Social media AI learns from your posts—here’s how to limit it

Social media platforms routinely use user content to train artificial intelligence systems, often without clear permission. Most companies assume that logging in grants broad consent, and opting out requires handling complex settings or lengthy privacy documents.

How platforms use your data

Amazon’s Twitch recently introduced a toggle allowing users to exclude their streams, videos, and chat logs from AI training. The setting is active by default, and the company has not disclosed which AI applications will use the collected material. Other platforms provide even less clarity about their practices.

Meta, parent company of Facebook, Instagram, and Threads, collects an extensive dataset. Its policy includes posts, images, user interactions, third-party data brokers, and publicly available online information. Private messages remain off-limits unless shared directly with Meta’s AI features. In the European Union, users can refuse participation, but elsewhere, the only option is filing a complaint if personal data surfaces in AI-generated responses.

Reddit does not develop its own AI models but has licensing agreements with OpenAI and Google for public posts. Even if these deals expire, AI companies have already harvested large volumes of Reddit content. The platform offers no way to opt out of this data sharing.

Where opting out is possible—and where it isn’t

Bluesky, built on the decentralized AT Protocol, does not use user posts for AI training. However, public data remains accessible to external scrapers, making opt-out mechanisms ineffective. X, formerly Twitter, allows users to block training for Grok and xAI by adjusting privacy settings, though the process involves multiple steps and is easy to miss.

YouTube, part of Google’s AI operations, takes an aggressive approach. Google has trained video-generation models using uploaded content without creator permission. While users can prevent third-party training, there is no option to exclude their data from Google’s own models. Standard users must modify settings at the Google account level, where options are scattered and often ineffective.

Most platforms default to including user data in AI training. Opt-out features, when available, are either legally required (as in the EU) or hidden behind multiple layers of menus. Deleting an account does not ensure removal from training datasets, as companies frequently retain data for extended periods, and third-party scrapers operate independently.

Twitch’s opt-out toggle is one of the few direct mechanisms available. However, it only applies to a user’s own channel. If someone appears in another stream or chat, the host’s settings determine whether that person’s data is used.

The core problem lies in how consent is obtained. Social media companies interpret passive use as permission, and few explain how data is repurposed. Privacy policies, often lengthy and dense, bury AI training clauses under vague terms like “improving services” or “personalization.”

Until policies change, users should assume any public post will be ingested by AI models. Opt-out options, where they exist, offer only temporary protection, as companies can alter policies at any time and third-party scrapers operate beyond platform control.

For those looking to minimize exposure, reviewing privacy settings across platforms is a practical step, though not a complete solution.