Main idea

When you give an AI several pieces of customer feedback on the same topic and ask for a summary, it looks for what they have in common, merges whatever looks alike and produces a well-written, nuanced compromise that gives the impression of having taken everything into account. That is exactly what it was asked to do.

The tension between a dock that works around data entry and an accounting team that needs it then vanishes into an elegant sentence. The AI makes no error: it carries out the summary instruction, and that instruction is precisely the one that smooths things over.

Why it matters

The temptation to smooth things over grows with AI, because smoothing becomes instant and well written. A summary instruction addressed to a model should therefore be read as a request for reconciliation, unless it explicitly says to keep the oppositions.

Nuances and limits

The flaw lies as much in the instruction as in the model: a model asked to list the contradictions lists them. But the default instruction, "give me a summary", remains the most common one.

Open questions

• What summary instruction keeps the oppositions without producing a list nobody reads?