What’s in a “Fact”?

Catherine Thompson · Head of Mary Legal, U.S.

Understanding whether an AI system’s output is trustworthy is not a fundamentally different exercise from understanding whether any other fact, argument or judgment call is trustworthy. And lawyers — trained, more rigorously than almost any other profession, to sit inside uncertainty rather than flee it — are unusually well equipped to do exactly that. Possibly better equipped than most.

I. Skeptics, and the Difference Socrates Made

The ancient Skeptics believed something genuinely radical: Certainty is not merely difficult, but structurally out of reach. The often wise response, therefore, was to stop reaching for it. The Pyrrhonists, a sub-group of the Skeptics, called this cessation of reaching epoché or “suspension of judgment.” They reasoned that for every claim there exists an equally persuasive counter-claim, so peace of mind lies in refusing to decide between them.¹

Socrates, a philosopher whose teachings inspired the Skeptical traditions, as well as American legal pedagogy generally, was famous for his admission “I was conscious that I knew nothing at all.”² While even the meaning of this admission is debated among scholars, what seems agreeable is that, for Socrates, this admission was not a resting place. It was the opening move of a method, elenchus, built on the conviction that truth exists and can be theoretically found through relentless cross-examination, even if Socrates himself hadn’t found it yet.³ In short, Socrates was not suspending judgment. He was working.

A great many lawyers approach artificial intelligence today sounding like Pyrrhonists — permanently withholding judgment because the tool feels unknowable — when what the profession might actually need, and may be unusually well trained to provide anyone using AI, is the Socratic method.

II. The Fear Underneath the Fear

Ask a room of practicing attorneys why they’re wary of AI, and two answers arise, with slightly different framings.

The first: lawyers are professional skeptics. We sell our credibility one “fact” at a time, and we cannot afford to build an argument, a filing, or a client recommendation on a fact that turns out to be wrong. A single bad citation invites sanctions; a single unverified fact in a declaration can cost a case, or a client’s trust, permanently.

The second: AI’s outputs are getting uncomfortably good at simulating the thing lawyers are actually trained to produce — judgment. Ask a large language model to analyze a contract, brief a case, or weigh a settlement position, and it will hand back something that reads like reasoned judgment, often persuasively, often instantly. That unsettles us not because the output is obviously wrong, but because we cannot easily tell when it is.

Strip the framing away, and both fears collapse into a single sentence: We don’t fully understand how AI decides what’s “true”, and we’re worried we lack the tools to find out.

Here is my answer to both fears, stated (as a form of Socratic pedagogy in itself) before I defend it: Understanding whether an AI system’s output is trustworthy is not a fundamentally different exercise from understanding whether any other fact, argument or judgment call is trustworthy. And lawyers — trained, more rigorously than almost any other profession, to sit inside uncertainty rather than flee it — are unusually well equipped to do exactly that. Possibly better equipped than most.

III. The Facts

The AI safety literature converges on two categories of risk that map, almost exactly, onto lawyers’ two biggest fears.

The first fear is called epistemic uncertainty: incompleteness or inaccuracy in whether the “facts” an AI system asserts actually represent the real-world decision space it’s being asked to reason about.⁴ A model grounded on an incomplete, stale or unrepresentative slice of reality produces outputs that sound confident regardless of how well they actually track the world they describe.

The second fear, the risk of poor decision-making, stems from the first: the possibility that a model, reasoning from epistemic uncertainty, produces recommendations that are biased, skewed or otherwise unsound — not because the model is malfunctioning, but because faulty factual grounding compounds into faulty judgment.⁵ That compounding is not abstract. Two recent studies of “hallucinations” in AI legal research tools illustrate (1) how systems marketed as “hallucination-free” continue to hallucinate at significant rates,⁶ and (2) that hallucinations can stem from systems tending to accept a false premise embedded in a user’s query rather than flag it — what the researchers term contra-factual bias.⁷ That is epistemic uncertainty and poor decision-making risk compounding in a single, observable failure: The model doesn’t know what it doesn’t know, and it decides anyway.

Layered on top of both risks is a third, more troubling pattern, cognitive surrender: Humans appear eager to hand off exactly the hardest parts of this work — clarifying what is actually uncertain, and exercising the judgment required to decide well despite that uncertainty — to tools least equipped to disclose when they’re guessing.⁸

IV. The Saber-Toothed Tiger

These fears are far from hypothetical, particularly for lawyers. A single bad fact is a credibility event; multiplied across a practice, or automated across a firm, it becomes a pattern that ends careers. And because so much of what lawyers do sits upstream of consequential decisions — custody arrangements, sentencing recommendations, safety orders, medical directives, business judgments other people will rely on — a bad decision built on a bad fact is rarely just embarrassing. It can be irreversible. Sometimes it is life and death.

AI does not remove that risk. It scales it. Speed, fluency and apparent confidence are exactly the qualities that let a faulty fact travel further and faster before anyone catches it. Feeling something like dread at that prospect is not an overreaction. Calling these anxieties a “saber-toothed tiger” for lawyer’s nervous systems is not hyperbole; it is a fair description of what a bad “fact,” moving at machine speed, can do to a life.

V. The Twist

Here is the twist: Lawyers are arguably better positioned than almost anyone else to face this particular tiger down — not despite our training as skeptics, but because of it.

Every first-year law student is drilled, through cold calls and case briefs, in a repeated exercise: Issue, Rule, Application, Conclusion, or “IRAC”, a scaffolding for constructing arguments. The Socratic method — Socrates’ method, not the Skeptics’ — supplied the process: Interrogate every link in the chain until you can say, precisely, what you know, what you’re assuming, and what remains genuinely uncertain.

