How the Machine Guesses

A generative AI system does not look anything up. It predicts the next word, then the next, each chosen because it is likely — not because it is true. That single fact explains both why these tools are useful and why they invent authority with total confidence. This lesson lets you feel the mechanism, watch it assemble a perfect-looking citation with nothing underneath, and then puts the discipline on top: the four-step framework and a sorting of what these tools can and cannot do.

How to use this: four parts, in order. Predict a sentence token by token. Watch the citation assembler do the same thing to a case cite. Walk the four-step framework — you already know it. Then sort eight tasks into appropriate use or not-without-verification.

Part 1 · Build a sentence the way the machine does

Pick the next words. The percentages are illustrative likelihoods — note that every option produces fluent legal English, and none of them involves looking anything up.

Pick the next token

Part 2 · Now watch it cite

The same mechanism, pointed at a citation: likely case-name shape, likely reporter, likely year. Every product below is fictional — and every one of them looks exactly right.

Press the button to assemble a citation, one likely token at a time.

What just happened: plausible name + plausible reporter + plausible page + plausible year. Nothing was retrieved; nothing exists behind it. This is why fabricated citations read perfectly — they are built from the shape of real ones. The only test that catches them is the one the machine never ran: retrieval.

Part 3 · The four-step framework — you already own it

The book’s framework for AI-assisted research is the research cycle you have run since Chapter 3, with one extra discipline bolted on: verify before anything enters your record.

Ask one small, precise question

Give the system a role, one discrete task, and the exact format you want back. A vague prompt invites a sprawling answer — and sprawling answers are where errors hide.

= the cycle’s Step 1, unchanged

Verify before you record

Check every claim against the primary source; never cite what you have not read. If verifying costs as much as researching it yourself, write that down — it signals a generative tool was the wrong instrument for this question.

= Step 3, with the discipline attached

Ask whether the question was actually answered

Fully, partially, or not really? The same checkpoint you run on every source. Synthesize the verified findings into your Research Log and annotated outline.

= the checkpoint you already run

Let the answer hand you the next question

Follow up to fill a gap the answer exposed, or move to the next cycle.

= Step 4, unchanged

Part 4 · Sort it: appropriate use, or not without verification?

Eight tasks from the chapter’s two-column table. Make the call on each.

Two kinds of tool

General-purpose chatbot

Generates from its training alone — no live connection to a legal database. Strong for vocabulary, orientation, brainstorming, and working over text you supply.

Verification tax: total. Every legal claim and every citation must be independently confirmed, because none of it was retrieved from anywhere.

Database-integrated AI (retrieval-backed)

Built into a legal research platform; retrieves real documents and generates over them. Citations link to sources that exist, which removes the fabrication problem — not the reading problem.

Verification tax: reduced, never zero. A real case can still be quoted for more than it holds, or be negatively treated. You still read what you cite.