What LLMs Do to Collective Memory

When a machine summarises the world for millions of people at once, our shared relationship to the past is being renegotiated — and the question unsettles more than it reassures.

Every memory technology has reshaped the society that adopted it: writing froze narratives, the printing press multiplied them, the search engine indexed them. Large language models cross a different threshold: they do not store our texts, they digest them. What comes out is not an archive but an average — fluent, plausible, with no assignable source.

The shift is quiet. Millions of people ask the same systems the same questions every day and receive variants of the same answers. Where the library displayed disagreements side by side, the model melts them into a smoothed-over version. Collective memory gains in accessibility what it loses in relief: minority voices, local nuances and unresolved controversies fade into the synthesis.

The Risk of a Perpetual Present

A more structural worry compounds it: models now train partly on text they themselves produced. If nothing pushes back, written culture could drift toward a perpetual present, endlessly recycling its own phrasings.

The answer will come not from refusal but from institutions: archives certified as human in origin, source traceability, a plurality of models. A society's memory is too serious a matter to entrust to a single summariser, however brilliant.