MorseSign inStart freeOpen Morse

Does Bluesky train AI on your posts? Its pledge, its Attie app, and the open firehose

The Morse teamPublished 11 min read

No, Bluesky does not train generative AI on your posts, by its own account, as of November 13, 2026. That promise sits in a November 2024 post from the company's account, not in its terms of service or privacy policy, which never use the word "train." Your public posts, likes, follows, and blocks also flow through an open firehose that anyone can read without logging in, so other people can and do collect them for AI. Bluesky's proposed opt-out signal, User Intents, still hasn't shipped. Bluesky's own AI app, Attie, has a separate privacy policy that lists training "models" among its purposes.

We read Bluesky's terms of service and privacy policy (both last updated August 14, 2025), its developer docs on the relay, Jetstream, and visibility settings, its developer guidelines, its robots.txt, the User Intents proposal and its GitHub discussion, Attie's beta privacy policy and terms (effective October 15, 2026), and Bluesky's own posts and blog on November 13, 2026. For the datasets built from Bluesky posts we relied on 404 Media's reporting. Bluesky deserves credit for a privacy policy that says plainly that your posts are public and your DMs are unencrypted.

Bluesky AI training: the key facts

Read November 13, 2026.

What it coversYour settingDefaultHow long, or who can get itSource
Bluesky training generative AI on your postsNone neededBluesky says it doesn'tA public statement, not a term in its contractsBluesky, November 15, 2024
Posts, profile, likes, follows, and blocksNone makes them private"Public by design"Broadcast on the firehose to anyone, no login neededPrivacy policy, relay docs
A signal telling AI companies "don't train on me"Doesn't exist yetProposed in March 2025, not shippedWould be voluntary for anyone reading itUser Intents proposal
Showing your posts to logged-out visitorsToggle to hide posts from logged-out usersPosts are shownBluesky honors it; other apps may notVisibility docs
Direct messagesNone"Unencrypted," and accessible "for Trust & Safety purposes"For "the lifetime of your account" or as long as neededPrivacy policy
Attie prompts, replies, and memory (opt-in beta)Don't use AttieUsed to "train and improve models and safety systems""Retained indefinitely," no self-service deletionAttie privacy policy

Does Bluesky use your posts to train AI?

Not for generative AI, Bluesky says. On November 15, 2024, Bluesky's official account posted: "We do not use any of your content to train generative AI, and have no intention of doing so" (Bluesky). The next post in that thread said Bluesky uses AI to help with moderation and in the Discover feed, and that "None of these are Gen AI systems trained on user content" (Bluesky). On November 26, 2024, Bluesky repeated that its stance "has not changed" (Bluesky).

The legal documents are broader. Bluesky's terms of service give it a license to "use your Content to develop, operate, and enhance Bluesky, the AT Protocol, and future products and services" (Bluesky terms). Content there means "anything you share on Bluesky Social," including direct messages. Neither the terms nor the privacy policy mentions AI training, either to allow it or to rule it out. Bluesky's August 2025 post about the update listed changes to age assurance, appeals, and dispute resolution, and nothing about AI (Bluesky). So the no-training promise is a public pledge that the terms neither repeat nor contradict.

The same terms say "We will never sell your Content without getting your permission first." Reddit, by contrast, licenses public posts to AI companies, and its User Agreement names AI training outright, as we cover in does Reddit sell your data to AI.

What does Attie, Bluesky's AI app, do with your data?

Attie keeps what you type indefinitely and lists model training as a purpose. Bluesky showed Attie for the first time at the Atmosphere conference in March 2026 as a stand-alone app built on Anthropic's Claude, which signs you in with your Bluesky account and builds custom feeds from plain-language requests (TechCrunch, March 28, 2026). Attie's terms say it's an invite-only beta for people 18 and older (Attie terms).

Attie's beta privacy policy, effective October 15, 2026, names Bluesky Social, PBC as the data controller (Attie privacy policy). Among its purposes, it lists: "To analyze usage patterns, diagnose errors, train and improve models and safety systems, and develop Attie's features and performance." It calls this "a core purpose of the beta program." It doesn't say which models, or whether they're generative.

What it collects goes beyond the people who use it:

  • Your prompts, Attie's replies, and session data, which are "retained indefinitely" as "a condition of your participation in the Attie beta program."
  • Persistent Memory, which "cannot be selectively deleted by you during the beta." The policy says there are no self-service tools to access or delete Attie data during the beta.
  • Other people's public posts. Attie "processes publicly available data from the AT Protocol network, including data about Bluesky users who have not used Attie," taken from the firehose for its summaries, profiles, and search.

