keepcalculator

06 Numbers and Powers

Reduce AI Hallucinations With Better Prompts and Source Checks

Card No. 06·78234 · Updated · Standard formula, shown below

Clear instructions and deliberate source checks can reduce unsupported claims in AI-generated answers, though they cannot guarantee that every answer is correct. Before you rely on a response, define what it should cover, ask it to show uncertainty, and verify important facts against trustworthy sources.

What You’ll Need

  • An AI tool that accepts detailed prompts. Features for web search, citations, or file uploads are useful but not essential.
  • The question or task you need answered, including any relevant context and boundaries.
  • Access to reliable reference material, such as official documentation, primary sources, or reputable specialist publications.
  • A few minutes to check consequential claims rather than treating fluent writing as proof.

Step-by-Step

  1. Decide what a good answer must do. Specify the task, intended audience, scope, and date or jurisdiction when relevant. For example, ask for a plain-language summary of a named policy as it stood on a particular date, rather than asking vaguely, “Explain the rules.” If the answer depends on location, version, or time period, state those details.
  2. Provide the facts and source material you already trust. Paste relevant excerpts or provide accessible documents, and label them clearly. Say whether the model should use only those materials or may consult other sources. If it cannot access a link or attachment, ask it to say so; do not assume that supplying a URL means it has read the page.
  3. Set evidence and uncertainty rules. Tell the model to distinguish facts from inference, avoid inventing missing details, and say when the evidence is insufficient. Ask it to flag assumptions and conflicting evidence. For factual work, request a citation or a short supporting excerpt for each important claim, not just a bibliography at the end.
  4. Define the output format. Specify whether you want a table, short summary, list of claims, or answer with sources. Set a useful length and ask for dates, names, measurements, and relevant caveats to be included. A format that separates “claim,” “evidence,” and “uncertainty” makes checking easier.
  5. Review the response for claims that need checking. Break a long answer into individual, verifiable statements. Prioritize precise numbers, quotations, legal or policy details, recent events, and claims that could affect a decision. Watch for confident wording that goes beyond the evidence the response provides.
  6. Open and verify each key source. Confirm that the cited page exists, supports the specific claim, and is relevant to the stated time and context. Prefer primary sources, such as original reports, official records, product documentation, or research papers, where appropriate. A search snippet or citation title alone is not verification.
  7. Cross-check consequential or disputed facts. Look for independent, credible sources that agree on the same detail. Two pages repeating the same press release do not provide strong independent confirmation. If reliable sources disagree, record the disagreement and avoid presenting one side as settled without explaining why.
  8. Correct, narrow, or withhold the answer. If a claim lacks support, ask the model to revise it using only verified material, or remove it. If it cannot be confirmed, label it as unverified or leave it out. Keep a note of the sources and the date checked when the result will be reused.

Common Mistakes

Asking for certainty instead of evidence: “Be 100% accurate” does not make an answer reliable. Ask for sources, uncertainty, and explicit limits instead.

Using vague prompts: Broad requests invite the model to fill gaps with assumptions. State the scope, context, and what counts as an acceptable answer.

Trusting citations without opening them: A citation can be irrelevant, incomplete, or incorrectly attributed. Check the actual passage and its context.

Confusing plausibility with truth: Polished language, detailed explanations, and precise-looking figures are not evidence. Verify important details independently.

Relying on one source for a consequential claim: A single source may be outdated or mistaken. Seek independent confirmation, especially for disputed or changing information.

Assuming a model knows what is current: Its stored knowledge may be out of date, and browsing may not be available. Give a relevant date and verify current claims directly.

Tips

  • Ask for a list of factual claims before asking for polished prose. Check the list, then request a draft based on the verified claims.
  • Use prompts such as: “If you cannot support a detail from the supplied sources, mark it unknown rather than guessing.”
  • Ask for exact source titles, publishers, dates, and links where available. Treat missing or vague references as a reason to investigate, not as confirmation.
  • For research, request the distinction between what a source says and what the model infers from it.
  • Keep prompts focused. Splitting a complex task into smaller questions makes unsupported steps easier to spot.
  • For time-sensitive topics, check publication dates and whether the source has been updated. For technical instructions, confirm that the documentation matches the product or version in question.

Troubleshooting

The model gives sources that do not exist. Open each link and search the source title independently. Remove unsupported citations, provide real reference material, and ask for a revised answer limited to it. If browsing is unavailable, request search terms or source types to check rather than fabricated references.

The source exists but does not support the claim. Compare the claim with the cited passage, including surrounding context. Ask the model to quote the relevant lines and explain the connection, then verify that explanation yourself. If the support is still weak, correct or remove the claim.

Sources disagree. Check dates, definitions, populations, and methods; apparent conflict may reflect different scopes. Prefer original and authoritative material for the question at hand, but report genuine disagreement and uncertainty instead of forcing a single answer.

The answer remains vague or overconfident. Narrow the question, supply better evidence, and ask for a claim-by-claim response with assumptions and unknowns. If the evidence cannot resolve the issue, do not treat another rephrasing as proof.

You cannot verify a high-stakes detail. Pause before acting or publishing it. Consult an appropriate qualified source or professional, and clearly state what remains unconfirmed. AI output alone is not a substitute for authoritative verification.

Frequently Asked Questions

Can a prompt eliminate AI hallucinations?
No. Specific instructions and source limits can reduce unsupported content, but models can still misread evidence or produce errors. Verify important claims.

Should I ask the model to cite every sentence?
Usually, citations for each substantive factual claim are more useful than a citation after every sentence. Make sure the sources actually support the claims.

Are official sources always correct?
No source is infallible. Official or primary sources are often best for what an organization published or a rule currently says, while independent sources can help check interpretation and context.

What if the model says it cannot find an answer?
Treat that as a useful limit, not a failure to overcome by demanding a guess. Search suitable references yourself or narrow the question to what the available evidence supports.

How many sources should I check?
There is no fixed number. Check enough reliable, relevant sources to support the claim; for consequential or disputed details, seek independent confirmation rather than counting repeated copies of one source.

Related

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →