What the notice settles and what it defers
The email OpenAI sent to users in the European Economic Area and Switzerland on the morning of August 15 draws the scope fairly sharply: ads start appearing this month on the Free and Go plans, Plus, Pro, Enterprise, Business and Education subscriptions stay outside, and users under 18, Temporary Chats and the ChatGPT Atlas browser see none. The company writes that ads will not be personalised at the start, that selection will rest on conversation context, general location and device type, and that past chats and memories are excluded; before personalisation begins, each user will be asked separately whether to opt in.[1]
The notice stops at one point. The company's privacy policy in force for Europe, dated June 4, 2026, contains no advertising disclosure at all, and the notice does no more than announce that an updated text describing how ads are selected, measured and controlled is coming; that text was not published alongside it. So what the phrase "conversation context" covers as an ad-selection input — the sentences you type, the topics inferred from them, how long that signal is kept — is not defined anywhere a reader can consult today. The subject of the consent arrives ahead of the text that describes it.[1]
The same gap on Twitch, from the other direction
Twitch has added a switch in account settings that stops a streamer's posts, streams and videos from being used to train Amazon's models. It was on by default, and the platform has not said when that use began; in the forum thread opened after the change, more than 16,000 creators objected. The terms of service in force since March 2024 grant the company and its sublicensees broad rights over user content while saying nothing about generative AI training, and the switch does not stop channel content being used for other purposes such as recommendations and AutoMod.[2]
What lines Twitch up with OpenAI is less the similarity of the products than the missing document the control rests on. On Twitch the switch sits over a backward gap: because nothing states what was used from when, there is no way to measure what turning it off takes back. At OpenAI the gap runs forward: the consent will be requested later, and the text defining what it covers does not exist yet. At Twitch and at OpenAI alike, the affected person holds no document with which to understand the claim, challenge it and seek a remedy. Bad faith is not the only reading; the Twitch statement and the OpenAI policy may appear in the coming weeks after an ordinary publication delay, and missing information stays missing information.[1], [2]
Which document closes the gap?
This time the document that would close it can be named exactly. If OpenAI publishes the updated European privacy policy before the ads start this month, the place to read is clear: whether it explicitly lists conversation content as an ad-selection input, and whether it states how long that signal is retained. A policy that states each of those turns the personalisation consent into a question that can be evaluated; a policy that only repeats that ads are chosen on conversation context preserves exactly the uncertainty the notice carries today.[1]
In this column on August 9 I looked at the advertising label on the pages Time opens to machines and wrote that the label sits on a version no reader ever opens. The problem there was that the label does not reach the reader; the problem here comes one step earlier, in the absence of the text that describes what the label means. The same line of thought at a different stage: whether a disclosure exists and whether it is usable are separate questions, and the second one always depends on a document someone can read.[3], [1]