Legal AI vendors talk a lot about hallucination. They claim to reduce it, prevent it, or catch it before it reaches you. Some claim to be "bias-tested" while others tout a low "error rate."
These terms get used interchangeably in sales conversations and marketing copy, but they describe different problems with different causes and different fixes.
If you can't tell them apart, you can't ask the right questions when evaluating a vendor, and you can't accurately assess where your actual risk sits.
This glossary breaks down the real causes of AI hallucinations, why they matter, and what you can do to protect your firm and your reputation.
Fabricated Cases
Definition: When an AI system generates a case citation that sounds plausible and confident but has no basis in real law.
This is typically what we first think of when someone mentions an AI hallucination.
Example: You ask an AI tool to summarize case law around an issue, and it gives you a citation to a case that doesn't exist, complete with a plausible-sounding case name, docket number, and holding.
Why it matters to you: This is one of the highest-stakes categories of AI risk in legal work right now. When a hallucinated citation ends up in a filing, it exposes you to real malpractice and sanctions risk.
How to protect yourself: Independently verify every citation before it goes into a filing: confirm the case exists and the reporter information matches. You can also rely on tools that build this step in automatically: with LOIS, every case referenced is run through a secondary verification layer and flagged as either verified (with a direct link to the real case) or unverified, so you always know which citations are safe to use and which need a manual follow-up before they go anywhere near a client or the court.
Mischaracterized Holdings
Definition: The case cited is real. But the AI describes its holding incorrectly, attributing a ruling or legal principle to it that the court didn't actually make.
Though fabricated cases get the most press, this form of hallucination can be even more dangerous because it’s more difficult to catch.
Example: An AI tool cites a real, well-known case correctly by name and reporter citation, but describes it as having ruled on an issue the court never addressed.
Why it matters to you: This can be harder to catch than a fabricated case, because the citation itself checks out. A quick lookup confirms the case is real, which can create false confidence. The error is in the substance, not the existence, of the citation.
How to protect yourself: You can’t stop at confirming a case exists. You have to read the actual holding, not just the case name. LOIS-verified citations link directly to the source case so you can compare the stated holding against the real opinion in one click.
Misattributed or False Quotes
Definition: The AI presents something as a direct quote from a judicial opinion, when it's actually paraphrased, fabricated outright, or, in a particularly deceptive version, pulled from a dissent rather than the court's actual ruling.
Example: An AI tool quotes language as coming "from the opinion" when that language actually appears in the dissent. The quote reverses what the case actually stands for.
Why it matters to you: This failure mode is especially risky because it can flip the apparent meaning of a case entirely. Citing a dissenting opinion as if it were the court's reasoning misrepresents the law. But even when the words accurately represent the general substance of a holding, a false quote is a glaring sign that a lawyer hasn’t done their due diligence.
How to protect yourself: Trace every quote back to its exact source and confirm which part of the opinion it came from. Verified citations in LOIS link to the specific line in the source document, making this distinction checkable at a glance.
Errors from Submissions or Exhibits
Definition: Unlike the case law failures above, this is a mistake made while processing your own real, uploaded materials, like a submission, exhibit, or transcript. The source is accurate but the AI misreads, miscounts, or misstates something from it.
Example: An AI tool correctly identifies a deposition transcript but summarizes a witness statement inaccurately.
Why it matters to you: An inaccurate summary of your own exhibit can lead you to build an argument on a fact that isn't actually there. The summary reads smoothly. The source is real. But misstated exhibit details in a filing can create an inconsistency between what's on the record and what's in the actual evidence. Opposing counsel can use this mistake against you, even if the error was unintentional.
How to protect yourself: Spot-check AI-generated summaries of your own submissions and exhibits against the original documents, particularly on dates, figures, and direct attributions. Grounded tools that link summaries back to the specific page and line of the source document make this comparison fast, rather than requiring a full re-read.
Bias
Definition: A systematic skew in outputs that reflects patterns in the AI's training data. This isn’t about a single mistake, but a consistent tendency across many outputs.
Example: An AI tool used for case triage consistently rates certain claim types as lower priority, independent of the actual facts of each case.
Why it matters to you: You can’t catch bias by checking just one or even a few outputs. It's a pattern that only surfaces across many cases over time, with fairness and discovery implications that a single spot-check won't reveal.
How to protect yourself: This requires ongoing monitoring rather than one-time verification. Periodically review AI outputs in aggregate for patterns. Ask any vendor how often they test for this, and how.
What Hallucinations Teach Us
AI hallucinations aren't just one problem with one fix. A fabricated case, a mischaracterized holding, a misattributed quote, and a bias pattern come from different causes, and require different methods to catch them. Treating "hallucination" as a single, solvable issue is part of what leaves firms exposed.
Legal technology researcher Damien Charlotin notes that AI hallucinations are teaching us something about the way law has been practiced for decades. Charlotin writes:
hallucinations are fascinating for what they tell us about the theory of the law (the chain of authorities we always relied on) and its practice (the time-worn habit of copying and pasting strings of citations without checking them). For years, we (me included) have cited without reading; but now the costs of that practice have become explicit. In other words, hallucinations expose the epistemic hygiene the legal profession has long lacked, and that is precisely why they deserve to be studied.
A new generation of legal AI tools is building practices of “epistemic hygiene” directly into the practice of law. LOIS runs every case reference through a secondary verification layer, flagging each one as either verified, with a direct link to the real case, or unverified, so you know what needs a second look before it reaches a filing. It offers a quick check on reasoning itself, not just citations, flagging where a conclusion may not follow from its supporting authority. It shows you where each argument is grounded, and points you directly to the specific holding in question instead of forcing you to search the full opinion.
None of this replaces a lawyer's judgment. It's built to make that judgment easier to exercise, consistently, across every citation in every filing.
Curious what this looks like in practice? Book a personalized demo today.
