AI transparency reports: what 10 AI companies say about government requests

The Morse teamPublished October 26, 20269 min read

As of October 2026, only two AI-first companies publish transparency reports on government requests for user data: OpenAI and Anthropic. Google, Microsoft, and Meta publish company-wide reports that fold Gemini, Copilot, and Meta AI in with email, photos, and social accounts, without counting the AI separately. We found no such report from xAI, Perplexity, Mistral, Character.AI, or DeepSeek.

We read each company's own latest report, data file, or law enforcement page on October 26, 2026, and copied every number exactly as the company gives it. Where a company publishes nothing, the table says so and links to the page that says when it hands data over instead. This isn't legal advice.

Which AI companies publish transparency reports?

Five of the ten do, and none of the five separates its chat assistant from the rest of what it runs. Read October 26, 2026.

"Content" means what you typed and what came back. "Non-content" means account details such as your name, email, and payment information. Google's chart already shows a January to June 2026 row of 65 requests, far below any period since 2009, so we treat July to December 2025 as its latest full period.

How many government requests does OpenAI get?

766 in the first half of 2026, counting content and non-content requests together, plus 28 emergency requests. That is more than double the 299 of the half-year before, and more than 20 times the 37 of the first half of 2024. Every row below comes from OpenAI's own report for that period.

OpenAI handed over content for 88 of the 146 content requests in early 2026, about six in ten, down from 23 of 26 a year earlier. The reports don't say why. Those 88 disclosures reached 130 accounts.

Does OpenAI give data to police?

Yes, with a warrant for chat content or a subpoena for account details, and without either in an emergency involving a risk of death or serious injury. We quote OpenAI's policy wording, and its rules on telling users, in can AI chats be subpoenaed.

OpenAI also goes to the police on its own, and its reports don't count that. In an August 26, 2025 post, it said conversations from users "planning to harm others" go to a small review team, and "if human reviewers determine that a case involves an imminent threat of serious physical harm to others, we may refer it to law enforcement." Its government request reports list only requests it received.

What does Anthropic's transparency report show?

Very few requests, and the newest report is ten months old. Anthropic's Transparency Hub, last updated July 23, 2026, lists four government request reports, and the latest covers July to December 2025. In that period Anthropic received 2 content requests covering 8 accounts and produced data for both, and 20 non-content requests covering 55 accounts, producing data for 7. It had no emergency requests.

Anthropic is the only company in the table that counts preservation requests: 7, covering 29 accounts. It defines them as requests to preserve an account's data "while the requesting agency obtains valid legal process." It reports national security requests "annually, subject to a one-year delay," so the most recent figure is 0-99 for all of 2024. Its report also says it notifies users "unless we believe we are legally prohibited from doing so, or in extraordinary situations such as emergencies." For how Anthropic keeps and uses Claude chats in the first place, see is Claude private.

Is Gemini in Google's transparency report?

Yes, but Gemini requests can't be told apart from the rest. Google's Global requests report counts 304,915 requests for July to December 2025, covering 748,475 accounts, with some information disclosed in 84% of them. Its downloadable data breaks US requests down by legal process: 27,521 subpoenas (90% with data produced), 22,060 search warrants (88%), and 4,261 emergency disclosure requests (84%). The only product split anywhere in the file is for Enterprise Cloud customers, as Google Cloud Platform or Workspace. Gemini, Search, Gmail, and YouTube share one count.

Google's Gemini Apps Privacy Hub, updated October 23, 2026, says Google processes Gemini data "to meet any applicable law, regulation, legal process, or enforceable governmental request." For US criminal cases, Google's policy on government requests says agencies must "get a search warrant to compel disclosure of the content of communications, such as email messages, documents, and photos." Gemini chats aren't among the examples, and neither page treats them as a case of their own. We cover how long Gemini keeps chats in is Gemini private.

Do Microsoft and Meta count Copilot and Meta AI requests?

No. Both publish large, detailed reports, and neither counts its assistant separately.

Microsoft's law enforcement requests report for July to December 2025 counts 27,412 criminal requests, 5,587 of them from the US. Its spreadsheet shows 5,647 rejected for not meeting legal requirements and 3,552 where no data was found. It also counts 745 emergency requests, and none of them led to content being disclosed. The spreadsheet's title row reads "Requests received for all Microsoft Services," and the report page gives Outlook.com, OneDrive, Xbox Live, and Skype as examples of consumer services. Copilot isn't named. Microsoft deserves credit as the only company here that separates, country by country, the requests it rejected from those where it found no data.

