How to Fix Claude Hallucinations & Fake Citations

Claude occasionally generates incorrect historical dates, fabricated academic citations, or non-functional web links. These errors, known as hallucinations, occur because large language models generate text based on statistical probability rather than a direct lookup database.

Because Claude does not have real-time access to the live web in its base state, it will try to predict what a plausible URL or source looks like when pressed. You can drastically reduce and almost entirely eliminate these hallucinations by structuring your prompts to enforce strict factual grounding.

1. Feed source documents directly into the prompt (Grounding) Never ask Claude to pull facts or citations from the open web or from its pre-training data if accuracy is critical. Instead, provide the exact source documents you want it to reference.

• On the Web App: Upload PDFs, TXT, or CSV files directly to the chat using the paperclip icon.
• Via the API: Pass the reference text inside XML tags like <source_document>.

Once grounded, instruct Claude to *only* use the provided data:

`text Use only the information contained in the <source_document> tags below to answer the question. Do not assume, extrapolate, or bring in external knowledge. `

2. Enforce strict negative constraints By default, Claude is designed to be helpful, which sometimes causes it to generate a "best guess" answer when it lacks information. You must explicitly override this helpfulness bias with negative constraints.

Add this exact block to your prompt instructions:

`text If the answer to the query cannot be found with absolute certainty inside the provided source text, you must say 'I do not have enough information to answer.' Do not attempt to guess, extrapolate, or suggest external sources. `

This gives the model permission to fail gracefully instead of hallucinating a plausible-sounding answer.

3. Force two-step verification (Chain of Thought) Forcing Claude to find and display its source material *before* writing its final response forces the model's self-attention mechanism to align with facts. This is often called Chain-of-Thought (CoT) prompting.

Structure your prompt to require quotes:

`text For every claim you make in your final response, you must follow this two-step process: 1. Locate the exact paragraph in the source text that supports the claim. 2. Quote that paragraph word-for-word in an <evidence> block. 3. Only after displaying the quote, write your analytical response. `

If the model cannot find a direct quote to output first, its probabilistic calculation shifts, preventing it from generating the hallucinated claim in the subsequent step.

4. Lower the Temperature setting (API & Workbench users) If you are accessing Claude through the API or the Anthropic Console, the `temperature` parameter controls the randomness of the outputs. A high temperature leads to creative but highly hallucinated responses.

• For tasks requiring factual accuracy, data extraction, or strict synthesis, lower your temperature setting to 0.0 or 0.1.
• This forces the model to choose the most highly probable, deterministic tokens, substantially cutting down on creative leaps and false citations.

5. Prevent URL and citation generation from thin air Claude cannot verify if a URL is active or dead. If you ask it to generate sources for a research topic, it will often concatenate real domain names with fake directory paths based on plausible URL structures.

To prevent this, use this prompt rule:

`text Do not output any hyperlinks, URLs, or markdown links unless they are explicitly written out word-for-word inside the source text provided in this chat. If a link is not in the source text, do not generate it under any circumstances. `

When to escalate If Claude is consistently generating completely incoherent data or ignoring direct, copy-pasted source text in clean, single-turn prompts, there may be an underlying API bug or a system-wide model performance degradation. Check the Anthropic developer forum or the developer status portal to see if other users are reporting anomalous model regressions.