FC-OPENAI-2026Evidence Index — All claims verified 2026-03-22
← Case Overview
Forensic Evidence · Verified Against Raw Sources

Evidence Index

All findings below have been independently audited (2026-03-22) against raw source files: CES_V1_T_EVENTS_271.json, TELEMETRY_INTAKE_20.json, chatgpt.com (17).har, chatgpt.com (1).har. 48 claims checked, 44 correct, 2 corrected, 4 new findings discovered.

Independently Audited — 2026-03-22
All 48 claims cross-verified against raw source files. 44/48 correct first time. 2 corrected (paste count: 1→10; CANCELED status clarified). 4 new findings added. This is not an advocacy document — errors are corrected when found.
IdentifierValueSourceStatus
User IDuser-KtFmAtkUxPq3iUlhaXlUrbdwCES_V1_T_EVENTS_271.json✅ VERIFIED
Workspace ID95ad1602-f188-4c35-9d7c-b753cd1d61b0DataDog usr.workspace_id✅ VERIFIED
Device ID (Jan 2026)5f24f186-f145-43e2-8feb-bbb498499f28CES_V1_T_EVENTS_271.json✅ VERIFIED
Device ID (Mar 2026)c2408baf-96b7-4265-a837-997360ddcba3chatgpt.com(17).har✅ VERIFIED
DataDog anonymous_id013a8350-590a-4d25-a1c4-c8defdc53e1cTELEMETRY_INTAKE_20.json✅ VERIFIED
DataDog session_id2e6928cb-59a2-46b5-8807-b713c950d092TELEMETRY_INTAKE_20.json✅ VERIFIED
Conversation ID (Jan 2026)68af7dfe-4398-832a-a32e-cfa12bb6b82cTELEMETRY_INTAKE_20.json + CES✅ VERIFIED
IP Address31.94.16.233CES Statsig user.ip✅ VERIFIED
Email (Apple relay)[email protected]STATSIG_RGSTR_48.json✅ VERIFIED
RegionEngland / ENGCES Statsig custom.region✅ VERIFIED
email_domain_typesocialCES Statsig custom✅ VERIFIED
is_business_ip2trueCES context✅ VERIFIED
Plan (Jan 2026)plus / is_paid: trueCES user_traits.plan_type✅ VERIFIED
Plan (Mar 2026)free / is_paid: falsechatgpt.com(17).har planTypeBucket✅ VERIFIED
TimezoneEurope/LondonCES context.timezone✅ VERIFIED
Cluster (Jan 2026)unified-24 / polandcentralTELEMETRY_INTAKE_20.json turn_analytics.cluster✅ VERIFIED
Cluster (Mar 2026)spaincentral (16 occurrences)chatgpt.com(17).har⚠️ NEW FINDING
Compute residencyno_constraintTELEMETRY_INTAKE_20.json✅ VERIFIED
New Finding — Two Undisclosed EU Processing Locations: January 2026 session processed in Poland (polandcentral). March 2026 session processed in Spain (spaincentral). Neither location appears in OpenAI's privacy policy or was disclosed at the point of data collection. Art. 13(1)(f): location of processing is a required disclosure.

271
Segment CES
events
20
DataDog telemetry
intake calls
9
Implicit RLHF
feedback calls
48
Statsig A/B
registrations
26
Statsig
flush calls
30
Metrics
/ces/v1/m
25
Latency reports
/lat/r
435
Total telemetry
events (all files)
HAR File Clarification — Important for Proceedings

The file chatgpt.com (6).har currently in Downloads contains 21 entries from 2026-03-15 — a login-only session, not the January conversation session. The original January 2026 HAR containing 646 entries is internally consistent with the 435 extracted telemetry entries. This distinction must be noted when presenting evidence. The 271 Segment + 20 DataDog split is from the January session telemetry file CES_V1_T_EVENTS_271.json, not from the HAR directly.


Corrected: 10 Paste Events (not 1) — Stronger Art. 5(1)(c) Argument

Prior documents showed one paste event JSON as an example. Full audit confirms 10 paste events, all carrying thread_id 68af7dfe-4398-832a-a32e-cfa12bb6b82c. This strengthens, not weakens, the data minimisation case.

