OpenAI backs EU AI content labeling code as August deadline closes in
The June 11th endorsement ties OpenAI's provenance work to a voluntary path for meeting EU AI Act transparency rules.
By Ryan Merket ยท Published
Why it matters
OpenAI's endorsement shows how AI labs are turning provenance into compliance infrastructure before EU transparency rules start applying on August 2nd.
OpenAI (@OpenAI) backed the European Commission's AI-generated content transparency code in a June 11th post, putting its image provenance work behind a voluntary compliance route just weeks before the EU AI Act's Article 50 transparency obligations begin applying on August 2nd (Commission FAQ).
The post was published by OpenAI, not by a named executive, and it reads like a policy position from a company trying to make compliance legible before the rules start to bite. OpenAI says it supports the Code of Practice on Transparency of AI-Generated Content and contributed to its development alongside hundreds of stakeholders. The European Commission published the final code in June 2026 as a voluntary framework for marking and labeling AI-generated content.
The timing matters. Article 50 obligations are scheduled to apply from August 2nd, and the Commission says the rules cover machine-readable marking and detection of AI-generated content, labeling of deepfakes, and disclosure for certain AI-generated publications. The code gives providers and deployers a common route for demonstrating compliance, subject to an adequacy assessment by the Commission and the AI Board. Companies that skip the code can still comply, but they will have to prove their own measures to market surveillance authorities.
For OpenAI, the endorsement turns provenance from a trust-and-safety project into a market-access instrument. The company behind ChatGPT and the OpenAI API has spent 2026 pushing deeper into default user workflows, from the GPT-5.5 Instant update to its Codex and infrastructure work. RuntimeWire also reported on OpenAI's Jalapeno inference chip and its reported 5% U.S. government stake proposal, two separate signs that OpenAI is negotiating with governments and cost structures at the same time it ships models. The EU transparency code sits in that same operating reality: the company cannot scale ChatGPT, Codex, image generation, and enterprise products globally without turning governance into product plumbing.
What OpenAI is actually committing to
OpenAI is not claiming that it can reliably detect every AI-made image, audio clip, video, or text passage once that media starts moving across the internet. Its post is more measured. OpenAI says its current approach (details here) combines C2PA Content Credentials for images, SynthID watermarks for images generated with ChatGPT, Codex, and the API, a public verification flow at openai.com/verify, standards work, and product safeguards such as policies, classifiers, reporting channels, and enforcement processes.
That stack is deliberately layered because each single technique breaks in predictable ways. OpenAI says metadata can be stripped, lost through uploads and downloads, or damaged by file conversions, resizing, and screenshots. Watermarks can weaken as content is transformed. Labels only help when users actually see them. The practical claim is narrower than the policy headline: OpenAI is aligning with the EU's preferred direction while admitting that provenance remains a work in progress.
The company's C2PA work predates this code. OpenAI says it began adding C2PA metadata to images created and edited by DALL-E 3 in ChatGPT and the API in 2024, joined the C2PA Steering Committee that year, and released its first public verification tool before the June 11th endorsement. Those details matter because the EU code is about operational compliance, not public statements of goodwill. A provider needs a machine-readable signal, a detection path, and documentation that regulators can inspect.
The code gives builders a compliance shortcut, if Brussels accepts it
The Commission's code separates the market into roles. Providers of generative AI systems are invited to sign the section concerning machine-readable marking and detection. Deployers are invited to sign the section covering disclosure and labeling of deepfakes and certain AI-generated or manipulated text published to inform the public on matters of public interest. An organization that acts as both provider and deployer can sign the full code.
The deadline mechanics are important for any company building on top of OpenAI, Google, Anthropic, Meta, Mistral, or another model provider. Initial signatories must submit forms by July 22nd at 18:00 CEST to be included in the list published before August 2nd (Commission FAQ). Companies can sign later, but the Commission says Article 50 still applies from August 2nd. The code does not replace the AI Act, and signing it is voluntary. Compliance with Article 50 is the legal requirement.
The Commission also says the code is still subject to an adequacy assessment. If that assessment is positive, signatories can rely on the code's measures to show compliance across EU member states. That is the incentive: a common recognized framework instead of fragmented evidence requests from different authorities. For a company with OpenAI's product surface area, that certainty has business value.
The hard part is distribution, not the watermark
OpenAI's endorsement exposes the unresolved technical bottleneck in AI-content transparency. A provenance signal is useful only if it survives outside OpenAI's own product boundary and can be read by platforms, publishers, marketplaces, tools, and end-user devices. That is why OpenAI points to C2PA, whose members work on interoperable standards for content credentials. The code's real test will be whether providers, deployers, platforms, and detection vendors can make labels survive ordinary internet behavior.
OpenAI has an advantage because it controls major generation surfaces. It can add metadata to images created in ChatGPT, tag outputs from the API, and offer a verification page. It has less control once users download, compress, screenshot, repost, or edit those outputs. The same weakness affects every provider pursuing watermarking or provenance. The EU code gives that messy problem a regulatory frame, then leaves implementation to companies that have every reason to make compliance cheap, defensible, and minimally disruptive to users.
That is the builder story inside the policy announcement. OpenAI's founders built the company around a mission to ensure artificial general intelligence benefits humanity; its current structure puts the nonprofit OpenAI Foundation in control of OpenAI Group PBC. In practice, the company is now forced to translate that mission language into repeatable systems for labeling, provenance, auditability, and user disclosure. The June 11th endorsement is one more step in that translation.
The open question is product scope. OpenAI says it will comply with requirements that apply to its relevant products, but its post does not enumerate every covered feature, product line, or deployment context. It also does not define every supported image format for the verification tool. Those omissions are normal in a policy blog post, but they are exactly where implementation risk lives. By August 2nd, the market will care less about who endorsed the code and more about which AI outputs carry durable signals once they leave the tool that made them.