OpenAI is making two changes to ChatGPT: Visual ads will appear during image generation in the United States, while text produced in the European Union (EU) will receive invisible watermarks. For you, these are very different interventions: The advertising is easy to see, but the text marker requires a suitable detection tool. Both affect how transparently ChatGPT handles commercial interests and machine-generated content.
What you can expect
OpenAI plans to test visual ads in the US later in October 2026, placing them alongside image-generation results. According to The Verge’s lead report, they will feature images of sponsored products or services. The ads are supposed to be labeled and kept separate from the generated image; OpenAI says they will not influence ChatGPT’s answers.
An everyday example makes the distinction clearer: You ask ChatGPT to create an image for a birthday invitation and see a commercial product advertised during or next to the result. That product should not become part of your generated image. Its visual proximity could nevertheless make the placement feel like part of the creative process or like a recommendation from ChatGPT.
The image ads will begin with a limited group of advertisers and only in the US. ChatGPT Plus, Pro, and Enterprise subscriptions will remain ad-free according to the available information. OpenAI first introduced ads in February, initially displaying a company name, logo, and product link in a sponsored area beneath the chat; TechCrunch says the advertising was later expanded to India.
At the same time, OpenAI is rolling out an invisible text watermark to eligible ChatGPT and Codex users on all plans in the EU. Codex is an AI tool focused on programming tasks. The rollout will take place over several weeks and is intended to address transparency rules under the European Union’s Artificial Intelligence Act, commonly called the EU AI Act.
What invisible watermarks do
The watermark, called textGrain, is not a symbol or visible footnote. It makes small, statistically recognizable adjustments to the language model’s word choices. A purpose-built detector can look for the resulting pattern even though you can read and copy the text normally.
Because the marker is embedded in word choice, it generally travels with text when you copy and paste it. In its overview of the EU text provenance rules, OpenAI describes where watermarking applies, how detection works, and why detector access will begin with researchers and expert organizations. The supplied sources do not announce a public tool that would allow anyone to reliably check any piece of writing.
OpenAI says the watermark does not identify an individual user and caused no meaningful decline in model performance during internal tests. These are provider claims and have not been independently verified in the supplied reports. OpenAI also says textGrain matched or exceeded other approaches it tested, including Google’s SynthID for text; according to THE DECODER, the company plans to make the technology open source.
Detection is still fragile. In one test described by OpenAI, the detection rate fell from about 92 percent to 66 percent after ten percent of the words were replaced with synonyms. Short passages, mathematical answers, and translated text are also more difficult to detect. THE DECODER reports roughly 95 percent detection for 400-token passages on psychology topics at a target false-positive rate of one percent, compared with around 80 percent for 200 tokens; tokens are small units into which a language model divides text.
The watermark is therefore a provenance signal rather than proof. It does not confirm that a text is accurate or establish ownership, the extent of human contribution, or purely human authorship. Conversely, the absence of a marker does not prove that a person wrote the text without AI.
Pros and Cons of ads and watermarks
Pros:
- Visible separation – The new ads are meant to be labeled and displayed outside the image ChatGPT generates.
- Greater transparency – The watermark can give qualified organizations an additional signal that text was generated by a machine.
- No user identifier – OpenAI says textGrain does not contain information that identifies an individual person.
- Ad-free subscriptions – Plus, Pro, and Enterprise accounts are expected to remain exempt from visual ads.
Cons:
- Commercial proximity – Even with a label, an ad next to an image result can feel like a recommendation from the system.
- Unreliable detection – Editing, translation, short passages, and certain types of content weaken the statistical signal.
- Limited verification – The detector will not be generally accessible at launch, so you cannot reliably inspect the signal yourself.
- Regional inconsistency – Ads and watermarks are launching in different regions, making expectations for the same product less clear.
