Ground Truth.
AI, checked against the source.

Learn · Beginner

Why does AI make things up?

Ask a language model for a quote, a citation, or a date it doesn't actually know, and it will often hand you one anyway — fluent, specific, and wrong. This is called hallucination, and it's the single most important thing to understand about why you can't take an AI's confident answer at face value. The unsettling part is that it isn't a glitch some future update will simply remove. It falls directly out of what these models are and how they're trained.

Why it happens

A language model is, at heart, a system for predicting plausible text. Trained on enormous amounts of writing, it learns what words tend to follow other words. When you ask a question, it doesn't look up an answer in a database — it generates the most likely-sounding continuation. Most of the time, the most likely-sounding continuation is also true, because truthful text is what it mostly saw. But when the model doesn't know something, it has no internal "I'm not sure" alarm to fall back on. Producing a confident guess looks, statistically, just like producing a real answer. Fluency and truth are different things, and the model optimizes for the first.

It gets worse: models can learn to repeat common human misconceptions, because those appear all over the training data. The TruthfulQA study showed that models often mimic popular falsehoods — the confidently-wrong things people say online — rather than the boring truth. And the training that makes a model agreeable and helpful can quietly push it toward telling you what sounds good over what's accurate, a tendency closely tied to how we do reward-based fine-tuning.

An analogy

Imagine a brilliant improv actor who has been told the show must never stop. Hand them any prompt and they'll produce a smooth, in-character response — whether or not they know anything about the topic. Asking them "what year did this obscure treaty get signed?" doesn't trigger "I don't know"; it triggers a confident, plausible-sounding year, because their whole job is to keep the scene going. A language model is that actor. The smoothness you find so impressive is exactly the mechanism that papers over the gaps.

Why it's hard to catch

The danger of a hallucination is that it carries no warning label. As a broad survey of the problem lays out, hallucinated text is grammatically perfect and internally consistent — it looks identical to a correct answer. Automated checks that watch for crashes or malformed output sail right past it. This is the same trap that makes AI agents so tricky: when a tool quietly fails, the model's instinct to always produce fluent language can weave the error into a believable story. And it's why a single AI-graded score is shaky — the grader in LLM-as-a-judge setups can itself be fooled by confident, fluent nonsense.

How people fight it

There's no cure, but there are real defenses:

Why it matters

Reliability is the gap between an AI demo and an AI you'd trust with real work. The whole debate over whether one model makes things up less than another is really a debate about hallucination — and it's genuinely hard to measure fairly. The takeaway is a posture, not a fix: when an AI gives you a smooth, certain answer, smoothness and certainty are not evidence that it's right. Sometimes they're exactly the symptom to worry about.

Key papers
Survey of Hallucination in Natural Language Generation (Ji et al., 2022)
TruthfulQA: Measuring How Models Mimic Human Falsehoods (Lin et al., 2021)
SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection (Manakul et al., 2023)