What operators reported / what the docs say
These notes are a privacy-boundary checklist drawn from OpenAI primaries, for teams reviewing ChatGPT ads and Sponsored Agents. They are not an in-house lab run.
Boundary map
| Data / control | Doc position |
|---|---|
| ChatGPT answers influenced by ads? | No. Separate systems; advertisers cannot shape responses |
| Chats, history, memories to advertisers? | No by default; never sold |
| What advertisers get | Aggregated non-identifying performance (views, clicks); more measurement may come later |
| User messages an advertiser via an ad | Advertiser sees those direct messages only |
| Who sees ads | Free and Go; not Plus/Pro/Business/Enterprise/Edu; not under-18 accounts as identified |
| Placement | Below end of a response; labeled sponsored; visually separated |
| Sensitive contexts | Ads not eligible near sensitive/regulated topics (health, mental health, politics per Help Center) |
| User controls | Settings > Ad Controls: turn off personalization, dismiss/feedback, why-this-ad, clear ads data |
| Sponsored Agents | Optional labeled conversation with a business-sponsored agent after an ad click; distinct from ChatGPT answers and from the originating thread (launch post) |
Review checklist
- Update DPIA / vendor review language: three objects (main chat, ad unit, Sponsored Agent thread).
- Confirm plan tier expectations for staff using Free/Go versus paid ad-free tiers.
- If enabling Sponsored Agents creatives: document who owns agent knowledge, Actions, and escalation when the agent is wrong.
- Re-read Help Center when Action handoffs to third-party apps are used; public Sponsored Agent-specific Action confirmation rules looked thinner than ordinary-ad controls in the materials we retrieved. Mark gaps UNKNOWN and ask OpenAI or counsel before enterprise rollout.
Failure modes to watch
- Users assume a Sponsored Agent is ChatGPT. Labeling is the control; train support staff on the difference.
- "Too relevant" ads create perceived privacy breach even when policy says only in-product signals were used. Keep the personalization toggle path in help macros.
- Aggregated reporting expanding over time. Re-check the Help Center measurement section on a schedule.