LLM (Large Language Model)

The mechanism explains the failure modes, so it is worth knowing. The model has no database to look things up in. It has statistical patterns absorbed from an enormous quantity of text. When it answers, it generates the most probable continuation rather than retrieving a record. That is why it can be fluent and wrong at once, and confident either way.

Two consequences follow, and both matter for visibility. Models have a training cutoff, so anything published after it is invisible unless the model can search live. And models generalise: information seen once in one obscure place rarely survives, while information repeated consistently across many sources becomes reliable to them.

For nonprofits that inverts the usual content advice. One excellent page is worth less than the same message stated consistently in many places.

Related reading: What AI can actually do for a four-person team