Critique a design that presents an AI answer with the same authority as a database lookup.

Brant & Vollmer is a mid-size litigation firm. Petra Falkenrath is a senior litigation associate there. Verdictly is an AI legal research tool the firm licenses alongside CourtWire, its paid, verified case-law database. Both tools show their answers inside the exact same card.

How to answer this, stage by stage

The interviewer isn't asking whether you can spot a UI that's a little busy. They're asking whether you can name the exact visual choice that's doing the lying.

The card that borrowed someone else's authority

For six years, Petra verified every legal research answer by hand against CourtWire's records before it went anywhere near a client memo, about 25 minutes a question, across roughly 30 questions a week. That's close to 12 and a half hours a week just checking.

Two very different kinds of paragraph. One card style, worn by both.

Verdictly drafted a first-pass answer to the same kind of question in under 15 seconds. Because its answer card used CourtWire's exact layout, font, and citation style, associates began treating a Verdictly answer as pre-verified the same way a CourtWire pull already was. Verification time on a Verdictly answer fell to about 3 minutes.

Here's the turn: the twenty-two minutes Verdictly saved per question was never the real story. The real problem is that an unverified paragraph, once it wears a verified record's exact clothes, walks straight into a client memo without anyone stopping to ask whether it should.

Four missing facts. Any one of them would have been enough to slow Petra down.

At its worst, this design flaw doesn't cost a firm an afternoon of embarrassment. A fabricated case citation filed in a real brief can draw court sanctions, a malpractice complaint, and a client's whole file being reviewed from scratch for anything else the tool touched.

What I would leave alone: CourtWire's actual verified case-pull cards don't need any redesign. They're correctly earning the authority they display, because a person confirmed the record was real before it ever reached the screen.

The lesson: a screen can lie about how trustworthy it is without a single wrong word on it, just by choosing to look like something else that already earned its trust honestly.

Now here is the same thing as a story

The short version above is what you'd say defending this fix to the firm's managing partner. Read this one for how the near miss actually happened.

The Brant & Vollmer research floor goes quiet around 6 p.m., after the partners leave and the associates start second-drafting memos due the next morning. Petra had worked litigation research there for six years and could smell a weak citation before she finished reading the sentence around it.

Verdictly arrived in the spring, and the first months were good. Petra would pull a first-pass answer, then walk every citation back to CourtWire by hand before it went into anything real. The tool was fast. Her own habits didn't change at first.

By midsummer, she was spot-checking maybe half of Verdictly's answers instead of all of them, since the ones she'd checked kept coming back clean. By early fall, she was checking only the ones that felt unusual, because the card looked exactly like a CourtWire pull whether it was right or not, and nothing on the screen ever told her which was which.

The AI answer sits in the one corner nothing should ever sit in: high stakes, dressed exactly like a fact.

The trigger was small. A junior associate asked Petra, in passing, whether a case Verdictly had cited that morning was "one of the real ones or one of the AI ones," and Petra realized she had no way to answer without opening a second window and searching for it herself.

She pulled forty of the firm's real research questions from the past quarter and ran each one back through Verdictly, checking every citation by hand this time. Six answers, fifteen percent, cited a case that was either misquoted or did not exist. One of those six had already been forwarded into a client memo two weeks earlier, caught only because a partner happened to recognize the case name as unfamiliar.

The fifth box is the one that actually caught the problem. The other four explain why nobody caught it sooner.

Brant & Vollmer didn't drop Verdictly. They rebuilt its answer card: a visible model version and run date, the actual source passage quoted beneath any citation, and a plain amber "not yet confirmed" strip that only clears once a person has checked the specific citation, not just opened the page.

AUDIT, five checks that separate a claim from a factNot a general AI-skepticism lecture. AUDIT is what tells you exactly which fact a good screen is hiding.

Any one of these three, alone on a screen, is worth a second look before it ships.

The recap, one line per letter: ask who paid means the vendor's own accuracy claim isn't evidence, uncover the eval set means no named test set makes a score meaningless, demand the version pin means an untraceable answer can't be fixed, isolate what's missing means the absence of confidence and sourcing is the real design flaw, and test it yourself means the firm's own forty questions are what finally surfaced the fifteen percent failure rate.

And if you want to be sure it really works, try it somewhere elseSame five checks, a hospital's medical coding AI instead of a law firm's research tool. A different missing signal breaks the second story.

Briarstone Health uses CodeAssure, an AI tool that suggests billing codes for patient encounters, displayed in the exact same grid and font as codes pulled straight from the verified CMS code database. Soren Halvorsen, a compliance auditor there, runs the same critique on CodeAssure's suggestion screen. Mapped onto AUDIT: ask who paid means CodeAssure's vendor cites its own internal accuracy number with no outside review; uncover the eval set means that number was never tied to a named, published set of real encounter notes; demand the version pin means the suggested-code grid shows no model version, so a wrong code can't be traced to a specific update.

The isolate step is where this story diverges from Petra's. Verdictly's screen was missing a confidence signal. CodeAssure's screen is missing something else: it never shows which part of the patient's encounter note the suggested code was actually drawn from, so a coder has no way to spot-check a code against the specific line that justified it. Soren's test-it-yourself step pulled fifty recent claims and hand-checked each suggested code against the full encounter note.

Same shape of fix, ten weeks later, in a hospital billing office instead of a law firm.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "strip the borrowed authority, add a version pin, a source excerpt, and an unconfirmed flag," and stop.
Cost: there's no budget this quarter for a full redesign. Say so honestly, and ship the unconfirmed flag alone first, since a single visible flag costs almost nothing and closes the biggest gap.
The model gets better, for real: if Verdictly's accuracy genuinely improves to near-zero fabrication, that's still not a reason to remove the flag. It's a reason to watch the fabrication rate even more closely, since a better model makes the rare miss even easier to miss.

Where people run it wrong.
They critique the model's accuracy and never look at the screen it's rendered on.
They accept a vendor's own reliability claim without asking what it was tested against.
They add a confidence score but keep the same borrowed visual weight, so the number gets ignored anyway.

How to use it live. When someone hands you a screen to critique, ask one question before anything else: if I covered up every word and just looked at the shapes, fonts, and boxes, could I tell which claims have been checked? If not, that's the actual flaw, not whatever the words say.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0