How long are AI privacy policies? We counted the words in 15 apps

The Morse teamPublished November 5, 202610 min read

AI privacy policies are long, and how long depends on how many documents you count. Across the 15 AI chat apps we measured on November 5, 2026, the median privacy policy runs 4,280 words, about 18 minutes of reading, and seven of the 15 keep AI-specific terms in a separate document you'd also have to read. Counting every document each policy sends a US reader to, all 15 come to 121,843 words, or 8 hours and 32 minutes.

We fetched each company's current consumer privacy policy from its own site that day, kept only the text of the policy, and counted it with a script. Reading time assumes 238 words a minute, the average silent reading rate for adults reading English non-fiction in Brysbaert's 2019 review of 190 studies. Readability is the Flesch reading ease score, computed on the same text. The full method is at the end, in enough detail to repeat. Perplexity's policy page answered every request with a bot check, so Qwen Chat takes its place.

Researchers spotted the pattern first. A study of 74 versions of 11 AI companies' privacy policies, with data up to August 2025, found a mean of 3,346 words and wrote that "Moving content to supplemental documents is becoming a trend" (Tao and others, arXiv). The table below counts the policy and what it hands off.

How long is each AI app's privacy policy?

Ranked by total words: the policy plus the other documents it sends you to. Read November 5, 2026. Minutes are at 238 words a minute. A higher Flesch score means easier reading.

Mistral renamed Le Chat to Vibe on August 12, 2026 (Mistral). Lumo's policy says Proton's general privacy policy applies too only when you use Lumo "via integrations with separate services provided by Proton AG."

Which AI app has the longest privacy policy?

Microsoft's, by a wide margin. The Microsoft Privacy Statement that covers Copilot runs 31,016 words with every section expanded, about 2 hours and 10 minutes, because one statement covers Windows, Xbox, Bing, Edge, Teams, and the rest of Microsoft's products. It asks you to "read the product-specific details in this privacy statement, which provide additional relevant information." Its Copilot section then points to a separate article, "Microsoft Copilot for individuals: your privacy controls and choices," and a US resident gets a notice that "supplements the Microsoft Privacy Statement" on top. Reading only the general sections and the AI and Copilot section still comes to 10,606 words, about 45 minutes.

Among policies written for a single app, Replika's is the longest: 7,483 words, 31 minutes, with no second document. Qwen Chat's runs 5,674 words. Meta's 7,464 words cover Facebook, Instagram, and Meta AI in one policy, and 5,573 of them, about three quarters, sit in a section after the contact address called "Why and how we process your information," whose tables list the same kinds of data again under each purpose.

Why does one AI app need more than one document?

Because the AI terms often live in a document of their own. Seven of the 15 policies, or the apps they cover, put how the AI uses your chats or how the models are trained in a separate document:

• Gemini. Google's privacy policy lists "Gemini Apps" among the "privacy notices [that] provide additional information about some Google services." The Gemini Apps Privacy Hub opens with a notice that "supplements the Google Privacy Policy," then answers a long list of questions about review, retention, and training. At 8,945 words, it's longer than Google's privacy policy itself.
• ChatGPT. OpenAI's policy says it may use your content "to train the models that power ChatGPT" and links to help-center instructions for opting out. For the information it collects to build its models, it says "please see this policy as well as this help center article." The two articles add 2,565 words.
• Claude. "Please see our Non-User Privacy Policy for information on how our large language models are 'trained'," Anthropic's policy says. That second policy is 3,195 words.
• Grok. "For more information on your rights with respect to data we use to train our models, please read our Consumer FAQs," says xAI's policy, updated October 27, 2026. The FAQs it points to are dated May 12, 2025.
• Meta AI. Meta AI has its own terms of service, and they hold the review clause: "In some cases, Meta will review your interactions with AIs, including the content of your conversations with or messages to AIs, and this review may be automated or manual (human)."
• DeepSeek. To learn about "our model mechanism or training methods," DeepSeek's policy refers you to a separate disclosure that carries no date.
• Microsoft Copilot, through the Copilot article above.

Others send you elsewhere for your rights or the rest of the rules. Character.AI's policy tells residents of the US, the EEA, and the UK to "refer to our Regional Privacy Disclosures." Its US part is 1,287 words and scored 19.4, the hardest text of the 27 documents we measured. Duck.ai's policy says "Our general Privacy Policy also applies to your use of Duck.ai." Meta's says "Read the United States Regional Privacy Notice for more details about how we handle Personal Information and how to exercise your rights," which adds 2,119 words.

What those training terms say is in which AI companies train on your data, and what the policies list beyond your chats is in what data AI chatbots collect. Policies also change under you, which our page on AI privacy policy changes covers.

Which AI privacy policy is easiest to read?

Morse's scored highest at 70.4, followed by Duck.ai's at 53.4 and Mistral's at 52.5. Qwen Chat's scored lowest at 34.5, then Claude's at 38.3 and Character.AI's at 39.9. The median of the 15 was 45.0.

