Built on an AI model API? Under the EU AI Act, you may have to watermark the output

On 5 October 2026, OpenAI said it would start adding an invisible watermark to eligible ChatGPT and Codex text output in the EU, and gave the EU AI Act as the reason. In the same announcement it said that for API customers, text watermarking is opt-in for select models and "will remain off by default in the API."

That second detail is the one that matters if you build products on a model API. Under the AI Act, the duty to mark AI-generated content does not attach to the model as such. It sits with the provider of the AI system that generates the content, and that is often the company that built the app.

This is an information service to help you orient, not legal advice. For the obligations tied to your own AI use, see which apply to your company.

The short answer

Your situation	Who carries the Article 50(2) marking duty	Source
You build a chatbot, writing tool or agent on a third-party model and offer it under your own name	You, as provider of the AI system	Art. 3(3); Art. 50(2); guidelines para. 11
You build it in-house and use it only internally, under your own name	You, as provider (putting into service) — check the B2B exception below	Art. 3(3); guidelines para. 11, 87
The model provider also built the system you use (e.g. its own chat app)	The model provider, as provider of that system	Art. 3(3); Art. 50(2); guidelines para. 27
You only use someone else's finished AI tool to create content	Not you for 50(2); but check Article 50(4) deepfake / public-interest text labelling as a deployer	Art. 3(4); Art. 50(4)
You only host or pass on content others generated	Not a deployer within the meaning of the AI Act; strongly encouraged, not required, to preserve marks	guidelines para. 16

Why the duty lands on the app builder

Article 50(2) is addressed to "Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content". They "shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."

A provider under Article 3(3) is whoever "develops an AI system … or that has an AI system … developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge." Wrapping a third-party model in your own product, under your own brand, fits that description.

The Commission's guidelines on Article 50 (approved 20 July 2026, C(2026) 5054 final) say so with an example: a company offering "a generative or interactive AI application (e.g. a chatbot, image generator, AI agent) on the Union market under its own name or trademark" is the provider responsible for Article 50(1) and/or (2). That holds whether the product is free or paid, and wherever the company is established. If what you ship is an agent rather than a chatbot, the same provider logic applies; see what the AI Act requires of AI agents.

The guidelines (para. 27) also note that Article 50 "does not explicitly apply to GPAI models". Model providers are encouraged to build marking in at model level, which the guidelines say is "particularly beneficial to facilitate compliance by downstream AI system providers". Encouraged is not obliged, and that is why the gap exists.

You may rely on the model's watermark, but you still own the result

The guidelines (para. 74) let a provider implement marking at different stages of the value chain, for example at the level of the AI system, at the underlying model or in the inference process. A provider "may rely on the marking solution implemented by an upstream model provider or a third party providing the solution … to the extent that the marking solution is compliant with Article 50(2) AI Act." The next sentence matters most:

"Such reliance is without prejudice to the responsibility of the provider of the AI system to demonstrate compliance with Article 50(2) AI Act."

The same logic applies to detection. If you depend on another company for the detection tool, the guidelines (para. 78) say you "must ensure that the detection solution complies with Article 50(2) and (5)". Article 50(2) requires outputs to be detectable, not just marked, and para. 75 says the means of detection must be available to the people exposed to the content.

The practical consequence:

If your model provider's marking is…	What that means for you
On by default for your outputs, and meets Art. 50(2)	You can rely on it, but keep evidence of what it is and that it meets Art. 50(2)
Opt-in / off by default	Turning it on is your decision; if you leave it off, your outputs may be unmarked
Not offered at all	You need your own or a third-party marking and detection solution
Marked, but detector access is restricted	Check how the Art. 50(2) detectability and Art. 50(5) requirements are met for people exposed to your output

Text is the hard case. OpenAI's own figures show why. At a 1% false-positive target, OpenAI reported that its detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology, with substantially lower rates for content such as mathematics. Replacing 25% of words with synonyms cut detection in 400-token passages to 17%. Article 50(2) asks for solutions that are effective, interoperable, robust and reliable "as far as this is technically feasible", taking into account the limits of each content type, the costs of implementation and the generally acknowledged state of the art. The guidelines (para. 81) treat technical feasibility as objective: you need not use a solution "not yet developed or available on the market", but feasibility "is not dependent on the specific resources and capabilities of individual providers". Being small does not by itself make a solution unfeasible. The costs of implementation are a separate factor that Article 50(2) and the guidelines (para. 85) allow to be weighed; how far that goes for a small provider is not settled.

