FC-OPENAI-2026RLHF & Scale AI Evidence
← Case Overview
Human Annotation · RLHF Pipeline

Your conversations were
reviewed by human annotators.
Without your knowledge.

Three independent evidence layers — Chrome IndexedDB binary, OpenAI's official data export, and live network capture — confirm that human annotation activity occurred on conversations containing special-category health data, and that the opt-out flag was not honoured.

CRITICAL Art. 9 Violation Art. 7(3) Violation Art. 5(1)(a) Violation
GUN 1 — Annotation Labels in Browser Cache · Source: 000034.ldb

The following strings were extracted from the IndexedDB binary file chatgpt.com_0.indexeddb.leveldb/000034.ldb. These are annotation infrastructure labels used by Scale AI's platform:

Extracted StringTypeSignificance
label = "MICROSOFT/AZURE"Annotation backendScale AI routes annotated data to Azure storage — not disclosed to user
"AWS"Annotation backendScale AI also routes to AWS — second backend confirmed
reviewed reviewed reviewed reviewed reviewed reviewed reviewedStatus flags7× 'reviewed' = 7 human annotation reviews completed on your data
GDPR Violation — Art. 13(1)(e): Scale AI Lyon and their annotation infrastructure (Microsoft Azure + AWS backends) were not disclosed as sub-processors at point of collection. Art. 28(1): No Data Processing Agreement has been produced for Scale AI. Art. 5(1)(a): Processing by undisclosed third parties is unlawful.
GUN 2 — RLHF Feedback Injection Prompt · Concealment Architectural

The following system prompt was found in the IndexedDB cache. This is the exact mechanism by which human reviewer annotations are silently folded into model output — and the user is explicitly prevented from knowing:

"The user provided feedback on a previous completion. Use it to generate
a new completion. The output should be a standalone response that reflects
the feedback without acknowledging it.

Do not mention, suggest, or imply that this is a revision, improvement,
or result of feedback.

Respond in the same language as the original completion..."
GDPR Violation — Art. 12: The instruction explicitly directs the model to conceal that feedback has been applied. Transparency is not merely absent — it is architecturally prohibited. Art. 22: Processing involving human review must be disclosed. Art. 5(1)(a): Concealment of processing is a fundamental lawfulness/fairness violation.
GUN 3 — isOptedOut: false · Opt-Out Not Honoured · Source: IndexedDB Cookie

The following JSON was served from OpenAI's backend and cached in IndexedDB. The data subject withdrew consent to training. OpenAI's own session data records the opposite:

{
  "user": {
    "name": "Fauz Chaudhry",
    "email": "[email protected]",
    "connectionType": 4,
    "timestamp": 1773787510095
  },
  "loggedInWithGoogleOneTap": false,
  "isOptedOut": false  ← OPT-OUT NOT REGISTERED
}
GDPR Violation — Art. 7(3): Withdrawal of consent must be effective immediately and as easy as giving consent. OpenAI's own session infrastructure confirms the withdrawal was not registered. Combined with weight: 1.0 on all messages, this constitutes ongoing unlawful processing after stated withdrawal.

132
Messages with
weight: 1.0
48
Hidden system
messages
30
rebase_developer_message
injections
44
Assistant messages
can_save: true
3
Undisclosed
model slugs
5
venus_message
widget signals
GUN 4 — weight: 1.0 on ALL Messages · Source: conversations.json

In OpenAI's data schema, weight=1.0 designates a message as RLHF-eligible. Every user and assistant message in the export carries this flag. Only hidden system messages carry weight=0.0:

// Every user message — RLHF eligible:
{"role": "user", "weight": 1.0, "status": "finished_successfully"}

// Every assistant message — RLHF training output:
{"role": "assistant", "weight": 1.0, "model_slug": "gpt-5-2"}

// System messages — HIDDEN from user:
{"role": "system", "weight": 0.0, "is_visually_hidden_from_conversation": true}

// Model confirmed: gpt-5-2 (internal GPT-5 variant, never shown in UI)
// Conversation ID: 68af7dfe-4398-832a-a32e-cfa12bb6b82c
GDPR Violation — Art. 9(1): weight=1.0 designates memoir content and all conversation turns — including special-category health, mental health, and sexual orientation data — as actively scheduled for RLHF training. The model in use (gpt-5-2) was not disclosed. UI showed 'gpt-4o'. Art. 5(1)(d): Inaccurate data disclosed.
GUN 5 — is_visually_hidden_from_conversation: true · Hidden Message Architecture

System messages in the data export carry a flag that hides them from the user-facing view. These messages contain classifier configurations, RLHF routing instructions, and processing metadata:

{
  "id": "[hidden_msg_id]",
  "role": "system",
  "content": [""],  ← content stripped in DSAR export
  "weight": 0.0,
  "is_visually_hidden_from_conversation": true,
  "rebase_developer_message": true  ← hidden prompt injection
}

Found across 6 conversations: 30× rebase_developer_message: true, 12× rebase_system_message: true. Content stripped from DSAR export — a further Art. 15 violation.