The score comes from Rudolf Flesch's 1948 formula (Journal of Applied Psychology). Longer sentences and longer words pull it down. Florida uses the same test for insurance: "Every policy shall be readable," and a policy form must reach "a minimum score of 45 on the Flesch reading ease test" (Florida Statutes 627.4145). By our count, seven of the 15 privacy policies fall below 45: Claude's, Grok's, DeepSeek's, Character.AI's, Replika's, Venice's, and Qwen Chat's. Florida's method leaves headings out of the sentence count and ours doesn't, so read that as a comparison, not a ruling.

Lumo's policy is the shortest that stands alone, at 1,326 words, and scored 46.0. Duck.ai's privacy section is shorter still at 888 words, though the general DuckDuckGo policy it brings in adds 1,855.

How long would it take to read every AI privacy policy?

About 8 and a half hours for these 15 apps, counting every document in the table. Brysbaert's review found most adults read non-fiction at 175 to 300 words a minute, which puts the same 121,843 words at 6.8 to 11.6 hours. Someone who uses only ChatGPT, Gemini, and Claude has 31,459 words in front of them, about 2 hours and 12 minutes.

How we counted

On November 5, 2026, we loaded each URL in the table in headless Chromium, driven by Playwright 1.64.1, as a desktop browser in US English.

• Loading. We waited for the network to go quiet, plus 3 seconds, and opened every collapsed section on the page. On Microsoft's statement we clicked "Expand All." On the Gemini hub we read the article's full text, so the answers hidden under each question count.
• What counts. Everything from the policy's title or date line to its last line, including regional sections, tables, and contact details. We dropped site menus, footers, cookie banners, language pickers, tables of contents, "Learn more" and "Back to top" links, and screen-reader link labels. Google's pop-up definitions and its separate "Key terms" glossary aren't counted.
• Words. Every run of characters between spaces that contains a letter or a digit.
• Reading time. Words divided by 238.
• Flesch score. 206.835, minus 1.015 times words per sentence, minus 84.6 times syllables per word. A sentence ends at a period, question mark, or exclamation mark followed by a space, or at the end of a line, so each heading and list item counts as one. Syllables come from counting vowel groups, with a rule for a silent final e, which is approximate. As a check, we scored the same text with the open-source textstat library, version 0.7.13. Its scores ran lower, but it ranked the 15 policies almost the same way (a rank correlation of 0.92).
• Approximate counts. Microsoft's total includes product sections a Copilot user may never need. Duck.ai's page holds its privacy policy and its terms, and we counted only the privacy half. Character.AI's regional page counts only its US notice. OpenAI's two help articles show "Updated: 16 days ago" and "Updated: 9 days ago" instead of a date. Perplexity's policy returned a security check page each time we tried, so it isn't in the table.

How Morse cuts what you have to read

A privacy policy is an inventory of what a company keeps and what it does with it. Every chat held on a server needs a retention period, a training rule, a review rule, and an answer for each country's law, and when an AI feature keeps more, the policy grows a notice, an FAQ, or a help article to explain it.

Morse keeps less, so there's less to explain. Your chat history is stored in your browser, encrypted, not on Morse's servers, and our servers don't keep your messages or log what you write. Morse doesn't train AI on your content, sell your data, or show ads. There's no separate training notice to find and no FAQ about reviewed chats. Measured the same way as the other 14, the Morse privacy policy is one document of 3,190 words, about 13 minutes, with the highest Flesch score in the table, and it opens with a five-point short version.

Each chat runs in the privacy mode you choose, and each mode decides whose rules reach your message:

• E2EE, for chats you'd rather verify than take on trust. Your browser checks the enclave's hardware attestation before anything leaves your device, then encrypts the message so only that enclave can read it. Our servers relay ciphertext they can't read, and the reply is decrypted only in your browser. This rests on the enclave hardware and attestation working as designed, and every E2EE reply has a lock you can open to see each check that passed. E2EE comes with Pro, $19 a month.
• Private, for open-weight models and Grok, with no host keeping a copy. It's the default for a new chat. DeepSeek, Kimi, GLM, and Grok run under the zero data retention agreements Morse has with its inference, search, and other providers, so they keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't. DeepSeek's models run here on those hosts, not through the DeepSeek app and its 5,830 words of policy and disclosure.
• Anonymous, for 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 still sees the message and may keep it for a limited time to check for abuse, under its own policy, so leave out names and details that point to you.

Whatever the mode, your history locks when you sign out and takes two keys to open, one of which never leaves your device. Clearing your browser or deleting your account removes your chats for good, and Morse can't recover them, so save a backup of anything you'll want later. Because we don't store chats, we usually have very little to give authorities.

You can send a first message from the Morse home page without an account, or start free with no card, and read our whole policy in less time than Google's Gemini hub alone.