OpenAI Got Ten Emergency Police Requests And Answered One

A Washington Post investigation this week found chatbot logs cited in a dozen court cases over the past two years, and the reaction to it has mostly been anecdotal. A student asked about damage to seventeen vehicles. A teenager suing social media companies had his ChatGPT history pulled into discovery. A man in Palm Beach County described plans to hurt an ex girlfriend and OpenAI called the police.

Those are stories. Underneath them there is a table, and OpenAI publishes it itself.

We pulled all four OpenAI Reports on Government Requests for User Data covering January 2024 through December 2025 and read them end to end. The numbers are not ambiguous and they have not been assembled in one place before.

Search Warrants For Chat Content, Four Periods

A content request is legal process, normally a search warrant, for what you actually typed and what the model actually said back. OpenAI reports how many it received, how many it complied with, and how many accounts had content handed over.

Eight to seventy five in eighteen months. That is 9.4 times the volume. Year over year, the second half of 2025 against the second half of 2024, content requests are up 5.8 times and the accounts whose chat content was handed to a government are up 4.9 times, 17 to 84.

Across all four periods, 153 accounts have had their actual conversation content disclosed to a government agency.

The Non Content Column Is Bigger And Nobody Looks At It

A non content request is a subpoena for the account: name, email address, payment information, transaction history. No conversations. It is also roughly three times as common.

Twenty nine to two hundred and twenty four. Both columns are roughly doubling every six months and have done so for four consecutive periods, which is the part that makes this a trend line rather than a spike.

In the second half of 2025 alone, 392 accounts had some category of data handed to a government agency: 307 non content, 84 content, one emergency.

The Emergency Line Is The Story

Emergency requests are the narrow channel. OpenAI describes them as disclosures made where it believes disclosure is necessary to prevent an emergency involving death or serious physical injury. No warrant required. Here is the whole history.

Read the last row twice. Ten emergency requests arrived. One was answered.

For the two prior periods the compliance rate on emergency requests was 100 percent. In the second half of 2025 it was 10 percent.

There are only two readings of that and both matter. Either law enforcement started sending emergency requests that did not meet the death or serious injury bar, which means the channel is being tested, or OpenAI tightened what it accepts, which means the channel is being defended. The report does not say which. It is the single most consequential number in the document and it comes with no explanation attached.

Compliance Rates, Which Are Not What People Assume

The common assumption is that a platform hands over whatever it is asked for. The four reports say otherwise, and they say it consistently.

• Content requests, share complied with: 75.0, 76.9, 88.5, 82.7 percent.
• Non content requests, share complied with: 62.1, 75.4, 68.1, 65.2 percent.

Roughly a third of subpoenas for account information produce nothing. Content warrants, which require a higher legal showing, succeed more often than the lower bar requests do. That is the correct shape and it is worth noting when a platform gets it right.

The Number That Is Not A Number

FISA requests and National Security Letters are excluded from every table above, because the United States government requires them to be reported only in aggregated ranges. OpenAI’s disclosure for July through December 2025 is that the total national security process it received was 0 to 249, affecting 0 to 249 accounts.

Against 224 non content requests and 75 content requests actually itemised, an unresolvable band of 0 to 249 could be nothing at all or it could be larger than every ordinary criminal request in the report combined. There is no way for a reader to tell, and that is by statutory design rather than by OpenAI’s choice.

Why This Is Different From The Email Era

People have been handing data to platforms for twenty years and the transparency report is an old genre. Two things make this one different.

First, there is no privilege. A conversation with a lawyer is protected. A conversation with a doctor is protected. A conversation with a chatbot about a legal problem or a medical symptom is a business record held by a company, discoverable in civil litigation and obtainable by warrant in criminal cases. Nothing about the intimacy of the interaction changes its legal status.

Second, the content is different in kind. Search history is a list of questions. Email is correspondence with another person, who knows what you wrote. A chat log is an unedited transcript of someone reasoning out loud, on the assumption of privacy, often about the exact subject that later becomes the case. In the New York Times copyright litigation a judge has already ordered OpenAI to preserve deleted chat logs, which OpenAI objected to on the ground that users were forced to forgo privacy protections it had put in place.

Deleting a conversation is a request to a company, not an erasure.

What The Curve Says About 2026

Four periods, each roughly double the last, on both content and non content. If that holds, the first half of 2026 report lands somewhere near 150 content requests and 450 non content requests, and somewhere north of 700 accounts.

That is an extrapolation and it is offered as one. The reports for 2026 are not out. What is not an extrapolation is the four rows that already exist, and the emergency line inside them that went from one request answered out of one, to one answered out of ten, without a word of explanation.

Sources: OpenAI Report on Government Requests for User Data, four editions covering January to June 2024, July to December 2024, January to June 2025 and July to December 2025, published at cdn.openai.com/trust-and-transparency. All figures read directly from the tables in those PDFs on September 5, 2026. Growth multiples, compliance rates and account totals are our arithmetic on those figures. The chatbot logs in twelve court cases figure and the Palm Beach County and social media discovery examples are from the Washington Post investigation reported the week of September 1, 2026. This site is not a law firm and this is not legal advice.