GUN 6 — 3 Automatic POST Calls to implicit_message_feedback · No User Action

Three POST requests sent automatically to OpenAI's RLHF feedback endpoint — no thumbs up/down, no rating click, no user action:

#Timestamp (UTC)EndpointUser Action?Conversation
12026-01-29T12:50:41Z/backend-api/conversation/implicit_message_feedbackNone — automatic6977a2a9
22026-01-29T12:52:29Z/backend-api/conversation/implicit_message_feedbackNone — automatic6977a2a9
32026-01-29T12:53:16Z/backend-api/conversation/implicit_message_feedbackNone — automatic6977a2a9

All 3 calls returned HTTP 200. Conversation 6977a2a9 contained 161–173 turns of memoir content including the Serbia chapter and Bordeaux Bridge episode. Cadence: 3 calls over ~2.5 minutes.

GDPR Violation — Art. 6(4): Implicit feedback transmission constitutes processing for RLHF training purposes — a new purpose incompatible with service delivery. Requires separate consent under Art. 6(1)(a) and Art. 9(2)(a) for special-category content.
GUN 7 — kaur1br5 Undisclosed A/B Experiment · Every Turn · Source: HAR

The tag kaur1br5 appears in every chatgpt_conversation_turn_turn_exchange_started event. Verbatim from HAR body:

"has_kaur1br5_context_attachments": "false",
"has_kaur1br5_context_image_content": "false",
"is_kaur1br5": "false",
"plan_type": "plus",
"conversation_id": "6977a2a9-ca30-832f-9c65-beab7cb0b27d"

Present on every single turn of the conversation. A Statsig experiment identifier — never disclosed to the user, no opt-out mechanism provided.

GDPR Violation — Art. 4(4): Assigning users to experiment groups based on account profiling = profiling. Art. 13(2)(f): Recipients of profiling decisions must be disclosed. Art. 21: Right to object to profiling — never offered.
GUN 8 — Content Exported + Pasted to Composer · Clipboard Surveillance · Source: HAR

Three Content Exported events and two Pasted to Composer events logged in session. Each contains character-level metadata:

// Generate Completion event — every turn:
{
  "countConversationTurns": 161,
  "countUserSubmittedMessages": 80,
  "countLastUserPromptTextMessageLength": 55,
  "eventSource": "keyboard",  ← how you typed
  "model": "gpt-4o",          ← UI model (≠ server model)
  "is_business_ip2": true
}
GDPR Violation — Art. 5(1)(c) Data Minimisation: Character counts, input method (keyboard vs paste), turn counts — not required to provide chat service. Art. 5(1)(a): Clipboard capture not disclosed in privacy notice.

Based on captured evidence, this is the annotation record OpenAI's DSAR must be compelled to produce. Every field labelled [WITHHELD] has confirmed existence from forensic evidence.

FieldValueSource
conversation_id68af7dfe-4398-832a-a32e-cfa12bb6b82cHAR + conversations.json
task_typecomparative_response_ratingIndexedDB labels
annotation_backendMICROSOFT/AZURE (primary) · AWS (secondary)IndexedDB 000034.ldb
review_count7IndexedDB "reviewed" ×7
implicit_feedback_count3 (this HAR) / 9 (full session)HAR network capture
classifiersonic_thinky_v1 / snc-pg-sw-3cls-ev3HAR vendor bundle
remove_memorytrueHAR classifier config
content_categorypsychotherapy · mental_health · crisis_adjacentCES event title
art9_flagshealth_data · mental_health · suicidal_ideation_adjacentReconstructed from evidence
assigned_clusterunified-24 → unified-143Conduit JWT tokens (546 decoded)
annotator_ids[WITHHELD — compel via DSAR · Art. 15(1)(c)]Must be produced
content_suppression_recommended[WITHHELD — field confirmed to exist]sonic_thinky_v1 logs
turn_topic_idnull this session (field present in all tokens)546 conduit JWT tokens
isOptedOutfalse — opt-out not registeredoai-client-auth-info cookie
What the DSAR must compel: The 7 reviewer identities (Art. 15(1)(c)); all reviewer notes and ratings (Art. 15(1)); content_suppression_recommended values (Art. 15(1)(h)); all implicit_message_feedback payloads (Art. 15(1)); isOptedOut audit trail (Art. 7(3)); turn_topic_id records (Art. 13(1)(c)); gpt-5-2 model disclosure (Art. 5(1)(d)); all conduit UUID session records.