Example: Irina, a lawyer — one knot, not the whole workbook

Irina is a composite character; the situation is ordinary. This is a short example: not a full pass through a workbook but one painful knot — checking, and cognitive debt. Checked: 2026-09.

The full passes are Marina the marketer and Pyotr the accountant. This one is different: what to do when AI is already built into the work, saves real hours, and is dangerous for exactly that reason.

Irina is 36 and handles contracts in-house. She uses AI every day: reviewing incoming contracts, drafting replies, searching case law. The saving is real — an hour per contract instead of three. So is the problem: she once cited, in a written opinion, a provision the model named confidently and which does not exist in that wording. A colleague caught it, not her.


What the mode audit showed (chapter 8)

Irina worked through the AI working mode audit and got an unwelcome but accurate result.

Question on the sheet Honest answer
Do I read the model's answer in full Diagonally, if it looks well structured
Do I check citations against the source I used to; lately only sometimes
Could I do this work as fast without AI Not any more
Has it become harder to formulate things myself Yes — I catch myself waiting for a draft

Mode: self-automator in analysis, centaur in negotiation.

Why this is an honest answer: the temptation was to write "centaur" everywhere, which reads better. Irina separated the areas: where she holds the context and argues, the mode is healthy; where she accepts finished work, it is not.

Where the debt actually accumulates

Not in the AI writing drafts. In the fact that the checking weakened unnoticed: first selectively, then "by how convincing it feels".

That is precisely the trap the cognitive debt checklist warns about: how convincing a text is and whether it is correct are different things, and the first grows faster than the second.

What she changed

  • No exceptions rule. Every reference to a rule or a case is opened in the source before it reaches a document. The model's wording can be used; its citation cannot.
  • Two separated modes. "Draft" can be fast and AI-assisted. "Signed opinion" starts with her own position in three lines, and only then the model — as an opponent, not an author.
  • One review a week without AI. Not on principle, but to see whether she can still formulate things herself. That is friction instead of a finished answer.
  • A skill of her own for contract review, with a required line — a quoted clause for every risk — so the check is built into the answer format.

The three-month check

Not "does it feel better" but measurable: how many times in a quarter a citation she did not open herself reached a document under her signature. Target: zero. In the first month, two — both caught by her own check.


Why this is not an invented problem

Citations that do not exist are a measured effect, not a scare story. The Stanford RegLab and Stanford HAI study "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools" (Magesh, Surani, Dahl, Suzgun, Manning and Ho; preprint 2024, published in the Journal of Empirical Legal Studies, 2025) tested purpose-built legal AI tools — Lexis+ AI and Westlaw AI-Assisted Research — and recorded hallucinations in 17% to 33% of answers. These are tools built specifically for law and sold on a promise of reliability.

The lesson for the worksheet is simple: the more convincing an answer looks inside your own field, the more it needs an outside check. Professional accountability does not transfer along with the work — the signature stays yours.

Next: the AI working mode audit · the cognitive debt checklist.

Example: Irina, a lawyer — one knot, not the whole workbook