Which Summary Did the AI Write?

Two summaries of the same meeting. One was written by a person in ninety seconds. One was written by a model in three. Read them before you scroll.

**A.** The team reviewed Q3 pipeline health and identified several key areas for improvement. Stakeholders aligned on prioritising high-impact initiatives while maintaining focus on core deliverables. Next steps include following up with relevant parties and revisiting timelines as needed.

**B.** Pipeline is down 18% on Q2. Two causes: the enterprise deals slipped a quarter, and nobody replaced the outbound that Marta was running before she left. Decision: Rui takes outbound from Monday. We look at enterprise again on the 14th.

B is the human. Not because it is better written — because it says something that could be wrong.

The three tells

**1. It never commits to a number.** A says "several key areas" and "high-impact initiatives". B says 18%, Marta, Rui, the 14th. A summary that cannot be contradicted has not summarised anything. It has described the shape of a meeting without its contents.

**2. Every sentence is the same length.** Read A aloud. Three sentences, all about twenty words, all built the same way. Human notes are lumpy — a fragment, then a long explanation, then two words. Rhythm is the cheapest tell and the hardest to fake without being told to.

**3. It ends with a next step that names nobody.** "Following up with relevant parties" is what a model writes when the transcript never said who. A person writes "Rui takes outbound from Monday" or writes nothing at all.

The prompt that fixes it

Summarise this meeting in under 120 words. Every claim must include a number, a name or a date from the transcript. If something was discussed but never decided, write "no decision" instead of describing the discussion. Do not write a next step unless the transcript names who does it. Vary sentence length.

The instruction that does the real work is the third one. Left alone, a model treats "we talked about X" and "we decided X" as the same kind of sentence, because grammatically they are. You have to tell it that one of them is worth writing down and the other is noise.

Why this matters beyond meetings

The same three tells appear in every AI draft you will read this week: the missing number, the even rhythm, the actor-free action. Once you can spot them in a meeting summary, you cannot unsee them in a product brief, a performance review or a strategy memo.

That is the actual skill. Not detecting AI — anybody can run a detector. Detecting empty, which is what the model produces when nobody told it that being specific is the job.

Want the calm version of AI news like this, once a week? Subscribe to the Sharp AI Hub newsletter →