ChatGPT embeds invisible image markers

OpenAI has started embedding hidden provenance markers in images generated through ChatGPT, Codex and the OpenAI API, adding a technical way for outside parties to identify AI-made images. The company said on May 19 that it is adopting a “multi-layered approach” that combines standardized C2PA metadata with Google DeepMind’s SynthID watermarking for images. (openai.com) The change matters because metadata alone can be stripped when images are reposted, recompressed or converted. OpenAI said the added watermark is meant to make provenance “more resilient,” while third-party coverage said the hidden signals can survive screenshots, compression and format changes. (artificialintelligenceact.eu) ### What exactly changed in ChatGPT’s image output? OpenAI said the new provenance signals apply to images generated through ChatGPT, Codex and the OpenAI API. In its announcement, the company said it is incorporating Google DeepMind’s SynthID watermarking “starting with images” from those products. (openai.com) TechTimes reported that every image generated through those OpenAI tools since May 19, 2026 carries two invisible signals embedded in the image data. A parallel report from International Business Times described the same rollout and said a public checking tool is available for those signals. (openai.com) ### Why use both metadata and an invisible watermark? C2PA metadata is already a common provenance method for AI images because it can record where content came from and how it was created or edited. (openai.com) OpenAI said that metadata is “an important foundation” but also said it is “not foolproof,” which is why it added watermarking as a second layer. PetaPixel described the combination as the two leading protocols for identifying AI images. That pairing matters because metadata helps compliant platforms and tools read origin information, while watermarking is designed to persist when file metadata is lost. (techtimes.com) ### Can other people actually detect the marker? TechTimes reported that third parties can detect the hidden provenance signals and use them to flag AI origin. The same report said a free public tool lets users check images for both signals. (openai.com) OpenAI’s own post did not frame the feature as a consumer-facing accusation tool. Instead, it described the rollout as part of broader “content provenance” work intended to support safer and more transparent AI-generated media. (petapixel.com) ### Why is August 2, 2026 part of this story? Article 50 of the EU AI Act is scheduled to apply from August 2, 2026, according to the AI Act text and EU-facing implementation materials. The provision covers transparency obligations for certain AI systems, including rules around AI-generated or manipulated content in some cases. (techtimes.com) California’s AI Transparency Act is also set to begin on August 2, 2026 after the state delayed its operative date to align with the EU timeline, according to legislative tracking and bill summaries. A California bill summary says the law generally regulates provenance-data disclosure in AI-generated content and requires a covered provider to make an AI detection tool available at no cost. (openai.com) ### Is this a legal requirement or a product choice? OpenAI’s May 19 announcement presents the marker rollout as a product implementation choice. (artificialintelligenceact.eu) The company said provenance only works if it survives beyond the first platform where content is created, and said conformance and layered signals make that possible. The regulatory backdrop is becoming more concrete. A Hogan Lovells note on European Commission draft guidance said the Commission issued draft Article 50 guidelines on May 8, 2026, while another legal analysis said the consultation runs until June 3 and the guidance is intended to apply alongside Article 50 from August 2. (calmatters.digitaldemocracy.org) ### What happens next? June 3, 2026 is the end date for consultation on the European Commission’s draft Article 50 guidance, according to legal summaries of the draft. (openai.com) August 2, 2026 is the date when the EU Article 50 regime and California’s transparency timetable are set to take effect, giving platforms, labs and detection-tool providers a near-term compliance milestone. (twobirds.com) (hoganlovells.com)