save or scupper us is whether or not these new technologies are deployed by an existing brand or editorial team with a reputation they wish to protect.
ChatGPT, Google and the other popular AIs come with no ‘warranty’ and the people that make them suffer no diminished reputation when the AI hallucinates.
However, I imagine anything put out by Westlaw, LexisNexis, ICLR will come with a stamp of approval from those companies. I don’t care whether they hand write each report with a quill or whether they deploy a bunch of AIs to generate 100% of the content: their need for quality control to protect their reputation will mean that they have to check their output before making it available to lawyers.