Writing a press release with AI? Supply the quote yourself.
Tested in GPT-5.5 (OpenAI API), Oct 2026
Most people paste a fact sheet and ask for a press release. Ours was for a fictional bakery, with the owner, date, address, menu and an opening offer, but no quote.
Real output · GPT-5.5 (OpenAI API) · Oct 2026
The draft came back with a quote from the owner, Dana Reyes. She never said it, because the fact sheet had no quote at all.
Real output · GPT-5.5 (OpenAI API) · Oct 2026
We ran the same prompt again and got a different invented quote. A press release quote gets published under a real person's name.
AI fills the gaps a press release usually has
If the facts don't include a quote, it writes a plausible one. Anthropic's prompting guide says to tell the model to use only the documents you give it. That reduces made-up content but doesn't eliminate it.
Source: Anthropic Claude docs, Reduce hallucinations, Oct 2026
Tested in GPT-5.5 (OpenAI API), Oct 2026
The fix has three parts: paste the owner's real quote, say to use only the fact sheet, and ask for a [NEEDED] marker wherever something is missing.
Real output · GPT-5.5 (OpenAI API) · Oct 2026
With the real quote pasted in, it used the quote word for word, added no other quotes, and flagged three gaps instead of guessing. You still fill the gaps and read every line. This was one run.
Paste the real quote first
Before you ask for a press release, write one real sentence from the person being quoted and paste it in. Then search the draft for quotation marks.
Try it: Search the finished draft for quotation marks and confirm each quote is real.
Sources and assumptions
- Reduce hallucinations (Claude docs, Anthropic): Under 'External knowledge restriction': 'Explicitly instruct Claude to only use information from provided documents and not its general knowledge.' The Note says: 'although these techniques significantly reduce hallucinations, they don’t eliminate them entirely. Always validate critical information.' The page's press-release example asks for a draft 'using only information from these product briefs and market reports' and to mark removed claims with empty [] brackets. (checked 2026-10-30)
- Prompt engineering guide (OpenAI API docs): General guidance only: 'It is often useful to include additional context information the model can use to generate a response within the prompt you give the model.' It does NOT specifically say to restrict the model to provided information, so don't cite it for that claim. (checked 2026-10-30)
Assumptions:
- The test was run in GPT-5.5 via the OpenAI API (label it 'GPT-5.5 (OpenAI API), Oct 2026'), not in the ChatGPT app and not in Claude. The Claude runs failed to execute in this research session, so no Claude output exists and the post must not claim one.
- Bakery, owner, press contact, address, phone and email are all fictional (the phone uses 555, the email uses the reserved .example domain).
- There were two 'before' runs and one 'after' run. The result shows what happened in this test, not a rate or a guarantee.
- The Anthropic guide is written for Claude. Present it as general prompting advice from an AI company's docs, not as proof about GPT-5.5. The proof for GPT-5.5 is the test itself.
- The 'after' output still needs a human to fill [NEEDED: release date], [NEEDED: state] and the background, and to check every line before sending.
The short version
- No quote in, invented quote out.
- Paste the real quote word for word.
- Ask for [NEEDED] markers, not guesses.
- Fill the gaps and read every line.