That training teaches something specific and useful lawyers can and must bring to the world of AI. An argument is not a “fact”. It is assertions of “facts” linked together through an exercise of judgment, and that linking can be examined, challenged and defended on its own terms, independent of whether any single fact in the chain is verifiably true. Go one level further, and the same is true of each asserted “fact” itself: Our justice system drills into us that a “fact” is not a discrete, static thing. It is the process by which someone arrived at it through adjudication — carefully weighing evidence, assumptions, doubts and alternatives.

A good lawyer, like a trustworthy AI system, is not a regurgitator of binary rules. Binary rules don’t require judgment; nor is the world in which we practice actually binary in most senses. What a good lawyer offers — and what a “trustworthy” AI tool must offer, if it is going to earn that description — is the capacity to navigate ambiguity: to hold multiple plausible readings of the facts, weigh them honestly, and disclose the confidence behind the conclusion rather than paper over the gaps.

That reframes the question lawyers should be asking about any AI tool. The wrong question is: is this system’s thinking process inscrutable? Most of them are, to some degree, and fixating there leads nowhere but paralysis — a kind of permanent epoché that helps no one. The right questions are narrower, and answerable: Can I scrutinize the AI system’s process? And, just as important: Am I actually doing so?

Pressure-testing a genuinely trustworthy AI tool, in other words, might look a lot like cold-calling a 1L, with the lawyer now in the professor’s chair:

ISSUE Can I tell what question the system believes it’s answering — and scrutinize whether that’s the right one?

RULE Can I identify what standard, source or authority is doing the work behind its answer — and test my confidence in it?

APPLICATION Can I see how the system moved from that standard to its conclusion — which facts it weighed (and how), which it didn’t, and why?

CONCLUSION Can I tell whether I’m looking at a single confident answer or a properly calibrated range of outcomes — and would the system tell me the difference if I asked?

If the answer to any of these is no, the tool hasn’t necessarily failed you. But you don’t yet know whether it will.

It is entirely normal — not a professional failing, and not unique to lawyers — to feel fear toward something we don’t fully understand. Every skeptic, ancient or otherwise, is reacting to exactly that discomfort. But if there is any group on Earth well suited to unpacking what is, what is not, what should be, and what should not be, it is lawyers. That is, after all, the whole of the job.

Socrates left no writings of his own; everything we know of his method survives only through Plato, itself a small demonstration of how carefully a fact has to travel before it can be trusted. But the line that has stayed with lawyers for 24 centuries states the underlying stakes as well as anything in the safety-science literature: The unexamined life is not worth living.⁹ In 2026, the corollary is just as literal: The unexamined “fact” is not one worth acting on.

Catherine Thompson is Head of Mary Legal, U.S., at Mary Technology, a legal AI company. Before joining Mary, Catherine practiced in international capital markets and asset management, and later high-net-worth family law, in D.C., Maryland, Virginia and California.

About Mary Technology

Mary is built for this form of review. It structures the matter record the client supplies, links material findings to specific source pages, flags conflicts and missing material it can identify, and keeps unresolved support questions visible for the lawyer. Mary doesn’t decide what the evidence means or what the lawyer should sign.

References

1. Katja Vogt, “Ancient Skepticism”, The Stanford Encyclopedia of Philosophy (Fall 2026 Edition) (describing the Pyrrhonist practice of epoché, or suspension of judgment).

https://plato.stanford.edu/archives/fall2026/entries/skepticism-ancient/

2. Id. (“Socrates’ commitment to reason — examination as the way to find out — inspires the skepticism of the Hellenistic Academy.”); Plato, Apology 22d (H.N. Fowler trans., Loeb Classical Library 1966).

http://www.perseus.tufts.edu/hopper/text?doc=Plat.+Apol.+22d&redirect=true

3. Gregory Vlastos, The Socratic Elenchus, 79 J. Phil. 711, 714 (1982); see also Christopher S. King, Wisdom, Moderation, and Elenchus in Plato’s Apology, 39 Metaphilosophy 345 (2008).

http://www.jstor.org/stable/24439908

4. Bonnie Johnson, Metacognition for Artificial Intelligence System Safety — An Approach to Safe and Desired Behavior, 151 Safety Sci. 105743 (2022) (citing Kush R. Varshney, Engineering Safety in Machine Learning, 2016 Info. Theory & Applications Workshop (ITA) 1 (defining epistemic uncertainty as incompleteness or inaccuracy in a system’s knowledge of its operational environment)).

Johnson: https://doi.org/10.1016/j.ssci.2022.105743

Varshney: https://doi.org/10.1109/ITA.2016.7888195

5. Johnson, supra note 4, at Section 5.3 (identifying biased, skewed or otherwise unsound decision recommendations as a primary AI system failure mode).

https://doi.org/10.1016/j.ssci.2022.105743

6. Varun Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, 22 J. Empirical Legal Stud. 216 (2025).

https://doi.org/10.1111/jels.12413

7. Matthew Dahl et al., Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, 16 J. Legal Analysis 64 (2024) (coining the term “contra-factual bias” at Section 5.2); Varun Magesh et al., supra note 6 (finding the same pattern, which the authors term “contrafactual bias,” in retrieval-augmented legal research tools, citing Dahl et al. 2024).

Dahl et al.: https://doi.org/10.1093/jla/laae003

Magesh et al.: https://doi.org/10.1111/jels.12413

8. Steven D. Shaw & Gideon Nave, Thinking—Fast, Slow, and Artificial: How AI Is Reshaping Human Reasoning and the Rise of Cognitive Surrender (Wharton Sch. Rsch. Paper, Jan. 11, 2026).

https://ssrn.com/abstract=6097646

9. Plato, Apology 38a (H.N. Fowler trans., Loeb Classical Library 1966).

http://www.perseus.tufts.edu/hopper/text?doc=Plat.+Apol.+38a&redirect=true