Network captures, browser cache extraction, and OpenAI's own data export prove that special-category mental health data was fed into the RLHF training pipeline — without consent, without disclosure, and with the opt-out flag set to false after withdrawal.
Network captures, binary cache extractions, and OpenAI's own data export — three forensic layers pointing to one conclusion.
label="MICROSOFT/AZURE" and seven instances of "reviewed" extracted directly from Chrome's IndexedDB — the annotation identifiers used by Scale AI's backend./implicit_message_feedback fired while simply reading a therapy conversation — no thumbs-up clicked, no rating given.isOptedOut: false despite a stated withdrawal of consent. Training continued unlawfully after opt-out.model_slug: gpt-5-2 in OpenAI's own export — therapy conversations processed by an undisclosed model while the UI displayed gpt-4o.All evidence files are SHA-256 hashed and RFC 3161 timestamped via FreeTSA.org. HAR files open in any browser network analyser. conversations.json is OpenAI's own official export format. You don't need to trust this site — verify the files yourself.