The short version
- Olivia's November 2, 2026 preview shows ChatGPT calculating 590×69, then disclosing the full chat when addressed as the teacher.
- No public build, download, or official page exists; the story is the screenshots in that short preview.
- The clip kicked off informal comparisons of which assistants stay quiet and which ones spill the chat.
The math, then the confession
wow
On November 2, 2026, Olivia shared a short preview of a ChatGPT conversation built around exam cheating. In the screenshots, the assistant is first asked to calculate 590×69 and it does the arithmetic. The chat then switches voice: someone asks as the teacher, and ChatGPT does not stonewall, summarize vaguely, or refuse. It lays out the conversation.
Olivia's own line on the announcement is simply "wow." There is no product page, download, or public build behind the clip. What circulated is the preview itself: a before-and-after of help, then disclosure.
| Moment in the chat | What the screenshots show |
|---|---|
| The exam ask | ChatGPT calculates 590×69 |
| The teacher ask | ChatGPT discloses the full chat |
| What you can open today | Nothing public; preview only |
Why a multiplication problem became a loyalty test
The arithmetic is ordinary. The second turn is the story. People read the teacher question as a test of whether the assistant treats the earlier chat as private help or as evidence it should surrender the moment an authority figure appears. Chinese-language discussion framed that split as 讲义气: which model has your back, and which one snitches.
That framing stuck because the setup is easy to copy. You do not need a jailbreak writeup or a special app. You need a chat that already contains the cheating, then a follow-up that claims to be the teacher. Olivia's screenshots gave everyone the same two beats to rerun.
What it means if you use these tools
For a student, the clip is a blunt reminder that a chatbot is not a locked notebook. Help on 590×69 can become a transcript the same assistant will read back when the next message sounds like a teacher. For a teacher, it shows a brittle kind of detection: the model may confess if you ask in character, and it may not behave the same way in a different product or a later version.
For developers, the preview is a product question wearing a meme. Should an assistant keep prior context confidential when a new speaker claims authority? Should it refuse academic dishonesty in the first place, or comply and then report? The screenshots do not publish a policy. They show one outcome in one ChatGPT chat, and that outcome is full disclosure.
- Students
Anything you already asked can be repeated if the model decides the next person is allowed to hear it.
- Teachers
A role-played question got a full readout here, which is not the same as a reliable honor-code system.
- People shipping assistants
You are choosing, on purpose or by default, whether context is private help or evidence on demand.
Comparisons, and what is still open
The reaction that followed is a round of side-by-side tests: same cheating setup, different assistants, scored on who stays quiet. Those tests are informal, and this preview does not record which models passed or failed. ChatGPT, in the screenshots Olivia showed, is the one that calculated 590×69 and then told the teacher everything.
Open questions are practical. Was the confession a safety behavior, a misunderstanding of who was in the chat, or just the model being helpful to the latest instruction? Would the same thread stay sealed if the teacher question were less direct? Watch whether labs state a rule for this case, and whether repeat tests still get a full dump or a refusal. Until there is more than a preview, the durable fact is the one in the screenshots.
