Most of the services you use every day now train AI models on what you post, write, upload or say to them. A few give you a switch. This is where each one is, what it does, and — because the switch is only half the story — what in the terms of service lets them do it in the first place.
The legal spine is the same everywhere. When you accepted the terms, you granted the company a licence to your content: in Instagram’s case a “non-exclusive, royalty-free, transferable, sub-licensable, worldwide licence to host, use, distribute, modify, run, copy, publicly perform or display, translate, and create derivative works” of it. Training a model is a use of that licence. The privacy policy then decides what personal data can go in, and that is the layer the opt-outs live in. You cannot opt out of AI training by editing the content licence; you can only use the controls the privacy policy gives you. So here they are.
Meta: Facebook and Instagram
What they train on. Meta says it uses public content from adults — posts, photos, captions, comments — and your conversations with Meta AI. It says it does not use private messages.
Where the switch is. There isn’t one in the United States. If you are in the EU, UK or Switzerland, there is an objection form: open Settings, go to Privacy Center, and search for “Generative AI” or “object.” Submit the form and Meta says it will honour the objection going forward. Meta describes this in its Privacy Policy.
What you can do in the U.S. Make future content non-public. Setting your Instagram account to private, or setting a Facebook post’s audience to Friends, keeps that content out of the public set Meta says it trains on. It does nothing for what was public before, and it is not an opt-out — it is a side effect of the audience setting. Instagram’s terms are graded D on our read partly because of how little the document leaves in your hands.
What they train on. Your profile, posts, articles and other content on the platform, for generative-AI features and, since November 2025, for training by Microsoft affiliates as well. LinkedIn expanded this to EU, EEA, UK, Swiss, Canadian and Hong Kong members on November 3, 2025; U.S. members were included from September 2024.
Where the switch is. Settings & Privacy → Data privacy → Data for Generative AI Improvement → off. It is on by default. Turning it off stops future use; LinkedIn says it does not affect training that already happened.
Adobe: Photoshop, Lightroom, Acrobat and the rest
What they train on. Adobe draws a line that caused a public fight in June 2024: it says it does not train its Firefly generative models on customer content stored in Creative Cloud, but it does run “content analysis” on your files to improve features, unless you turn it off. Adobe explains the distinction in its content analysis FAQ, and our read of Adobe’s terms shows the clauses that make it possible.
Where the switch is. Sign in at account.adobe.com → Privacy and personal data → Content analysis → off. If you have a business or enterprise licence the setting is controlled by your admin and, Adobe says, is off by default for those accounts.
X (Twitter)
What they train on. Your public posts and your interactions with Grok — the questions you ask it, and its answers. X’s November 15, 2024 terms update made your content available for training its models, and a later update extended that to “third-party collaborators.”
Where the switch is. Settings and privacy → Privacy and safety → Grok & Third-party Collaborators → untick the box that allows your posts and Grok interactions to be used for training. There are two consents on that screen — one for X’s own models and one for third parties — and you have to untick both.
ChatGPT
What they train on. Your conversations, unless you say otherwise. Free, Plus and Pro accounts are opted in by default; Team, Enterprise and API use are excluded by default.
Where the switch is. Settings → Data controls → Improve the model for everyone → off. A Temporary Chat is also excluded, and OpenAI keeps it for up to 30 days for abuse review rather than adding it to history. OpenAI lays out the controls in its data controls FAQ.
The ones without a switch
Some services give you no toggle at all, and it is worth knowing which. Slack trains on customer data unless a workspace owner emails feedback@slack.com asking to opt the whole workspace out; an individual member cannot do it alone. Reddit offers no opt-out: it has licensed its archive to Google and OpenAI outright, and the content licence in its terms is what made those deals possible. Google Gemini does not have a training toggle as such; the control is Gemini Apps Activity, which you can turn off so conversations are not kept or reviewed. In each case the answer lives in the terms, not the settings, which is why we read the terms rather than the menus.
Three things every switch has in common
They only look forward. No opt-out we have seen retrains a model to forget you. Deletion is the only lever on the past, and every service keeps deleted content for a window first — Instagram’s data-retention clause says up to 90 days.
They are on by default. Each of these is an opt-out, not an opt-in. The company is counting on the setting staying where it shipped.
They sit under a licence you cannot edit. The content-licence clause is in nearly every set of terms we have read, and it is the reason a privacy toggle is the only control you have. If a company you rely on has no toggle, the clause is the whole answer, and the decision is whether to keep posting there. The full set of reads shows which companies have the clause and what it says, quoted from the document.
Menus move. These paths were checked in September 2026; when one has shifted, search the service’s help centre for the setting name, which changes far less often than its location.