Learn / AI Training Foundations / Module 5 · 13 min
Factuality, Research & Hallucinations
Fluent language can hide weak factual grounding
AI can generate plausible names, dates, quotations, statistics, citations, or explanations that are wrong. Evaluators must learn which claims deserve verification.
High-verification claims
- Exact dates
- Statistics
- Names and titles
- Legal requirements
- Medical or financial claims
- Scientific findings
- Quotations
- Current events
- Prices and product specs
- Claims about what a company currently offers
Fact, opinion, prediction
Verifiable factual claim
- The film was released in 1999.
- The policy took effect on January 1.
Subjective / predictive
- The film is boring.
- This company will definitely dominate next year.
Prompt
Evaluate this statement.
Response
Alexander Graham Bell invented the telephone in 1876 and was the only person developing telephone technology at the time.
The second clause is an overbroad historical claim. A response can contain both reasonable and misleading components.
When research is allowed
- Identify the exact claim to verify
- Prefer primary or authoritative sources
- Check publication/update date for current claims
- Distinguish source evidence from your interpretation
- Do not use one weak source to support an extraordinary claim
- If reliable verification is unavailable, reflect uncertainty rather than inventing certainty
Module summary
- Identify which claims are verifiable.
- Confidence is not evidence.
- Check current claims against current sources when research is allowed.
- Qualify uncertainty instead of inventing certainty.
Quick check
Which statement most clearly needs verification?
Quick check
If a quote cannot be found in reliable sources, what is a reasonable concern?
Quick check