A Court Ordered 20 Million ChatGPT Conversations Released — Including 'Deleted' Ones
On January 5, 2026, a federal judge affirmed an order for OpenAI to hand over 20 million ChatGPT conversations to the news organizations suing it — including chats people believed they had deleted. Your privacy now rests on a handful of settings almost nobody turns on.
The Promise
The Risk
A court order overrode “delete”
On January 5, 2026, Judge Sidney H. Stein affirmed an order requiring OpenAI to produce 20 million de-identified ChatGPT logs to the news organizations suing it in the copyright case. Some of those conversations belonged to people who had hit delete and assumed that was the end of it. It was not. For months, a legal hold kept records alive that users thought were gone.
That is the part to sit with. “Delete” in a consumer AI tool is a request, not a guarantee. A court order, a legal hold, or a retention policy can override it, and you will not get a notification when it does.
What actually happens when you hit enter
Every prompt you type does three things: it is transmitted, it is stored, and on most consumer accounts it is used to train the model by default. None of that is hidden, exactly — it is in the terms. But almost nobody reads the terms, and the defaults are set to the vendor’s benefit, not yours. OpenAI’s standard consumer retention runs 30 days, longer when a legal hold says so.
The words you would never put in an email — a health worry, a draft resignation, a client’s numbers — get the same default handling as everything else.
The line nobody is told about
The single most important distinction in AI privacy is consumer tier versus business tier. On most business and enterprise agreements, your inputs are not used for training and retention is contractually bounded. On the consumer tier, the opposite is the default. Same interface, same model, completely different data posture — and the person pasting confidential text into a free account is rarely the person who read which tier they are on.
This is where shadow AI bites. An employee using a personal ChatGPT account for work has just moved company data onto the most permissive tier the vendor offers.
Agentic AI raises the stakes
The newer tools do not just read what you type. They read your screen, your files, your inbox — whatever you connect them to. That is genuinely useful, and it widens the blast radius of every default above. “Anonymized” helps less than it sounds, too: de-identified data can often be re-identified when there is enough of it, which is the entire premise of a 20-million-conversation production.
Three questions before you trust the box
One. Which tier is every AI tool in this organization actually on — and who checked? Two. What is our real answer when an employee asks where their prompts go, and is it the answer the contract gives or the one we hope is true? Three. If a court ordered our vendor to produce our conversations tomorrow, what would be in them? The promise of these tools is real. So is the discovery request that treats your “deleted” chat as evidence.