What does not need marking

Article 50(2) itself excludes systems "to the extent" they perform "an assistive function for standard editing" or "do not substantially alter the input data provided by the deployer or the semantics thereof". The guidelines (paras. 90–92) draw the line with examples:

Generally no marking (exception)	Marking required
Grammar correction, spellchecking, minor stylistic polishing	AI-generated summaries of text
AI-generated translations of text	Paraphrasing or rewriting that changes style, structure and meaning
Transcriptions of conversations	Generating new content (standard editing "does not involve generating new content", para. 90)
Formatting and format conversion	Composite images or video clips, or any other substantial alteration of the content

If one system does both, the exception covers only the content that was altered in a minor way (para. 92). A writing assistant that both fixes typos and drafts new paragraphs has to mark the drafted paragraphs.

The guidelines also describe a narrow business-to-business case (para. 87) where no marking is required. It applies only if all three conditions are met: the output is strictly technical (for example engineering designs, technical instructions or internal documentation processes), only a limited, pre-defined group of professionals in the provider's and deployer's organisations sees it, and it is not intended to leave the company, with safeguards against foreseeable misuse. Public-facing and consumer-facing systems are expressly excluded. Most SaaS products that generate text for customers will not qualify.

The deadline depends on your launch date

Your generative system was placed on the market…	You must comply with Article 50(2)	Source
Before 2 August 2026	By 2 December 2026 (transitional period; Art. 50(2) itself has applied since 2 August 2026)	Art. 111(4)
On or after 2 August 2026	From 2 August 2026, no transitional period	Art. 113

The transitional period covers only the Article 50(2) marking and detection duty. The guidelines (para. 153) say that for a system that is partly interactive and partly generative, the Article 50(1) duty to tell people they are talking to an AI has applied since 2 August 2026. Outputs generated before 2 August 2026 do not need marking retroactively (para. 154). Note that Article 111(4) as enacted, and recital 38 of Regulation 2026/1744, refer only to systems "placed on the market" before 2 August 2026. The Commission's guidelines (para. 153) describe the transitional rule as covering systems "placed on the market or put into service" before that date. The guidelines are non-binding, so whether a purely in-house system put into service before 2 August 2026 gets the extra time is not settled; do not assume it does. For the rest of what lands that day, see the 2 December 2026 deadline.

Who checks, and what is at stake

For a system built on another company's model, supervision normally sits with national market surveillance authorities, not the AI Office. Under Article 75(1) as amended, the AI Office is exclusively competent for providers of AI systems built on a general-purpose AI model where the model and the system are developed by the same provider or by companies in the same undertaking (with carve-outs, for example Annex I products), and for AI systems that form part of a designated very large online platform or search engine. If your app is built on a general-purpose AI model from a company in your own group, or forms part of such a platform, the AI Office is normally the supervisor for your provider obligations, subject to the Article 75(1)(a) carve-outs. See who enforces the EU AI Act.

Breaches of Article 50 fall under Article 99(4)(g): administrative fines of up to EUR 15 000 000 or 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups, Article 99(6) caps each fine at whichever of the two is lower; new Article 99(6a) gives SMCs the same lower-of cap for fines under paragraphs 4 and 5. Where national authorities handle the case, Member States set the penalty rules, and Article 99(1) as amended requires them to take into account the interests of SMEs (including start-ups) and SMCs and their economic viability. Where the AI Office is competent, it can impose fines itself under Article 75c, applying Article 99(3) to (7). For context, see how much AI Act fines can be.

Signing the Code of Practice on Transparency of AI-Generated Content, which the Commission assessed as adequate in an opinion of 8 July 2026, is one route the guidelines recognise for demonstrating compliance with Article 50(2) and (5) (paras. 84, 146–148). The Commission's opinion on the Code adds that adherence "does not constitute conclusive evidence of compliance". Not signing is also a legitimate choice, but non-signatories are expected to show compliance by other adequate means and may receive more information requests (para. 148).

What to do this month

1. List every feature that generates text, images, audio or video, internal tools included. Sort each into "exception" (editing, translation, transcription), the narrow internal B2B case (para. 87, all three conditions met) or "marking required".
2. For each model API you use, write down whether marking is on, off by default or unavailable, and who provides detection. That record is the starting point for the evidence you would need if an authority asks.
3. Decide whether to opt in. If your provider offers marking off by default, enabling it may be the simplest route; check that it covers every model you use and that detection is available to the people exposed to your output. If you leave it off, write down why and what you use instead.
4. Check your launch date against 2 December 2026. Systems placed on the market on or after 2 August 2026 are already in scope.
5. Keep Article 50(1) and 50(4) separate in your plan. The "you are talking to an AI" notice and deepfake/public-interest labels are different duties; see the Article 50 transparency obligations and the disclosure template.

To see which obligations and which dates apply to your own systems, answer a few questions about your AI use. To be told when the marking guidance or the Code of Practice changes, join the waitlist.

The official text is Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744. The Commission's guidelines on Article 50 are non-binding. This article is an information service to help you orient — it is not legal advice.