How do I stop an AI writing tool from inventing facts or fake quotes?

You can reduce made-up facts and quotes, but no prompt or setting can eliminate them reliably. Treat the tool as a drafting assistant, not a source of truth; even citations can be fabricated or fail to support the claim. (nvlpubs.nist.gov)

A practical workflow

1. Give it trusted source material. Ask it to work only from documents or links you provide, rather than relying on its memory. For current facts, use a tool that searches and shows source links.
2. Set a strict rule: “If the sources don’t support a claim, write ‘not established by the provided sources’—don’t guess.”
3. Ban invented quotations. Ask for exact quotes only when the wording appears verbatim in a supplied source, with a page number or link. Otherwise, have it paraphrase and label the text as a paraphrase.
4. Require evidence claim by claim. Ask for a table with each factual claim, its supporting source, and the exact passage or page. Then check that the passage really supports the claim.
5. Verify independently before publishing. Open the cited source; check names, dates, numbers, and quote wording against the original. A plausible-looking citation is not proof. Human review remains important. (cdc.gov)
6. Separate drafting from checking. Have the tool draft from sources, then run a separate pass asking it to flag unsupported claims—while still checking those flags yourself.

Prompt you can reuse:

Use only the sources I provide. Do not add facts from memory or invent citations. For every factual claim, cite the supporting source and identify the relevant passage or page. Use quotation marks only for wording copied exactly from a source. If a claim or quote cannot be verified from the sources, say “unverified” or omit it. Do not fill gaps by guessing.

Grounding a model in retrieved sources can help, but it doesn’t replace checking that the output matches those sources. (ai.google.dev)

The most reliable approach combines three things: give the tool real source material, constrain what it's allowed to do, and verify everything before publishing. No setup guarantees zero errors.

1. Supply the facts yourself. Grounding is the main fix. The most effective way to prevent hallucinations is to give the LLM the specific information it needs, rather than relying only on its training data. In practice, paste in your notes, documents, or verified articles, or use a tool that retrieves sources for you (retrieval-augmented generation). Some writing tools let you add live URLs for this purpose. Grounding reduces errors but doesn't eliminate them. A model can still produce output that contradicts accurate retrieved context.

2. Write prompts that restrict the tool.

• Tell it to use only the material you provided, and that any statement must be traceable to that material. Prompt instructions like "only derive information from the provided context" or "cite specific passages for each claim" help.
• Explicitly forbid invented content. Most hallucinations happen because the user never explicitly ruled them out.
• Allow it to say it doesn't know. Models can abstain when uncertain, which reduces fabrication.
• Require citations. When the model knows it must cite, it resists inventing unsupported details.
• Ask it to mark claims it couldn't support, so you can review them. One approach is to list every unsupported claim at the end of the draft.
• For quotes, tell it to use only quotations you've pasted in, and to write placeholders like [QUOTE NEEDED] otherwise.

3. Verify before publishing. This step matters most, especially for quotes and citations.

• Treat every specific claim as unconfirmed until you find it in a source the AI didn't write. Treat every specific claim as unconfirmed until you have seen it somewhere the AI did not write.
• Check each quote against the original recording, article, or transcript. Search the exact wording and confirm the speaker and context.
• Check references in the actual databases or publisher sites. AI tools confidently cite sources that don't exist, or attribute quotes to people who never said them.
• Don't ask the same tool to vouch for itself. A model that invented a citation will usually defend it when asked, because it has no memory of what's real.
• Use claim-detection features if your tool has them. Some flag facts, quotes, and sources for human review.
• Watch for subtle distortions, not just invented facts. For example, a source's hedged "may contribute to" can quietly become "contributes to."
• Check dates. Models have knowledge cutoffs, so recent figures may be outdated or missing.

Practical tip: If you're using a general chatbot, the simplest effective workflow is: paste your sources, instruct it to use only those, require a source reference for every factual sentence, and then check each reference yourself. For high-stakes work (medical, legal, financial, academic), human review is essential.

