How to Keep AI From Hallucinating: Sources and Checks

To keep AI from hallucinating, give it relevant sources, tell it to leave unsupported details unanswered, and verify the claims you will rely on. These steps reduce invented answers; they cannot guarantee an error-free response. A fluent paragraph still needs evidence when it names a requirement, number, source, or project decision.

fenbs helps with the project side: it keeps checked decisions, rules, and work context outside a single conversation so another assistant can read the same record. You still decide which facts are verified. Start with the following routine for any AI app you use at work.

What should you do before asking for an answer?

1. Define the deliverable and its scope. Name the audience, relevant product or project, and the date or version when it matters.
2. Supply a small set of authoritative sources. Say which document controls if two sources disagree.
3. Require a source location for factual claims and a clear label for assumptions. Allow the answer to say that the material is insufficient.
4. Check important claims against the original sources before publishing, sending, or acting on the answer.
5. Save the verified result with its source and remaining uncertainty, so the next conversation starts from checked facts.

Say you are drafting instructions for a customer refund. Provide the approved policy and ask for a plain-language draft that preserves its conditions. Asking for a friendly refund policy from scratch invites the assistant to fill gaps that only your company can settle.

Which sources should you give the AI?

Use the document closest to the claim: the approved project decision, a vendor's documentation, the original research, or the organization responsible for the requirement. Check that it applies to your question. An older policy or a guide for a different product can be authentic and still be the wrong evidence.

Grounding means giving the answer a reference to work from. OpenAI's accuracy guidance↗ (opens in a new tab) distinguishes missing context from inconsistent behavior and warns that irrelevant retrieved material can obscure useful information. Adding more documents is therefore not a substitute for choosing relevant ones.

Ask the assistant to identify the source passage before drafting its conclusion. If it cannot access a link, provide the relevant excerpt or obtain the document yourself. A URL in a prompt does not establish that the assistant read its contents. A compact source-backed project note fits the AI context fenbs provides.

How do you make uncertainty acceptable?

Specify what to do with missing evidence, rather than only saying to be accurate. Tell the assistant to separate statements supported by the material from its suggestions. Give it permission to ask a focused question when one missing detail changes the answer.

Answer using the supplied sources.
For each factual claim, name the document and section that supports it.
Separate supported facts, assumptions, and suggestions.
If a required detail is absent, say what is missing. Do not invent it.
If sources conflict, show the conflict and ask which source controls.

OpenAI's prompting guidance on ambiguity↗ (opens in a new tab) recommends explicit uncertainty instructions and clarifying questions where requirements are underspecified. Treat this as a useful instruction to test, rather than proof that a particular response is correct.

Avoid asking for a confident answer after the source has run out. If the policy does not state a refund deadline, the useful result is a question for the policy owner. A plausible deadline creates a new rule nobody approved.

How can you check an answer efficiently?

Review claims individually. Open every source you will cite, locate the supporting passage, and compare its meaning with the draft. A real citation can support a narrower statement than the assistant wrote. Check names, dates, calculations, exceptions, and words such as must or always especially closely.

• For a number, find the original value and independently repeat any calculation.
• For a quotation, compare the actual wording and surrounding passage.
• For a software instruction, check the relevant documentation and verify the behavior in an appropriate test environment.
• For a project decision, locate the person's recorded approval; an earlier AI suggestion does not count.

You can ask another assistant to flag unsupported claims, but its agreement is not independent evidence. Let it help find what needs checking, then inspect the source yourself. Building this review into everyday AI work is easier than trying to reconstruct evidence after someone questions the result.

Why do corrected facts go wrong again later?

A correction buried in one conversation is a weak starting point for another. Your next assistant may receive an old summary, a superseded document, or none of the correction. Save what changed, the supporting source, and any limit on where the correction applies.

For the refund example, record that the policy leaves the deadline unspecified and that a person must approve any proposed deadline. Do not save the rejected number as though it were useful background. This is one practical reason to keep an AI brain focused on reliable work context.

How to do this with fenbs

1. Create a free account and add the project whose answers need a consistent factual starting point.
2. Add AI context notes with verified project facts and their source references. Keep unresolved questions explicit.
3. Record the person's chosen verification requirements in Decisions and rules, such as requiring evidence before calling a task complete.
4. Connect a supported assistant and ask it to read the AI context first. Rules appear before progress and project notes.
5. Review its result, correct the reusable record where needed, and have it save Where we left off with the remaining checks.

fenbs does not fact-check sources or eliminate hallucinations. If you only need to verify one public fact, opening its primary source may be enough; fenbs fits the ongoing project whose checked context several sessions need.

How can you tell whether the routine is improving?

Keep a small set of questions with known answers, including one the sources cannot answer and one involving conflicting material. Repeat them after changing the instructions or source packet. OpenAI's workplace prompting guide↗ (opens in a new tab) recommends testing against known answers; include appropriate abstention in your own checks.

For app-specific steps, see ChatGPT hallucination checks; for boundaries on actions, see keeping AI under control. Explore more ways to provide useful evidence in the context engineering guides.