OpenAI Used My Therapy Conversations to Train AI Without Consent

OpenAI used my therapy
conversations to train AI.
The network traffic proves it.

This is the forensic case file for Fauzia Chaudhry v OpenAI Inc. Network captures, browser cache extractions, and OpenAI's own data export confirm that special-category health data was processed by Scale AI annotators for RLHF training — without explicit consent, without disclosure, and with the opt-out flag set to false after stated withdrawal.

The Evidence — 10 Independent Findings

Scale AI Annotation Labels in Browser Cache

Strings label = "MICROSOFT/AZURE", "AWS", and "reviewed" ×7 extracted from Chrome IndexedDB binary 000034.ldb. Seven human annotation reviews confirmed.

CRITICAL · Art. 13(1)(e)

RLHF Feedback Injection Prompt — Concealment by Design

System prompt cached in IndexedDB: "Do not mention, suggest, or imply that this is a revision, improvement, or result of feedback." Transparency architecturally prohibited.

CRITICAL · Art. 5(1)(a)

isOptedOut: false — Opt-Out Not Honoured

OpenAI's own session cookie oai-client-auth-info records "isOptedOut": false after stated withdrawal of consent. All messages carry weight: 1.0 RLHF eligibility.

CRITICAL · Art. 7(3)

weight: 1.0 on All Messages — RLHF Pipeline Active

Every user and assistant message in the official data export carries weight: 1.0. Only hidden system messages carry weight: 0.0. 132 RLHF-eligible messages confirmed including memoir content.

CRITICAL · Art. 9(1)

3 Implicit Feedback Calls — No User Action

Three POST requests to /backend-api/conversation/implicit_message_feedback sent automatically during the captured session — no thumbs up, no rating click, no user action whatsoever.

CRITICAL · Art. 6(4)

kaur1br5 — Undisclosed A/B Experiment on Every Turn

The field is_kaur1br5: false appears in every turn_exchange_started event in the HAR. An undisclosed Statsig experiment assigned to control group — never disclosed, no opt-out provided.

HIGH · Art. 22 + Art. 13(2)(f)

remove_memory: true — Paid Feature Secretly Disabled

Sonic classifier config extracted from vendor bundle: "remove_memory": true. The classifier strips memory context before running — disabling a paid Plus feature without disclosure.

HIGH · Art. 22(1)

gpt-5-2 Server vs gpt-4o UI — Model Deception

The UI displayed gpt-4o. The official conversations.json export confirms model_slug: gpt-5-2. An undisclosed internal GPT-5 variant processed the psychotherapy and memoir conversations.

CRITICAL · Art. 5(1)(d)

302 Article 9 Events — Health Data in RLHF Pipeline

302 special-category data events across all HAR files. Conversation "Psychotherapy mode questions" (ID: 68af7dfe) triggered the RLHF rating event. Mental health content confirmed in active annotation pipeline.

CRITICAL · Art. 9(2)(a)

snc-pg-sw-3cls-ev3 — Potential Joint Controller

Classifier snc-pg-sw-3cls-ev3: prefix "snc" consistent with Anthropic model naming. If Anthropic co-determined purposes of this classifier, they are a joint controller under Art. 26 GDPR. No joint controllership agreement published.

HIGH · Art. 26

OpenAI used my therapyconversations to train AI.The network traffic proves it.

OpenAI used my therapy
conversations to train AI.
The network traffic proves it.