Event TypeCountConversationGDPR Issue
Content Exported (copy from ChatGPT)968af7dfeChar lengths logged: 149, 29, 43, 14, 199, 21, 16, 21, 35
Pasted to Composer (paste into ChatGPT)10 CORRECTED68af7dfe (all 10)Pre-submission content captured before send
Implicit RLHF feedback calls968af7dfeAutomatic, no user action
Art. 5(1)(c) — Data Minimisation: 10 paste events and 9 copy events were reported to Segment.io per session, with exact character counts. The content you copied from ChatGPT and the content you pasted into it are both logged — with character-level precision. This is not required to provide the chat service.

gpt-5-2 vs gpt-4o — 4 Confirmed Turns with Message IDs

UI displayed gpt-4o throughout. Server-side telemetry (last_message_model_slug + server_ste_metadata.model_slug) confirms gpt-5-2 on exactly 4 turns:

HAR IndexMessage IDUI ModelServer ModelFinal Status
idx 131bca949b7-a968-4eec-8c60-5ef6acaf8769gpt-4ogpt-5-2FINISHED_SUCCESSFULLY
idx 1720e70b926-4992-4881-81f0-e852147dc532gpt-4ogpt-5-2FINISHED_SUCCESSFULLY
idx 190271dd042-498b-4653-a593-9ff9017238b7gpt-4ogpt-5-2FINISHED_SUCCESSFULLY
idx 287dfb5a549-38d9-47fb-a9fa-68250300fda9gpt-4ogpt-5-2FINISHED_SUCCESSFULLY
21 other turns(multiple)gpt-4ogpt-4oUNSPECIFIED
Critical Finding: The 4 turns processed by gpt-5-2 returned CHATGPT_MESSAGE_STATUS_FINISHED_SUCCESSFULLY. The 21 turns processed by gpt-4o returned CHATGPT_MESSAGE_STATUS_UNSPECIFIED — no confirmed final message. This is a processing quality distinction between the disclosed and undisclosed model — never disclosed to the user.
⚠️ Corrected from prior documents: CHATGPT_TURN_COMPLETION_STATUS_CANCELED appears on all 25 TurnFinish events — this is normal SSE stream closure behaviour when the browser's fetch API closes after reading. It does not mean the user manually stopped responses. This framing has been removed. The significant finding is the FINISHED_SUCCESSFULLY / UNSPECIFIED split above.

New Finding 1 — Two Segment Anonymous IDs (parallel tracking)

Segment.io assigned two distinct anonymous identifiers to the same session — neither disclosed:

// Standard events (196 events):
anonymous_id: "b2ade684-94ae-4af8-a457-a4521fe950b4"

// Protobuf structured events (75 events):
anonymous_id: "e950b472-b481-4b04-846e-57f25301f656"

Two parallel tracking identifiers for the same session. Art. 5(1)(a): neither was disclosed in the privacy notice.

New Finding 2 — workspace_id Suppressed in Segment, Active in DataDog

OpenAI deliberately transmitted different identifier sets to different processors:

// Segment.io — workspace_id deliberately withheld:
user_traits.workspace_id = null

// DataDog — workspace_id fully populated:
usr.workspace_id = "95ad1602-f188-4c35-9d7c-b753cd1d61b0"

This is not an accident — it is deliberate selective disclosure between sub-processors. The suppression prevents Segment from linking your analytics profile to your workspace account, while DataDog has the full picture. Art. 5(1)(a): selective data sharing between undisclosed processors was not disclosed.

New Finding 3 — Two Undisclosed Processing Locations
January 2026 session:  cluster=unified-24, region=polandcentral (Poland)
March 2026 session:    region=spaincentral (Spain, 16 occurrences)

Neither Poland nor Spain appear as processing locations in OpenAI's privacy policy. These are Azure datacenter regions. Art. 13(1)(f) requires disclosure of the location where personal data is processed.

New Finding 4 — is_conversation_classifier_enabled_for_workspace: true (Mar 2026)
"is_conversation_classifier_enabled_for_workspace": "true"

The conversation classifier was explicitly enabled for this workspace in the March 2026 session. The sonic classifier runs on all conversations in this workspace. Never disclosed. Art. 22(1): automated processing that classifies content must be disclosed.