When Facebook decides what counts as your data

On 15 June 2026, with the news cycle already full of louder stories, Meta confirmed that it is rolling a new "AI Mode" into Facebook. The feature, the company said, is built to pull from public information across its platforms — the posts, comments, photos and engagement signals that billions of users have generated over more than a decade. The framing inside the announcement was calm: a smarter search box, a more conversational Facebook, a tool that knows the difference between asking about a recipe and asking about a friend's holiday. The substance beneath the framing is louder. A platform that has spent years building the world's most exhaustive archive of human expression is now reorganising that archive as raw material for its own machine intelligence.

That is the beat the company did not hit. Every time a user typed a status, left a comment, uploaded a photo set to public, or ran a Page, they were contributing to a corpus. For most of Facebook's life, that corpus was treated, in the language of privacy notices and product copy, as content addressed to other humans. AI Mode reclassifies it: the same public post is now also training data, retrieval context, a citation in an answer a chatbot might give to a stranger. The user did not change their behaviour. The contract did.

The quiet redefinition of "public"

The announcement, as reported by TechCrunch on 15 June, emphasises that AI Mode draws on content that is already public on Meta's platforms. That is technically true and rhetorically evasive. A great deal of the material on Facebook was made public under a different theory of what "public" meant — a theory in which public meant visible to other people, not legible to a model, not retrievable in a chat window, not recombined into answers the original author never imagined and cannot audit. The Indian Express's coverage of the same rollout, published the same week, makes the same point in plainer language: the public surface of the platform is now feedstock for a product Meta sells.

This is not a novel problem. Search engines have indexed public web pages for decades; the difference is the depth of inference. A search index returns a list of links. An AI assistant returns a synthesised answer, fluent enough that users stop clicking through, confident enough that the source disappears into the response. The author of the underlying post is not named in the synthesis. The platform has quietly converted user behaviour into a predictive inventory it can deploy against any query.

Consent after the fact

The structural concern is consent, but consent stripped of the moment it would have to be given. The cleanest version of the critique is procedural: at what point was the user asked whether their public posts could be used to train, retrieve against, and answer questions inside a generative product? The practical answer, on the evidence available, is that the question was folded into a longer terms-of-service update and a settings panel most users will never open. The Indian Express's reporting notes the breadth of the rollout; the technical depth of what "public content across the platform" entails in practice is left for the user to reconstruct.

Compare this to the more careful framing now standard in jurisdictions with functioning data-protection authorities. The European Union's GDPR, the United Kingdom's data regime, and the patchwork of state-level laws in the United States have all spent the last decade sharpening the line between data a user shares with another person and data a user shares with a service provider. The line on Facebook has just been redrawn, in the opposite direction, by the platform itself. The user, again, is the last to know.

A counter-narrative worth taking seriously

The defence is real and should be stated. Generative products do need access to large, diverse corpora to be useful in languages and dialects that the open web underserves. Opt-in-only corpora skew English, skew professional, skew recent; they under-represent the kinds of conversational, vernacular, and locally-specific posts that have always been Facebook's strongest material. There is a genuine argument that excluding this corpus from AI training would entrench the biases that already plague English-dominant models. Meta is not wrong to point out that public data is the substrate on which any honest answer to a global user base has to be built.

That argument, however, is an argument about how to use the corpus, not about who decides. The decision to repurpose more than a decade of public content as AI input was made by a single company, in a single announcement, in a week when most of the political and regulatory class was looking elsewhere. The Indian Express and TechCrunch both reported the rollout; neither described a meaningful consultation with the users whose material is now in scope. A defence grounded in corpus quality does not answer the governance question. It just relocates it.

What is at stake

If AI Mode succeeds at the scale Meta clearly expects, the precedent is set. Other platforms — TikTok, X, YouTube, LinkedIn, Snapchat — are watching, and several are already building parallel products. Each new entrant will face the same temptation and the same deficit of user-side process. The public web, and the public parts of private platforms, will be absorbed into proprietary model weights, retrievable at inference time, attributable to no one, monetisable by the platform. The user becomes, structurally, an unpaid contributor to a product the user does not own and cannot audit.

The counter-case is that users can leave. In theory, a user unhappy with the new terms can set their account to private, delete old posts, or close the account entirely. In practice, the cost of exit is enormous: years of friendships, messages, group memberships, marketplace listings, local-business Pages, and accumulated reputation that exist only inside the platform. "You can leave" is a real choice for some users and a fiction for most.

The honest reading is that Meta has, with one announcement, retired a long-standing implicit distinction between human-readable and machine-readable public content. The category of "public on Facebook" now means, in operational terms, "available to Meta's AI." That is a change worth more than the announcement received. It is also a change that, in the absence of regulatory pushback or organised user response, will be quietly inherited by every platform that watches Facebook move first.

This publication finds that the right response is not panic but process: clear disclosure of which content is in scope, an explicit opt-out that does not require deleting the account, and an audit trail for any post or comment that is surfaced in an AI response. None of that requires new technology. It requires a decision Meta has not, on the evidence, made.