The policy says Anthropic "does not train its general-purpose models on data submitted through the Attie API," and that Bluesky doesn't let Fireworks AI, another model host, train on Attie content. It also says Attie's search provider, Exa Labs, has terms that let it use query data for "training and fine-tuning models." Attie's terms add a "perpetual" license for Bluesky to use what you type into it "for the purpose of operating, developing, and improving Attie and related products and services" (Attie terms).

Can AI companies scrape Bluesky posts?

Yes, and they don't need to scrape. Bluesky's developer docs say its relay merges "every post, like, follow, and reply" across the network into one stream, and "No authentication is required; the firehose is public" (Bluesky relay docs). Bluesky's Jetstream service serves the same stream as plain JSON, and its docs describe "replaying a month of history" as well as live events (Bluesky Jetstream docs).

The website is open too. As of November 13, 2026, bsky.app's robots.txt blocks only the compose page, names no AI crawler, and points anyone "considering bulk or automated crawling" to the firehose (bsky.app robots.txt). We explain what robots.txt can and can't stop in how to block AI training on your website.

On March 20, 2026, Bluesky announced a $20,000 grant to fund Hubble, a mirror of every public repository in the network that promises "complete snapshots for researchers" (AT Protocol blog). Hubble says it will delete records you delete and won't store private app data such as mutes and bookmarks.

Bluesky's rules for developers don't mention AI. Its developer guidelines ban spam and require that "All services must have a method for deleting content a user has requested to be deleted," and say nothing about training or datasets (Bluesky developer guidelines).

Has anyone made AI datasets from Bluesky posts?

Yes, several, starting in November 2024. On November 26, 2024, 404 Media reported that a machine learning librarian at Hugging Face had published one million public Bluesky posts, "collected from Bluesky Social's firehose API, intended for machine learning research" (404 Media). He removed it after the story, writing that "this approach violated principles of transparency and consent in data collection."

That evening, Bluesky posted that it was looking at a robots.txt-style consent setting, and that "Bluesky won't be able to enforce this consent outside of our systems" (Bluesky). On December 3, 2024, 404 Media reported larger datasets on Hugging Face in its place: two million posts, eight million posts, and one of 298 million posts, "non-anonymized" and full text (404 Media).

Bluesky and AI: a timeline

DateWhat happenedSource
May 22, 2024DMs launch without encryption; Bluesky says it intends to add end-to-end encryption "down the line"Bluesky
November 15, 2024Bluesky: "We do not use any of your content to train generative AI"Bluesky
November 26, 2024One-million-post dataset reported, then removed; Bluesky says it can't enforce consent outside its systems404 Media, Bluesky
December 3, 2024Datasets of up to 298 million posts reported on Hugging Face404 Media
March 8, 2025User Intents proposal opened for discussionGitHub
July 19, 2025Draft pull request to show the AI preference in the Bluesky app; still openGitHub
August 14, 2025Current terms and privacy policy published, with no AI training clauseBluesky
March 20, 2026Bluesky funds Hubble, a full mirror of public dataAT Protocol blog
March 28, 2026Attie shown at the Atmosphere conferenceTechCrunch
April 4, 2026The proposal's author says he's tempted to ship "the non-AI bits" firstBluesky
October 15, 2026Attie's beta privacy policy takes effect, listing model training as a purposeAttie

What happened to Bluesky's User Intents proposal?

It's still a proposal, 20 months on. Bryan Newbold, who works on protocol and standards at Bluesky, published "User Intents for Data Reuse" in March 2025 (proposal). It would let you set four preferences, each to allow, disallow, or undefined: generative AI, protocol bridging, bulk datasets, and public archiving. Your choice would be a public record in your account, and the proposal compares it to robots.txt: something "good actors are expected to abide," but "not legally enforceable."

The generative AI intent would cover training large language models and "human language translation," with no exception for research or nonprofits. It would not cover classification, embeddings, recommendation systems, or moderation. Undefined would be the default, and the proposal says apps "should not set intents automatically for users."

On April 4, 2026, Newbold wrote that he had published demo lexicons and a demo site "last summer," and that the AI preferences are "churning as the regulatory, industry, and technical situation matures" (Bluesky). The draft pull request that would add it to the Bluesky app has been open since July 19, 2025 (GitHub). Bluesky's spring 2025, fall 2025, and spring 2026 protocol roadmaps don't mention it. Coverage in March 2025 described the setting as in the works (TechCrunch), but as of November 13, 2026, the Bluesky app has no AI training preference to set.

What's public on Bluesky, and what's private?