To prevent an AI writing tool from inventing facts (hallucinating) or fabricating quotes, it is essential to move away from treating AI as a "knowledge base" and instead treat it as a reasoning engine that requires your active supervision.

Because AI models are designed to prioritize fluent and plausible-sounding language over strict factual accuracy, they will often "fill in the gaps" when they lack information. You can significantly reduce this behavior using the following strategies:

1. Refine Your Prompting (The "Guardrail" Approach)

How you ask your questions dictates the quality of the response.

• Forbid Guessing: Explicitly instruct the AI: "You must not generate any citation, reference, or quote that is not verifiable. If you cannot find a source for a specific claim, state: 'I could not find a verifiable source for this claim.' Never invent a source or quote".
• Require Evidence: Use "Source-Grounded" prompting. Tell the AI: "For every factual claim, provide a link to the original source or describe where that information can be found".
• Chain-of-Thought (CoT) Prompting: Ask the AI to "think step-by-step". For example: "First, identify the core claims. Second, look for evidence to support them. Third, summarize the findings. If you are uncertain about any step, state your confidence level".

2. Use "Retrieval-Augmented" Workflows

The most effective way to stop hallucinations is to provide the AI with the facts you want it to use.

• Grounding: Instead of asking an AI to write about a topic from its memory, paste the relevant articles, data, or documents into the prompt and instruct it: "Answer only based on the provided context. If the answer is not in the text, say you don't know".
• Work with Snippets: If you need a quote, provide the full text or transcript to the AI and ask it to extract the quote from that specific source, rather than asking it to "recall" a quote from a famous figure.

3. Implement a Verification Layer

Never publish or submit AI-generated text without a "human-in-the-loop" verification step.

• Traceability: If an AI provides a statistic or a quote, always perform a manual search to confirm it exists. If you cannot find it through a search engine, assume it is an AI hallucination.
• Cross-Reference: Check key facts against at least two independent, reputable sources.
• Confidence Check: If the topic is niche, sensitive, or professional, be extra vigilant. AI is most likely to hallucinate when it is forced to answer questions about topics it was not specifically trained on or where information is scarce.

4. Understand Model Limitations

• Knowledge Cut-offs: Remember that models have a "knowledge cutoff" date. If you ask about current events or very recent developments, the model is highly likely to fabricate details to remain "helpful".
• Model Tier: More advanced models (e.g., the latest iterations of GPT-4, Claude Opus, or Gemini Pro) generally hallucinate less than older or free-tier models. If you are doing high-stakes work, use the most capable model available to you.

Summary Checklist for Your Workflow:

1. Ground the AI: Provide the data/text you want it to use.
2. Give Clear Boundaries: Tell it to admit ignorance if it doesn't know the answer.
3. Verify Everything: Treat AI output as a draft that must be fact-checked, especially quotes and citations.

You can reduce invented facts, but no prompt can eliminate them. The safest approach is to limit what the tool may draw on, then verify the result before using it.

Try a prompt like this:

Use only the sources I provide below. Don’t add facts from memory or invent sources, statistics, or quotations. Attach a source to each factual claim. For direct quotes, copy the wording exactly and identify the source; if you can’t verify the exact wording, don’t quote it. If the sources don’t support a detail, write “not established by the provided sources” or leave it out.

[Paste your sources here.]

Then use a verification pass:

• Check every name, date, number, citation, and other specific claim against the linked source—ideally the original or another authoritative source.
• For a direct quote, find the words in the original document, transcript, or recording and check the wording and context. A plausible quote or a citation that exists isn’t enough; the source must support the exact claim. ¹³
• Remove anything you can’t verify. Don’t treat the tool’s confidence—or its own assurance that a claim is accurate—as evidence.
• If the tool supports source-grounded writing or search, provide the documents or links yourself and ask it to stay within them. Still check the sources yourself.

A good rule: use AI for drafting and organization, not as the authority for facts or quotations. ¹

Sources: ¹ OpenAI Help Center: Does ChatGPT tell the truth? · ³ Purdue Northwest Library guide: AI hallucinations