Everyone’s talking about hallucinated case citations in filings. But hallucinations are different for eDiscovery solutions. Here’s why.

The latest high-profile court filing containing hallucinations has made big news this week (you can check it out here if you haven’t heard about it). Sounds like this premier firm has taken responsibility for it, which is good. We’ve seen cases where that wasn’t the case, like this one.

The extensive discussion about hallucinated case citations in filings (many through the use of public AI models like ChatGPT ad Claude) and the rapidly growing list of cases involved (compiled in Damien Charlotin’s site here) is causing many in the legal community to shy away from using generative AI for legal use cases, including eDiscovery. But hallucinations are different for eDiscovery solutions. Here’s why and what legal professionals need to do about it.

Where Hallucinations Are the Same

Before we talk about why they are different, let’s talk about where hallucinations are the same for both public AI models and eDiscovery solutions:

• At their core, both public AI models (like LLMs) and AI-enabled eDiscovery tools rely on probabilistic methods rather than true “understanding” – these models don’t actually understand anything.
• Both environments can produce outputs that appear authoritative while being incorrect.
• Hallucinations in both contexts are highly sensitive to the quality of the input data: bad prompts = bad outputs. That doesn’t mean good prompts guarantee good outputs, but it helps.
• Neither public AI models nor eDiscovery systems are capable of fully validating their own outputs. There must be human oversight (aka, “human-in-the-loop”).

That last one is critical and the crux of why we’re seeing so many hallucinated case citations in filings. We’ll get back to that in a bit.

Where Hallucinations Are Different

There are at least three ways in which hallucinations are different in eDiscovery solutions than they are for public AI models:

1. Source of Truth vs. Generative Construction: Public AI models generate new content and may produce outputs that have no basis in reality, effectively fabricating information. eDiscovery systems operate on a defined corpus of actual evidence. So, hallucinations in eDiscovery are typically not fabrications – they’re more likely to be misinterpretations, misclassifications, or overgeneralizations of real data. That’s an important distinction: public LLM hallucinations often involve invented facts, while eDiscovery hallucinations involve errors in how existing information is understood or categorized.
2. Ability to Verify Against Evidence: A defining strength of eDiscovery workflows is the ability to trace outputs back to underlying evidence. Most implementations of GenAI I’ve seen in eDiscovery involve links back to the source evidence for verification – while the source evidence may not always match GenAI’s interpretation of it, you can at least trace back to determine whether it does or not. Conversely, public LLM outputs may not relate cleanly to verifiable sources, and citations can sometimes be fabricated or unverifiable. This makes eDiscovery inherently more auditable and transparent, especially if the workflow is properly designed.
3. Measurement and Validation Frameworks: eDiscovery benefits from well-established, court-accepted validation methodologies. Metrics such as recall, precision, and elusion, combined with statistical sampling and benchmark control sets, provide a structured way to evaluate system performance (as I discussed in our new whitepaper on evaluating methods for review that eDiscovery Today released on Monday). Public AI models, by contrast, lack universally accepted frameworks for measuring truthfulness or accuracy. Their evaluation is often subjective or dependent on the specific task, making consistent validation more challenging.

It’s a People Problem, Not a Technology Problem

Notice I didn’t say “lawyer problem”, I said “people problem”. Earlier this week, I conducted an analysis of Charlotin’s database for a presentation. 790 of 1,333 cases in the database (at the time) solely involved pro se parties, not lawyers (they solely accounted for 495 cases). Pro se parties who use public LLMs to draft case filings are de-facto legal professionals – they’re just as responsible for checking the results as lawyers are.

Regardless of what type of party is involved and what type of technology they’re using, mistakes can occur. It can even happen with TAR processes. In the 2018 case In re Domestic Airline Travel Antitrust Litigation (discussed in The Sedona Conference^® TAR Case Law Primer) the plaintiffs determined that defendants’ TAR process resulted in recall of 97.4%, but precision of 16.7%, meaning defendants overproduced hundreds of thousands of documents that weren’t responsive (they computed their own metrics, but did so incorrectly).

Trust, but verify. But don’t blame GenAI technology hallucinations as a reason not to use the technology – especially for eDiscovery. The proverb “A bad workman always blames his tools” is applicable to technology as well. Understand how to use the tools and verify the results. It’s not that hard.

So, what do you think? Do you agree that hallucinations are different for eDiscovery solutions? Please share any comments you might have or if you’d like to know more about a particular topic.

Image created using DALL-E 3, using the term “robot lawyer wearing a suit experiencing hallucinations”.

Disclaimer: The views represented herein are exclusively the views of the author, and do not necessarily represent the views held by my employer, my partners or my clients. eDiscovery Today is made available solely for educational purposes to provide general information about general eDiscovery principles and not to provide specific legal advice applicable to any particular circumstance. eDiscovery Today should not be used as a substitute for competent legal advice from a lawyer you have retained and who has agreed to represent you.