Why AI makes things up

Let's go back to Priya's worst Tuesday. She'd asked the AI to punch up a slide: "add a statistic showing email outperforms social for B2B." It obliged instantly: "B2B email campaigns convert 4.3× better than paid social (McKinley Digital Benchmark, 2023)." Clean. Specific. Citable. Into the deck it went.

Mid-presentation, her CFO, who fact-checks for sport, asked for the source. There was no McKinley Benchmark. No 4.3×. The AI had manufactured it. Before we dissect how, your verdict:

Remember lesson one: an LLM has no fact database: it generates every answer by predicting what text plausibly comes next. Priya asked for a statistic with a source. The model has read millions of sentences shaped exactly like that: a percentage, a firm's name, a year. So it predicted one, a sentence with the perfect shape of a statistic, assembled from patterns rather than retrieved from anywhere.

That's a hallucination: plausibility filling a gap. Not lying. Lying requires knowing the truth. Not a bug. The machine did its one job. Watch it happen:

Good news: hallucinations aren't random. They cluster wherever you ask for something specific the model is unlikely to have solid patterns for: precise figures, named studies and citations, niche topics, anything after its training data was collected.

Safe ground is the opposite: summarising a document you pasted in, rewrites, brainstorms, well-trodden concepts. Before trusting an answer, ask: was the model likely to have really seen this, or is it completing a pattern?

Priya now runs three cheap defences. Ask for sources, then check one: a source that doesn't exist is a hallucination unmasked in thirty seconds. Give it permission to not know: add "say 'I don't know', don't guess" to any request for facts.

And verify anything load-bearing. Going in front of a CFO, a client, a contract? Independently checked. Drafts and brainstorms don't need it; facts you'll be quoted on do.