What this means for trust and privacy
For ads, OpenAI promises safeguards intended to keep placements out of emotionally vulnerable, sensitive, or otherwise unsuitable conversations. Selected advertisers can also set exclusion terms, according to THE DECODER’s report. DoubleVerify and Integral Ad Science are expected to assess brand suitability without receiving access to real user conversations.
OpenAI is also expanding how advertisers can measure results. The company’s announcement describes new measurement tools, attribution partnerships that connect actions such as clicks and purchases, and systems for evaluating whether an advertising environment is suitable. TechCrunch lists AppsFlyer, Adjust, and Branch among the attribution providers, as well as Haus, Measured, and WorkMagic for geographic experiments.
This measurement infrastructure makes ChatGPT more attractive to advertisers, but it also raises questions about user data. The sources do not explain in detail which personal data from an individual ChatGPT session enters every measurement process. The reports therefore support neither a promise of complete anonymity nor a claim that individual users are necessarily being tracked.
OpenAI says advertising will not influence answers. That is a central commitment, but the supplied material does not provide an independent audit of actual placements. Trust will therefore depend not only on technical separation, but also on whether labels remain clear and sensitive conversations genuinely stay free of ads.
The privacy issue is different for watermarks. The signal is not supposed to contain a name or account identifier, but it may classify a text as likely generated by a particular system. In education and the workplace, that finding could easily be overinterpreted even though OpenAI itself warns that detection is not reliable in every case.
What this means for you
If you are a beginner, the most useful first step is simple separation: Treat a visible placement as advertising, not as part of ChatGPT’s answer. If you use ChatGPT text in a job application, report, or school assignment, you should also avoid assuming that copying or lightly rewriting it will conclusively hide or establish its origin. Checking the content and being transparent about the tools you used remain more informative than a single detector result.
If you use ChatGPT more extensively, you can document which passages you adopted, verified, and rewrote yourself. That can be especially useful if an employer, school, or university later asks how a document was produced. A detector can provide one additional signal at most; it cannot replace a traceable record of your work.
Consider a communications employee who uses ChatGPT to create a long draft and then revises its wording. The watermark may be present in the copied draft but become weaker after extensive editing. That does not show that the employee contributed nothing, nor does a failed detection prove that the final document was produced without AI.
For Switzerland, the situation is currently less clear than it is for the EU and the US. The sources announce neither the launch of visual ads nor default text watermarking for Swiss ChatGPT accounts. Users in Switzerland should therefore not assume that their text already carries the same EU watermark; the reports also provide no information about specific differences among German, French, Italian, and English output.
Customers using OpenAI’s application programming interface, a technical connection to its models, can voluntarily enable watermarked output for selected models worldwide; it is off by default. THE DECODER contrasts this with Anthropic, whose watermark is reportedly mandatory worldwide and across all access methods. OpenAI’s initial approach is therefore regional and less restrictive.
The two changes serve opposing goals: Advertising is meant to make ChatGPT more commercially useful, while watermarking is intended to provide greater transparency about generated text. Clearly separating ads from results is reasonable, as is a provenance signal that does not identify a person. The unresolved risks are whether OpenAI can consistently exclude sensitive advertising contexts and how often a statistical watermark will be misinterpreted or rendered ineffective by ordinary editing.
Sources
- OpenAI is sticking more ads in ChatGPT – The Verge, 2026-10-05
- OpenAI will start watermarking ChatGPT’s text in the EU – TechCrunch, 2026-10-05
- OpenAI is adding text watermarking in ChatGPT and Codex – The Verge, 2026-10-05
- OpenAI versieht ChatGPT-Texte mit unsichtbarem Wasserzeichen – THE DECODER, 2026-10-05
- OpenAI launches visual ads that appear alongside image generation results – TechCrunch, 2026-10-05
- Our approach to EU text provenance rules – OpenAI, 2026-10-05
- OpenAI testet Anzeigen in ChatGPT, die während der Bilderzeugung erscheinen – THE DECODER, 2026-10-05
- Building advertising for the way people use AI – OpenAI, 2026-10-05
Image: Hanna Pad via Pexels