OpenAI announced on June 11, 2026 that it supports the European Commission’s Code of Practice on Transparency of AI-Generated Content. The Commission published the code on June 10, 2026 as a voluntary path for providers and deployers to help meet the EU AI Act’s transparency obligations for marking and labelling AI-generated content.
The timing is the practical hook. The Commission says Article 50 transparency obligations apply from August 2, 2026. OpenAI is positioning its provenance work, including C2PA metadata, SynthID watermarks, and openai.com/verify, as part of the compliance architecture forming around generated media.
The code sits between voluntary process and legal duty
The European Commission describes adherence to the code as voluntary. The underlying AI Act transparency requirements are legal obligations. That distinction is easy to blur, and it is the main thing readers should keep straight.
The code is meant to help providers and deployers comply with obligations around marking, detecting, and labeling AI-generated content, including deepfakes and certain AI-generated text publications. The Commission says the code has two sections: provider rules for marking and detection, and deployer rules for labeling deepfakes plus AI-generated or manipulated text.
In practical terms, the EU is trying to turn a broad legal requirement into operational patterns that companies can implement and auditors can understand.
OpenAI is leaning on provenance infrastructure
OpenAI says its support builds on provenance work that began with C2PA metadata in DALL-E 3 images in 2024. Its current approach includes C2PA Content Credentials for images, SynthID watermarks for supported OpenAI-generated images, a public verification experience at openai.com/verify, open-standards work, and policy enforcement for deceptive uses.
The important detail is redundancy. OpenAI says C2PA metadata is useful because it can carry information about origin, edits, and signatures. It also says metadata can be stripped, lost through uploads and downloads, or broken by transformations such as resizing and screenshots. That is why OpenAI frames provenance as a multi-layered system rather than a single label.
Verification is becoming a product surface
OpenAI’s public verification tool is part of the shift from policy statement to user-facing product. If a supported image carries provenance signals associated with OpenAI-generated images, the verification experience can help people inspect that signal.
That does not solve the whole problem. Labels only help when people see them. Watermarks and metadata only help when they survive the path from generation to publication. Bad actors can try to remove, transform, or launder content through other systems. OpenAI acknowledges those limits in its post.
Still, the EU code gives companies a reason to make provenance visible, documented, and repeatable. For AI providers, content transparency is moving from trust-and-safety language into product requirements, developer documentation, and compliance reviews.
The compliance burden will spread beyond model labs
The Commission says the code is for both providers and deployers of generative AI systems. That means the work does not stop at the model API. A media company using generated images, a campaign using AI-manipulated video, or an enterprise publishing AI-generated public-interest text may also need labeling practices.
For builders, the immediate work is inventory. Which systems generate images, audio, video, or public-facing text? Which outputs already carry metadata or watermarks? Which publishing workflows strip metadata? Which user interfaces need labels?
What to watch next
The next checkpoint is the Commission’s assessment of the code and any guidance on the scope of Article 50 obligations. The second checkpoint is signature behavior: which AI providers and deployers sign, and whether their product surfaces change before August 2.
For OpenAI, watch whether provenance expands beyond image workflows into more formats and more obvious user-facing labels. The company’s current post focuses heavily on images, C2PA, SynthID, and verification. The harder ecosystem question is how well those signals survive once generated content leaves the original product.
For broader model context, see our AI model leaderboard. For related company coverage, see OpenAI.
