What is an AI hallucination?

What it actually means

The name is unfortunate — it suggests seeing things, as if the model had perception to fail with. What actually happens is more mundane: the model is a prediction engine for plausible text. When it reaches a fact it doesn't reliably hold, it doesn't stop or hedge; it continues with the most statistically likely continuation, which can be a beautifully formatted fabrication. Ask for a citation and it will happily produce one — authors, journal, volume number, all fluent, none real — because "generate a citation-shaped string" and "generate a true citation" are the same task to a model with no truth oracle.

Hallucinations cluster in predictable places: obscure facts the training data barely covered, specific numbers and names, recent events past the model's knowledge cutoff, and anything about your private documents the model has never seen. They are rare where the internet is dense and loud, and thickest where the internet is thin and quiet — which is precisely the terrain of professional work.

Why it matters when you're picking a tool

Because vendors diverge on mitigation, and the difference is a buying criterion. The strongest mitigations are architectural: grounding answers in retrieval with citations (see RAG) so claims come with receipts; refusing to answer when retrieval comes up empty; and explicit "verify" passes. Web-connected assistants hallucinate less on current events than offline models precisely because they check before speaking. For high-stakes domains, this is the whole product: the best AI legal tools are, at bottom, hallucination-control systems with a legal UI.

The real-world stakes are documented, not hypothetical: courts on both sides of the Atlantic have sanctioned lawyers for filings citing decisions that never existed — model-fabricated case law, submitted unchecked. Every one of those was preventable by the four checks below.

The 2026 reality check

Progress is real: grounding, better training and web access cut benchmark hallucination rates substantially year over year, and "I don't know" appears more often than it used to — that's a feature, not a regression. But three traps remain open. Long-tail obscurity: niche entities and small jurisdictions are where fabrication still thrives. Confident compression: summarizing your own document, a model can blend what the document says with what it plausibly would say. And citation theater: sources that exist but don't support the specific claim — subtler than pure invention and more common. Fluency is now a design goal, which means it can no longer be read as a credibility signal.

Quick checklist

• Treat every specific fact — number, name, date, case — as a claim to verify, not a fact you received.
• Prefer tools that cite sources by default, and actually open them: a real source that doesn't support the claim is still a hallucination.
• Re-ask critical questions in a fresh session; inconsistent answers across sessions are a fabrication flag.
• Never ship or file anything where a wrong fact costs money or credibility without a human check against the primary source.

Where you'll hit it

Hallucination risk is a running thread in the best AI chatbots, Perplexity vs ChatGPT (citation-first vs chat-first design), and the AI legal tools ranking. Related terms: RAG, context window — or back to the full glossary.