A Verified Answer to the Wrong Question Is Still Wrong

There is a specific kind of confidence that comes from checking your work. You verified the citation, cross-referenced the clause, confirmed the case actually exists and says what the draft claims it says. Every box is ticked. And the answer can still be completely wrong, because verification checks whether an answer is accurate. It does not check whether it is an answer to the right question.

Most of the public conversation about AI in legal work is stuck on the first problem: hallucinated citations, fabricated case law, drafts that read fluently and cite confidently while referencing cases that were never decided. That problem is real and worth taking seriously. But it is not where the risk actually starts. It starts one step earlier, at the moment someone opens a tool and types a prompt before they have decided, precisely, what the legal question in front of them actually is.

Framing comes before drafting, always

Senior counsel does not start with the document. They start by scoping the matter: who the parties actually are and what they actually want, what the real risk is versus the risk that was assumed at the outset, which facts are settled and which are still contested, and what a good outcome would even look like given all of that. Only once that scoping is done does drafting begin, and by the time it does, the draft is answering a question that has already been pressure-tested.

Prompting an AI tool skips straight to drafting. The tool has no way to scope a matter, because scoping requires judgment about what matters and what does not, informed by experience with how this particular kind of dispute or deal tends to actually unfold. What the tool receives is whatever question the person typing happened to frame in the moment, usually the first version of the question that came to mind, not the version that survives scrutiny.

This is not a hypothetical failure mode. It is the ordinary way that fluent, well-cited, technically accurate output ends up useless or worse, because everyone downstream, including the person who prompted it, trusts a confident answer more than they interrogate the question it was answering.

Why fluency makes this harder to catch, not easier

A badly written draft invites scrutiny. Something about it signals unfinished, and a reader slows down. A fluently written draft does the opposite. It reads exactly like something a competent associate would produce, structured correctly, cited correctly, using the right register and the right conventions of the practice area. That fluency is precisely why the underlying framing error survives all the way to a partner's desk without anyone stopping to ask whether the question itself was the right one.

Tools like Harvey and Luminance are genuinely good at this kind of output, and that capability is exactly the problem. When output sounds right, the instinct to sense-check it quietly weakens. Nobody decided to stop checking. The checking just stopped happening by default, because it no longer felt necessary, and unnecessary is a feeling, not a fact.

What framing the question actually looks like

• Name the actual parties and their actual interests, not the abstracted version of the dispute or deal
• State the real risk in one sentence before drafting anything, and check it against what you assumed at the outset
• Separate settled facts from contested ones, a draft that treats a contested fact as settled is answering a different case than the one you have
• Ask what a good outcome looks like from the client's actual position, not the generically correct legal answer
• Only then hand the framed question to a drafting tool, and treat the draft as a first pass against that framing, not a finished answer

The honest limitation

None of this means AI tools are unsafe to use in legal work. Used after the framing step, with a verification checkpoint built in rather than assumed, they compress review time meaningfully, in some reported cases by as much as ninety percent. The limitation is that framing cannot be delegated the way drafting can. It requires exactly the kind of judgment, about which facts matter, which risks are real, and what the client actually needs, that experience builds and that a language model, however fluent, has no access to.

A verified answer to the wrong question is not a partial success. It is a confident, well-cited, fully checked version of a mistake, and it is more dangerous than an obviously bad draft precisely because nobody feels the need to question it. The discipline that protects against this was never about catching AI's errors after the fact. It is about doing the framing work first, the way senior counsel always has, before any tool gets involved at all.

A short worked example

A client asks for a termination clause reviewed for enforceability. Prompted directly, an AI tool will produce a competent, well-cited analysis of whether the clause as written is enforceable in the abstract, and it will be accurate. The trouble is that the real question was never the abstract one. The real question, once you scope the matter properly, might be whether this specific clause is enforceable against this specific employee, in this specific jurisdiction, given a pattern of informal side agreements that were never put in writing but that a court would likely consider relevant.

Framed and answered correctly, that second question might produce a completely different recommendation than the first, even though both answers would cite the same statute correctly and neither would contain a single fabricated case. The document that answers the wrong question is not sloppy. It is polished, verified, and aimed at a target that was never the client's actual problem, and the only thing standing between those two outcomes was five minutes of framing before anyone opened a drafting tool.

A short answer to what the wrong question can cost is not always dramatic. Sometimes it is a clause that technically holds up but leaves a real exposure unaddressed because nobody asked whether the abstract legal question matched the client's actual situation. Sometimes it is a filing that is internally consistent and still fails because it argued the strongest version of the wrong claim. In both cases, everyone involved could point to a verified, well-cited document right up until the moment it mattered.

That is the whole argument, stated plainly one more time: verification is a quality check on the answer you already have, not a check on whether it is the right question to have answered. Firms that treat the two as the same thing will keep producing accurate documents that solve the wrong problem, and no amount of citation checking fixes that, because citation checking was never designed to catch it.

Read next: What Senior Counsel Does Before Opening Any AI Tool