Google opened its SynthID Detector to anyone with an account on 7 October 2026, ending the restriction that had limited an earlier version to media professionals. Two days earlier, OpenAI set out its plan to add an invisible watermark to eligible ChatGPT and Codex text in the European Union “over the coming weeks”, and said it is not making text watermarking a global default at launch.
Both announcements describe the same technique: a machine-readable signal embedded in AI output that compatible software can look for later. The limits attached to each one decide how much a check result is worth, and they sit in the details rather than the launch framing.
Google’s SynthID Detector now takes anyone with an account
Google’s announcement, published by Pushmeet Kohli, vice president for science and strategic initiatives at Google DeepMind, says the company is expanding access to everyone with the tool “available globally in English starting today”.
The detector checks whether an image, video or audio file was made with AI from Google or its partners. Google names OpenAI, NVIDIA and Kakao, and says Apple support is coming “soon”, with no date attached.
Ars Technica reported that the tool now runs on a dedicated site, synthid.com, and that users sign in with a Google, OpenAI or Apple account. The output is simpler than an earlier internal tool: a message telling you whether SynthID was used to produce or edit the media, where the internal version highlighted the regions carrying SynthID pixels.
Ars reported that each user gets “a daily usage quota of approximately 10 image, video, and audio checks”, quoting Google. Google’s blog post does not mention a quota. The Gemini help page describes a similar in-app allowance of about 10 image checks, 10 video checks up to five minutes in total, and 10 audio checks up to three hours in total in any 24-hour rolling window. Ars added that the system may lock a user out sooner for checking many very similar images, and that Google engineers said the cap exists to stop people tuning SynthID removal tools.
Google’s own documentation describes three outcomes rather than two. The help page says a SynthID watermark can be detected; not detected, “but it could have been created by other AI systems”; or unclear, because there were not enough details to watermark or a likely too-small edit.
One inconsistency matters here. The Gemini help page, read on 9 October, still says Gemini “can currently only recognize content created by Google AI tools”, which conflicts with the partner coverage described for the new site. The most consistent reading is that partner watermarks are handled on synthid.com while the in-Gemini check remains Google-only.
OpenAI’s text watermark turns on in the EU and stops there
OpenAI’s post is dated 5 October 2026, a day before most coverage; Ars Technica’s report ran on 6 October. The company says it will add an invisible watermark to eligible ChatGPT and Codex text output “across all plans in the EU only” and states: “We are not making text watermarking a global default at launch.”
API customers worldwide can opt in, with the feature off by default. The detector that reads the mark will be “initially limited to approved researchers and expert organizations”, with no public access at launch. OpenAI says it reports only whether an OpenAI watermark is detected, without identifying the user or revealing prompts or conversations.
The mechanism, textGrain, “adds an invisible statistical signal to the model’s word choices”, and the detector looks for that signal. OpenAI says it plans to open-source the approach.
The accuracy figures are OpenAI’s own. At a 1% target false-positive rate, the company reports about 80% detection in 200-token passages and about 95% in 400-token passages for psychology-type content, with substantially lower rates for mathematics, where the model has less freedom in word choice. Editing degrades the signal quickly: in a 400-token passage, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. OpenAI’s caveat is that “strong performance under ideal conditions does not guarantee reliable detection in everyday use”.
Ars reported the same work as a “92 percent successful detection rate” that falls by “almost 30 percent” after a 10% change and by “almost 75 percent” after a 20% change. OpenAI’s degradation figure uses a 25% change, so this article follows OpenAI’s numbers. No independent test of textGrain’s accuracy was found.
Article 50 is why OpenAI’s default applies in the EU only
Article 50 of the EU AI Act, Regulation (EU) 2024/1689, applies from 2 August 2026. Paragraph 2 requires providers of AI systems that generate synthetic audio, image, video or text to ensure outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated”, using solutions that are “effective, interoperable, robust and reliable as far as this is technically feasible”.
Paragraph 4 puts a separate duty on deployers to disclose deepfakes, and to disclose AI-generated or manipulated text published to inform the public on matters of public interest, unless the content went through human review or editorial control with a person holding editorial responsibility.
The European Commission’s FAQ on Article 50 adds two details that matter for anyone relying on these signals. For systems placed on the market before 2 August 2026, the marking and detection obligation applies from 2 December 2026, and content generated before 2 August 2026 does not need retroactive labelling. The FAQ also states that a deployer “cannot simply rely on the machine-readable marking embedded in the content by the provider under Article 50(2)” to meet the paragraph 4 disclosure duty. Enforcement sits mainly with national market surveillance authorities, and fines can reach EUR 15 million or 3% of worldwide turnover.
