How do I stop AI from making up facts?

The five moves in detail

Move 1: Add the “admit uncertainty” instruction

Single line at the start of your prompt:

If you don’t have enough information to answer confidently, ask me a specific follow-up question rather than guessing. Don’t invent statistics, dates, names or quotes.

This works. Both Claude and ChatGPT are tuned to follow it. Costs nothing.

Move 2: Ground in a source document

Don’t ask “what does Australian Consumer Law say about refunds?”. Upload the actual ACL text and ask “based on the document I uploaded, what does it say about refunds?”.

The AI then quotes the document. Hallucination rates drop close to zero because the model is constrained to what’s actually there.

This is the single highest-use move for any business-critical AI work. It’s why we use Claude Projects + Custom GPTs with knowledge files for repeated workflows.

Move 3: Force structured output

Free-form prose gives the AI room to fabricate. Structured output (JSON, tables, lists) forces explicit field-by-field answers that are easier to verify.

Bad:

Summarise the financial position of this company.

Better:

Return a JSON object with these fields:

• revenue (number, AUD): only fill if a specific number is in the source
• revenue_year (integer): only fill if specified
• profit_margin (number): only fill if specified
• notes (string): any caveats

For any field where the source doesn’t give a specific number, use null. Do not estimate or fabricate.

Source: [paste]

Forcing the schema makes fabrication visible.

Move 4: Use search-augmented mode for time-sensitive facts

Both ChatGPT and Claude have web search modes. For anything recent (current pricing, news, product releases, current personnel), enable search. The model fetches a real source and grounds its answer.

In ChatGPT this is on by default in 2026. In Claude.ai click the web icon. In Claude Code use the WebSearch tool.

Without search, the model’s training data has a cutoff and it’ll fabricate to fill the gap.

Move 5: Verify before publishing

The hard truth: even with all four moves above, you should still verify any specific number or name before publishing. AI is a draft tool, not a research tool. The verification step takes 30 seconds and is the difference between embarrassment and trust.

Build a checklist for your editorial workflow:

• Every statistic has a source link?
• Every person/company named is real and accurately described?
• Every quote is from a real source?
• Every date is correct?

For client-facing work, run this before publishing. Always.

The risk categories ranked

From highest to lowest hallucination risk:

Content type	Risk
Specific statistics (“65% of Australians…”)	Very high
Citations to academic papers / studies	Very high
Specific dates of historical events	High
Specific quotes (“As X said in 2024…”)	High
Specific product features that don’t exist	Medium
General factual claims	Medium
Conceptual explanations	Low
Creative writing	None (by definition)
Code that compiles	Low (you can test it)

Treat the top four with extreme suspicion. Verify everything.

The Australian-specific gotcha

AI models are largely trained on US data. They confidently make up Australian-specific facts: ABN numbers, ATO references, AUD amounts, Australian addresses, AHPRA rules.

For Australian content, source-ground harder. Upload actual ATO documentation, actual ASIC rules, actual current pricing. Don’t trust the model’s training-data knowledge of Australia.

See also

• How to write better AI prompts for the foundational prompt patterns.
• Hallucination for the technical definition.
• How to use AI for SEO for the verification workflow we use in editorial.