What AI does not do: applicability, strategy, accountability

AI for lawyers and contracts · Lesson 2 / 20

The boundary runs by type of question, not by difficulty

A model answers the question of what the text says very well, and the question of what it means for us very badly. The first question is closed by the document. The second needs facts the document does not carry: how the parties have behaved before, whether speed or protection matters more today, what you are willing to trade, what is customary in your industry and jurisdiction. The model knows none of that and, worse, does not signal that it does not know.

Three zones that stay human

Applicability. The same liability cap can be perfectly reasonable in one context and gut the entire commercial logic in another. Circumstances decide, not wording.
Negotiation strategy. Which edits to push hard, which to trade away, where not to fight at all — that is bargaining with people you know, not a text analysis problem.
Accountability. The signature, the approval and the consequences remain yours. Saying that the model said so does not work with your management, your counterparty or a regulator.

A practical test

Before accepting any model output, ask:
1. Can this be verified against the document text? -> usable
2. Does it need facts outside the document?        -> I verify myself
3. Does it need a citation to law or case law?     -> treat as invented
   until I have opened the primary source
4. Does it affect money, deadlines or exit?        -> double check

A useful mental model: treat the AI as a very fast, very well-read first-year assistant who will never say the words "I don't know". Anything you would hand such an assistant without checking afterwards, you can hand the model. Everything else, you cannot.

Insight. The model cannot stay silent. For it, no answer is worse than a wrong answer, so it fills emptiness with plausible prose.

Common mistake. Accepting a verdict of low risk with no reasoning. Demand a clause number and a mechanism of harm behind every risk rating, or the rating rests on nothing.

Pro tip. Add to your prompt: if the text does not contain enough information, write NOT DETERMINED and name the missing fact. That converts a share of fabrication into honest gaps.

Cheat sheet

What the text says is for the model, what it means is for you.
Applicability, strategy and accountability do not delegate.
Demand a clause and a harm mechanism behind every rating.
Give the model permission to answer not determined.

1. Which question does the model answer most reliably?

2. What do you do with output that depends on facts outside the document?

3. Why does the model not report its own ignorance?

Task — checked by AI

Take the last contract you negotiated. Write down five conclusions you reached about it. Sort each into one of three buckets: verifiable against the text, requires external facts, requires a citation to law. Then describe what changes in your workflow if the first bucket goes to the model and the other two stay with you.

← Back

🔒 Answer the question correctly to move on to the next lesson.