You upload a picture of an empty room and type one line: fill this room with everything you know about me. The room comes back furnished with objects picked out of your own past conversations, and you can see, in one glance, how much an assistant has been keeping about you.
What is actually going on
The prompt has been spreading for a few days. A popular account on X told followers to try it and post the results, and technology news sites picked the trend up on 8 November. Most people who tried it were curious rather than worried. Then they saw the room.
Memory in a chatbot works like a shop assistant who keeps a small card on you behind the counter. Every so often they write a note — you prefer the window seat, you have a dog, you were asking about a mortgage — and the next time you walk in they read the card before saying hello. It is helpful, and it is also a record.
OpenAI, which makes ChatGPT, says the assistant can draw on past chats, saved memories, the instructions you gave it, files in its Library and even connected apps such as your Gmail. It does not keep every word, and it decides for itself what is worth remembering. So the room is not a transcript. It is a summary of what the assistant judged important.
That is why the picture unsettles people. Nothing was hacked. Every object came from something the person typed willingly, months ago, and then forgot.
Why it matters to you
A summary of you is what a scammer needs, because a scam works when the story fits. A message that mentions your dog’s name, the town your sister lives in and the holiday you booked in June reads like it came from someone who knows you.
Researchers at the University of Texas at Arlington measured how little it takes. They studied 200 public Instagram accounts and found that as few as ten posts were enough for AI tools to write personalised scam emails about a recent trip or a concert the person had mentioned. Volunteers rated the fake emails highly, and standard safety filters caught between 0 and 19 percent of the requests.
Those emails did not need a chatbot’s memory. They needed public posts. Put the two together — what you publish and what your assistant remembers — and the story becomes very convincing.
Consumer agencies have started saying this out loud. On Monday, New York’s Division of Consumer Protection warned that anything typed into a chatbot should be assumed to be collected, possibly used to train future models, possibly read by a human reviewer, possibly exposed in a breach.
What you can actually do
None of this needs a technical fix. It needs ten minutes and one decision.
- Ask the assistant, in your own words, what it remembers about you. It will give you a short summary. Read it the way a stranger would.
- Open Settings, then Personalization, then Memory. That is where the summary and the saved memories sit, and where you can delete a line that is wrong or that you would rather not keep.
- If you want it to forget the summary, use Delete and turn off memory. Be clear about what that does: it removes the remembered information in the summary and switches memory off. It does not delete your old chats.
- Look for the switch called Reference chat history. Turning it off schedules the material remembered from past chats for deletion within 30 days. Saved memories stay until you delete them, so do both.
- Do not post the room picture, or a screenshot of your chat history, in public. A private account helps, but anything you post can be screenshotted and passed on by someone else.
- Keep ID numbers, bank details, PINs, passwords and medical results out of the chat in the first place. Information cannot leak from a place it was never written to.
- If a message arrives that quotes something you told an assistant, treat the detail as a coincidence until you have checked the sender another way — the number on your bank card, the app, or a call you make yourself.
Common misunderstandings
Myth: it is only a picture
The picture is a picture. The objects in it are notes. Posting it is closer to publishing a page from your diary than to sharing a drawing.
Myth: deleting the chat deletes what it learned
A saved memory can live apart from the conversation that produced it. OpenAI’s own guidance says deleting a chat does not necessarily remove a memory created from it, and that logs of deleted memories may be kept for up to 30 days for safety and debugging.
Myth: I have nothing worth remembering
Nobody needs your bank password. A name, a street, a week away from home and the school your child attends are enough to write a message you would answer.
What to watch for next
Expect more of these prompts. They are entertaining, they spread quickly, and each one asks you to publish what a machine worked out about you. Some are harmless, and you cannot tell which is which before you post.
Expect more scams that quote your own words back at you, too. A real detail, a believable reason to click and a request that feels urgent is a cheap combination to produce and a hard one to spot.
The one thing to remember
An assistant that remembers you is safe only up to the last thing you told it and the last thing you posted. Look at the memory list once, clear what you do not need, and keep the room picture to yourself.
Image credit: “Empty apartment room with corner windows” by aismallard, CC BY-SA 3.0, via Wikimedia Commons.