The Commission’s voluntary Code of Practice on Transparency of AI-generated Content had around 190 signatories by the end of July 2026, and the Commission has published guidance on the Article 50 obligations and a set of EU labelling icons.
OpenAI frames its EU rollout as a response to the machine-readable marking requirement and says it wants to learn from real-world use before going wider. Press coverage attributes the EU-only default to AI Act compliance, and Ars noted that Anthropic enabled watermarking globally while OpenAI defaults it only where regulators require. No source found in this research shows a non-EU jurisdiction requiring the same thing, so the compliance explanation rests on OpenAI’s stated rationale and press inference.
Watermarks, Content Credentials, legal labelling and AI-writing detectors are separate things
An invisible watermark is a signal inside the content itself: pixels in an image or video, a waveform in audio, or statistical patterns in word choice for text. A compatible detector reads it, and for text that often means the provider’s key or approved access. Watermarks survive some transformations and metadata stripping, and degrade with heavy edits, small edits, short or translated text, and cropping.
Content provenance, in the form of Content Credentials built on the C2PA standard, is signed metadata documenting origin and edit history for both AI and non-AI content. The Gemini help page describes it as a digital passport and notes that Content Credentials do not tell you whether a file is “definitely” AI-generated. The metadata is cryptographically hard to fake and easy to strip through editing, saving or screenshotting. OpenAI says it is C2PA-conformant, attaches Content Credentials to supported image outputs and also embeds SynthID in supported images and audio.
Legal labelling is the third. Article 50(4) puts the human-facing disclosure duty on the deployer, and the Commission’s FAQ is explicit that a deployer cannot discharge that duty by pointing to the provider’s machine-readable marking. Marking is machine-readable and sits with the provider; labelling is human-readable and sits with the deployer.
Third-party “AI-writing detectors” form a fourth category, and they often get folded into these discussions. They infer likelihood from text alone, without the provider’s key or cooperation, which makes them different from both the OpenAI textGrain detector and the SynthID Detector. Google’s help page lists non-watermark methods such as visual artifacts, reverse image search and EXIF or metadata analysis separately from SynthID and Content Credentials. No third-party detector was tested for this article.
Which claims are independently verified and which are vendor figures
Google says more than 180 billion images and videos have been watermarked since 2023, along with “240,000 years of audio content”, and that verification in Search, Gemini and Chrome now handles over 1 million requests a day. Those figures come from Google and have been repeated by reporters; they have not been independently audited.
The strongest independent test found is Ars Technica’s July 2026 experiment with image SynthID, using a script that repeatedly compressed and resized an image. Full-frame watermarks survived 300 generations; a 20% crop after 300 cycles broke detection, and a 50% crop broke it at about 250 iterations. That is one journalist’s scripted test, not a peer-reviewed evaluation.
Google’s original SynthID paper, cited by Ars, says the system is not designed to withstand adversarial attacks. Claims circulating that SynthID has been “hacked” could not be verified, and Pushmeet Kohli told Ars his team could not replicate them. Meta’s Content Seal watermark is a different system, and Ars, citing Reuters, reported in July 2026 that cropping often removed it.
For textGrain, the accuracy figures above are vendor evaluations. No independent, peer-reviewed evaluation of either tool’s real-world accuracy was found, and no hands-on testing was performed for this article. The textGrain figures come from OpenAI’s own summary of its technical report, and the check quota rests on Google’s statement to Ars Technica and the Gemini help page.
How to read a result from either tool
A hit from SynthID Detector tells you that a specific provider’s watermark is present, which points to that provider’s tooling. OpenAI’s post lists what its own watermark does not do: it does not measure human contribution, establish ownership or responsibility, identify the user or verify accuracy.
A miss is not a verdict on the file. Google’s help page says no watermark found means the content was not created or edited by Google AI, and adds that it could have been created by other AI systems. OpenAI states that “The absence of a detected watermark does not prove human authorship.” Meta’s systems use different technology, and open-weight or self-hosted models do not carry these signals, so plenty of AI-generated material will never be identified by these checks. Ars also flags the reverse risk: as awareness grows, people can dismiss genuine footage as fabricated, a problem sometimes called the liar’s dividend.
The two launches leave users in very different positions. Anyone with an account can now check an image, video or audio file at synthid.com. Text is the harder case, because OpenAI’s mark is arriving in the EU and the detector that reads it is closed to everyone outside its approved researchers, so there is no public way to check whether the mark is present.