To stop ChatGPT from making up sources, you must explicitly define where the AI should look for information and forbid it from using external knowledge if it cannot find a direct match. This issue happens because the model is designed to predict the most likely next word rather than verify factual truth. By restricting its source material and giving it a clear way to say "I do not know," you can eliminate fabricated links and citations entirely.

It is incredibly frustrating to read a beautifully written report from your AI assistant, only to find that every single link leads to a dead page. You waste valuable time checking links that do not exist, feeling like you cannot trust the tool for serious business research.

Why does ChatGPT make up links and sources?

ChatGPT invents links and sources because it operates on patterns of language rather than an active database of the live internet. When you ask it to provide a citation, it generates text that looks exactly like a real citation, even if the underlying web address is entirely fictional.

When you type a prompt, the AI reads your words and calculates what the most logical next word should be. It does this based on millions of books and web pages it has already read. If your prompt asks for a source about remote work productivity, the AI knows that a standard source looks like a URL from a site like Forbes or Harvard Business Review. It then writes a URL that looks perfectly normal, because that is what its training suggests should come next. The AI is not trying to deceive you. It simply does not understand the difference between a real link that exists on a live server and a string of text that looks like a link.

The technical term for this behavior is hallucination. It is easier to think of the AI as an eager assistant that wants to please you. If you ask for a source, it will write a string of characters that matches the shape of a web link, such as a domain name followed by realistic looking words. It does not actually check if the web page is live unless you explicitly instruct it to use its search tool. Even then, its default behavior is to guess if it runs into a wall.

How to stop ChatGPT from making up sources by setting strict boundaries

You can stop ChatGPT from making up sources by giving the AI a strict set of rules that governs what it is allowed to say when it lacks information. The most effective rule is to provide the exact reference material yourself and forbid the model from stepping outside of that box.

This method is called grounding. By grounding your prompt in a specific source text, you remove the need for the AI to search its memory. You are turning the task from a closed book test into an open book test. When the AI has the book open in front of it, it does not need to guess. It can simply copy the exact quote and URL you provided.

Most people ask the AI to find sources for a topic. This is a weak approach because the AI has to search its vast training data. Instead, you should provide the text or search results yourself. If you are not sure how to structure these instructions, tools like The Prompt Engineer can help. This service asks you for your target audience, goals, and constraints, then builds a custom, copy-and-paste prompt that forces the AI to stick strictly to your source documents. This ensures the AI only works with the truth you provide.

Here is an example of how to change your approach.

Weak prompt

Write an article about the benefits of remote work. Include three real statistics and link to the original studies.

Stronger prompt

You are a research assistant. Write a short article about the benefits of remote work using ONLY the facts provided below.

Constraints:
1. Only use statistics that are explicitly mentioned in the text below.
2. Do not search your training data or make up any new statistics.
3. If you cite a source, use the exact name and URL provided in the source section.
4. If the source section does not contain enough information to support a claim, do not make the claim.
5. If you do not know the answer or lack the source, write "Information not available in provided sources." Do not invent any links.

Source Text:
[Insert your verified source text, articles, or data here with their real URLs]

The essential rules for accurate AI citations

To get accurate citations from an AI, you must explicitly program the rules of engagement into your prompt. This means telling the model how to behave when a source is missing, how to format links, and what to do if it is unsure.

Without these guidelines, the AI will fall back on its default setting, which is to keep writing at all costs. An AI will always prioritize writing a complete paragraph over admitting that it does not have the data.

"An AI will always prefer a believable lie over a blank space unless you explicitly give it permission to say 'I do not know.'"

Here is a simple checklist you can copy and paste into your system instructions or main prompts to keep your assistant honest:

  • Source Constraint: Only cite documents, authors, and links that are explicitly provided in our conversation history.
  • No Guessing: If you cannot find a specific link in the provided text, do not create one. State clearly that the link is missing.
  • Format Verification: Format all links as plain text so I can easily verify them. Do not hide URLs behind anchor text.
  • Negative Constraints: Never use placeholders, fictional domain names, or estimated page numbers.

How to verify links before you publish

Verifying your links before you publish is your final line of defense against misinformation. Even with excellent prompts, you must always double check every URL, author name, and book title before sending your work to a client or publishing it online.

A quick way to do this is to keep a master document of your actual reference links. When the AI generates the draft, cross reference its citations against your master list. If you notice a URL that looks slightly different, flag it. You can also ask the AI to print out a bibliography at the end of the text, listing every URL in full, plain text. This makes it much easier to copy and paste them into a browser tab to ensure they are live.

Common questions

Why does ChatGPT write links that look real but do not work?

ChatGPT writes realistic but broken links because it is trained on patterns of web addresses, not an active map of the internet. It knows that a link about marketing usually starts with a domain name and ends with descriptive words. It simply generates words that fit that pattern without checking if the page actually exists.

Can ChatGPT search the live internet to find real sources?

Yes, modern AI models can search the web if they have browsing features enabled. However, even with web access, they can still hallucinate if they fail to find a direct match. You must still instruct them to only use the exact links they visited during that specific search.

What is the best way to handle a hallucinated source?

The best way to handle a hallucinated source is to delete it immediately and prompt the model again with stricter rules. Do not try to correct the model in the same chat thread without resetting the context. Start a fresh conversation with clear constraints to keep the AI from repeating its mistakes.

Does this issue happen with all AI writers?

Yes, this issue happens with all major language models, including ChatGPT, Claude, and Gemini. They all share the same underlying technology, which prioritizes smooth, natural writing over factual verification. Applying strict constraints is necessary regardless of which tool you choose.

The short version

  • AI models hallucinate links because they predict patterns of letters rather than checking a live web database.
  • To stop this, you must explicitly forbid the AI from using external knowledge outside of the text you provide.
  • Always give the AI a safe exit route by telling it to write "Information not available" instead of guessing.
  • Never trust anchor text; make the AI output raw, plain text URLs so you can easily test them before publishing.

Related reading