In this article
You can’t check every fact, but you can build defense in depth into your workflow so that hallucinations are caught before they reach the outside world.
Here’s a practical hierarchy of controls, from the prompt outward:
1. In the prompt itself: force uncertainty
Instruct the AI to flag its own limits:
“For every factual claim, assign a confidence score (High/Medium/Low). For any claim rated Medium or Low, append [VERIFY] and cite the weakest source. If you are unsure about a date, statistic, or name, write ‘I am uncertain about this’ rather than guessing.”
This turns the model into its own first-pass fact-checker. Humans then only need to review the flagged items, not the entire output.
2. Split the task: generation vs. verification
Don’t ask one model to write and self-critique in the same pass. Use a two-step workflow:
Step 1: Generate the draft.
Step 2: Feed the draft back into the model (or a different model) with a verification prompt. This forces the model to switch from creative generation to analytical verification mode.
“Review this text. Extract every specific claim (names, dates, numbers, quotes). For each, state whether it is supported by your training data, contradicted, or unknown. List any claim that is invented, exaggerated, or unsupported.”
3. Constrain the output to known ground
If you can’t check everything, reduce the surface area:
Ground the prompt. Attach source documents and instruct:
“Only use information from the attached documents. Do not use outside knowledge. If the documents are silent on a topic, say ‘Not covered in source material.'”
Require citations:
“Every paragraph must end with a citation to the specific source document and page/section. If you cannot cite it, remove the claim.” Hallucinations hate citation requirements.
Ban superlatives and vague intensifiers
“Do not use words like ‘first,’ ‘only,’ ‘always,’ ‘never,’ ‘best,’ or ‘most’ unless you can cite a specific source for that absolute claim.” These words are where reputational risk lives.
4. Sampling strategy: ensemble and compare
When stakes are high, run the same prompt through the same model 3 times or through 2 different models. Then use a comparison prompt:
“Here are three versions of the same report. Identify any factual claims that appear in one version but not the others, or that contradict each other. Flag these for human review.”
5. Human oversight at the edges
Since you can’t check every fact, be strategic about which humans check which facts:
- Would this end up in a headline or lawsuit. (Names, dollar amounts, legal allegations, medical advice, quotes attributed to real people.)
- “That’s a surprise”. Humans are intuitive. Be especially wary of good suprises.
- Audit 10% of outputs in more detail. If you’re finding errors, you know to increase the spot checks.
6. “Reputation” check
Consider this for any client-facing content:
“You are a research assistant, not a subject-matter expert. Your job is to summarize and organize information, not to impress the reader with novel insights. If you do not know something, say ‘I don’t have that information’ . Do not construct a plausible-sounding answer. The user’s reputation depends on your honesty about information you can’t access or fact-check.”
This helps to reframe the model’s objective from “be helpful and a people pleaser” to “be trustworthy”.
Bottom line
Make hallucinations visible (confidence scores, citations, contradictions across runs) and concentrate human review on the highest-risk claims (surprising facts, absolutes, attributed quotes).