Skip to main content

Hallucinations & How to Reduce Them

Intermediate

A hallucination is when a model states something false with complete confidence. It's not lying and not broken — it's the flip side of how LLMs work: they generate plausible text, and plausible isn't always true (see What Is an LLM?). You can't prompt this away entirely, but you can drastically reduce it and catch the rest.

Why it happens​

The model predicts a likely continuation. When it doesn't "know" something, the most likely-looking continuation is often a confident, well-formed — and wrong — answer. There's no built-in "I'm unsure" signal unless you create room for one.

The high-risk zones​

Be most skeptical when output involves:

  • Citations, quotes, and references — fabricated papers, fake URLs, misattributed quotes.
  • Specific numbers, dates, and stats — plausible but invented figures.
  • Niche or very recent facts — beyond what the model reliably learned.
  • APIs and library details — methods or parameters that don't exist.
  • People and legal/medical specifics — high stakes, easy to get subtly wrong.

The reduction toolkit​

Stack these — each one helps:

  1. Ground it in sources. Paste the source text and say "answer only from the text above; if it's not there, say so." This is the core idea behind RAG.
  2. Give it an out. Explicitly allow "If you're not sure, say 'I don't know'" — it dramatically reduces confident guessing.
  3. Ask for reasoning and citations. "Quote the exact sentence that supports each claim." Unsupported claims become obvious.
  4. Lower the creativity for factual tasks where the model exposes a temperature control (see Sampling Controls).
  5. Use tools. For math, current data, or lookups, give the model a calculator/search/tool instead of trusting recall.
  6. Cross-check. Ask the same question two ways, or have a second pass critique the first.

The mindset that actually protects you​

:::warning Verify what matters — always No prompt makes output 100% reliable. For anything consequential — a number in a report, a citation, a command you'll run, a medical/legal/financial detail — check it against a trusted source. Treat AI as a fast first draft, not a final authority. This is the heart of Responsible Use. :::

A simple rule: the cost of being wrong sets the amount of verification. Brainstorming? Trust freely. Publishing a statistic? Verify every time.

Next​