Nearly everything you do in the app is public. Bluesky's privacy policy says "most user activity is public by design," including "your posts, profile, likes, following, and blocks" (Bluesky privacy policy). Its protocol team said in August 2026 that, by design, the data stored on the protocol today is public: "every post, every follow, every like, every block" (AT Protocol blog). A block on Bluesky is visible to anyone reading the network.

Direct messages aren't posts, but they aren't end-to-end encrypted either. The privacy policy says Bluesky stores and processes your DMs, and "These are unencrypted and can be accessed for Trust & Safety purposes." Mutes and bookmarks also stay off your public record, according to the Hubble announcement.

Bluesky's toggle to hide your posts from logged-out users changes what its own website shows, and marks your pages "noindex and nofollow." Its docs say "all Bluesky posts are therefore technically accessible to anyone," and that other apps decide for themselves whether to respect the setting (Bluesky visibility docs).

Does deleting a Bluesky post remove it from AI datasets?

Not from copies already made. Deleting a post sends a delete event down the firehose (Bluesky Jetstream docs), and Bluesky's terms say that when you delete your account it will "notify other services and Developer Applications on the AT Protocol" (Bluesky terms). The same paragraph says some posts "may continue to exist on these services that are outside our control," and "complete deletion across the network may not always be possible." A dataset downloaded and posted elsewhere doesn't hear the delete. If you're in the EU or UK, objecting to AI training under the GDPR is a route to take with whoever holds the copy.

How Morse keeps your unposted thoughts private

On Bluesky, anything you post goes into a public repository and out over a public stream, and no setting, pledge, or proposal can call it back from someone who has already saved it. So the writing you'd rather keep to yourself, the draft thread, the vent you decided not to send, or the question you didn't want your followers to see, belongs in a private chat.

On Morse, your chat history is stored in your browser, encrypted, and Morse's 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 (Morse privacy policy). Pick a privacy mode for each chat:

  • Private, for the post you're still deciding whether to write. Every new chat opens in Private. It runs open-weight models like DeepSeek, Kimi, and GLM, plus xAI's Grok, and Morse has zero data retention agreements 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.
  • E2EE, for words you'd never put in a DM. Your browser checks the enclave's hardware attestation, then encrypts your message so only that enclave can read it, and Morse's servers relay ciphertext they can't open. The guarantee depends on the enclave hardware and attestation working as designed. E2EE chats come with Pro, at $19 a month.
  • Anonymous, for asking GPT, Claude, or Gemini without a profile attached. Morse's proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider sees it without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy, so leave out names and details that identify you.

Your history locks when you sign out, and unlocking it takes two keys: one that never leaves your device, and one Morse's server gives you when you sign in. Clearing your browser or deleting your account erases it for good, and Morse can't bring it back, so keep a copy of anything you need. Because Morse doesn't store your chats, it usually has very little to give authorities. The Free plan is $0 with 15 messages a day, Pro is $19 a month, and Max is $49 a month.

What to do on Bluesky today

  • Posts: assume anything you post can end up in a dataset, since the firehose is public and User Intents hasn't shipped.
  • Deleting: delete what you regret soon, before more copies are made, and know that copies already saved elsewhere may stay.
  • Logged-out visitors: turn on the toggle to hide your posts from logged-out users. It asks search engines to skip your Bluesky pages, but it doesn't touch the firehose.
  • DMs: keep anything sensitive out of them. Bluesky says they're unencrypted.
  • Attie: if you join the beta, treat everything you type as kept indefinitely, with no way to delete it yourself.
  • Writing you'd rather not publish: draft it in a Morse chat instead of the composer.

Start a chat from the Morse home page without an account, or sign up free with no card.

Common questions

Does Bluesky sell your data?

Bluesky says it doesn't. Its privacy policy, last updated August 14, 2025, says "We do not sell Personal Data collected about you," and its terms say "We will never sell your Content without getting your permission first."

Can I keep my posts out of Bluesky's Discover feed?

Yes. Bluesky's developer docs describe a toggle that hides your posts from algorithmic recommendations, so they appear in Discover only to people who already follow you. It's published as a public record that other feeds may honor, but Bluesky leaves that to each feed's operator.

Does Bluesky use AI to check your age?

Where the law requires an age check, Bluesky's privacy policy lists facial age estimation, "a selfie or real-time facial scan used to estimate age via machine learning technology," among the methods its vendors may use. It says Bluesky does not keep "any biometric data or facial images" given to those vendors.

Will Bluesky get private posts?

The protocol is getting non-public data, but not encrypted data. Atproto Spaces opened as a developer alpha on August 20, 2026, and Bluesky's protocol team says spaces "give you access control not confidentiality."

Ask what you actually want to ask.

80+ models, characters, and images, with every chat encrypted and kept on your device. Free to start.

Start a private chat

Read next