Meta's latest integrity report, updated October 16, 2026, says global government requests for user data "increased by 11.1% from 374,516 to 415,977" in the second half of 2025, and its report page shows some data produced for 77.80% of them. Of the 80,009 US requests, "78.1% of which include non-disclosure orders prohibiting Meta from notifying the target user." Meta's report FAQ says the report covers "Facebook, Messenger, Instagram, Whatsapp, and Oculus." Meta AI isn't on that list. See is Meta AI private for what Meta keeps from your Meta AI chats.

What do AI transparency reports leave out?

Most of what a user would want to know. Read side by side, the five reports have four gaps:

1. Which product. No report separates chats with an assistant from everything else the company runs. OpenAI's covers "users of OpenAI services," so ChatGPT, the API, and its other apps share one line. Google, Microsoft, and Meta put their assistants in with email, cloud files, and social accounts.
2. Whether chats were what police wanted. "Content" covers any content. A Google search warrant may have sought Gmail, and a Microsoft one OneDrive. None of the reports says how many content requests sought AI conversations.
3. What the company volunteered. OpenAI's own referrals to police, described in its August 2025 post, aren't requests, so they don't appear in its tables.
4. Lawsuits. OpenAI's policy says it "is not intended for requests for information or service of legal process by OpenAI users or other civil or criminal litigants." The 20 million ChatGPT logs ordered handed over in the New York Times case came through a lawsuit, not a government request, so no transparency report shows them. We lay out that order in the ChatGPT court order on data retention.

Which AI companies publish no transparency report?

Five of the ten: xAI, Perplexity, Mistral, Character.AI, and DeepSeek. Each says in its own policy that it hands data to authorities when the law requires it, but none says how often.

• xAI. Its law enforcement page, which now names the company SpaceXAI LLC, says non-public user information "will not be released except as required by appropriate legal process such as a subpoena, court order, or warrant." Its request form offers preservation requests, search warrants, and National Security Letters. X's information requests report lists data only through July to December 2021.
• Perplexity. Its privacy notice, last updated July 8, 2026, says it "may share data to cooperate with law enforcement, participate in a legal process, or for legal compliance."
• Mistral. Its privacy policy, effective September 3, 2026, lists "competent courts" and regulators such as France's CNIL among those it may share data with.
• Character.AI. Its privacy policy, effective July 1, 2026, says it may disclose information "to respond to a warrant, subpoena, court order, or other applicable laws and regulations."
• DeepSeek. Its privacy policy, last updated February 10, 2026, says it shares data to comply with "legal process or government requests" and may give "your IP address" to police in an emergency. It also says it will "directly collect, process and store your Personal Data in People's Republic of China." More in is DeepSeek safe.

What Morse has to give a government request

Every number above counts requests to a company that holds your chats on its servers. Your Morse chat history is stored only in your browser, encrypted, and locked when you sign out. Our servers don't keep messages or log message content. Your account holds your sign-in details, plan, balance, and a usage ledger with no content. Because Morse doesn't store chats, it usually has very little to give authorities.

Each chat has its own privacy mode, so you choose who handles each message:

• Private, for chats you'd rather no host keep. It's the default for a new chat. Open-weight models like DeepSeek, Kimi, and GLM, plus Grok, run on hosts with zero data retention agreements, so they keep nothing. Where a host confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
• Anonymous, to use GPT, Claude, and Gemini without the provider knowing it's you. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider may keep the message for a limited time to check for abuse, under its own policy, so leave out names and details you'd rather keep to yourself.
• E2EE, for anything no one should read. Your browser checks the secure enclave's proof that it's genuine, then encrypts each message so only that enclave can read it. Our servers relay text they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed.

Your history is on your device, so that's the place to protect. Unlocking it takes two keys, one that never leaves your device and one our server gives you when you sign in, so a copy of your browser data alone can't be read. Lock your device when you step away. Clearing your browser deletes your chats for good, and we can't recover them, so save anything you need. Morse doesn't train AI on your chats, sell your data, or show ads.

Start a chat from the Morse home page without an account, or sign up free with no card. End-to-end encrypted chats come with Pro, at $19 a month.