Google and Meta signed the EU’s AI labelling code. Their detectors miss each other’s marks

Google opened SynthID Detector to everyone on 7 October. It checks images, video and audio for a hidden mark, but only marks from four companies: Google, OpenAI, Nvidia and Kakao. Meta’s own checker, shown in July, reads only Meta’s mark. Both firms signed the EU code on AI labelling, which the European Commission published on 10 June. Four months later, no single tool can tell an advertiser where an image came from. Under the AI Act, the job of labelling an AI-made ad falls on whoever made it.

Google and OpenAI signed the EU code. TikTok, X and Apple are missing

The code is voluntary. It has two parts. The first tells AI providers how to mark and detect what their tools produce. The second tells the companies that use those tools how to label it for the public. It also offers an optional EU icon in three versions. By the Commission’s update of 24 September, 95 firms had signed the first part, including Google, Meta, Microsoft, OpenAI, Mistral and Anthropic. The second part had 192 signatories, among them Lufthansa, Lenovo and Getty Images. TikTok, X and Apple were not on either list.

Law firm Freshfields, in its reading of the final text, says signatories must use at least two layers of machine-readable marks: signed metadata and a watermark. They must also offer a free detection tool, and those tools should work with each other by 2 February 2027. Until then, each company checks mainly its own marks.

Google’s checker misses Meta’s mark, and Meta’s misses Google’s

An AI watermark is not a logo in the corner. It is a hidden pattern inside the file that people cannot see and software can read. The second method is C2PA metadata, a signed note inside the file that records which tool made it and when.

Meta added its own watermark, Content Seal, to images from its Muse Image model in July. Meta says the mark survives cropping, compression and screenshots, and will reach video later. Its checker at meta.ai handles images only and has a daily limit. Engadget found it cannot read SynthID or C2PA. Indicator found that images from meta.ai carry no C2PA metadata at all, even though Meta sits on the C2PA steering committee.

Google puts SynthID into images, video, audio and text from its own tools. It says more than 180 billion images and videos now carry the mark. The public detector works worldwide, in English only, and does not check text. Google says Apple files will follow. According to the tool’s own description, quoted by MacRumors, it cannot tell content made by AI from content only edited by AI.

OpenAI gives images from ChatGPT both C2PA metadata and Google’s SynthID, so Google’s detector sees them too. OpenAI’s own verify page checks images and audio, but only from OpenAI. On 5 October it started a text watermark called textGrain. In ChatGPT it will arrive over the coming weeks, and only in the EU. In the API it is off by default and available worldwide. Only approved researchers and organisations can use the text detector.

TikTok has read C2PA since May 2024 and labels AI content made with other tools. Creators must label realistic AI video themselves. By November 2025 TikTok had labelled more than 1.3 billion videos. It is testing a hidden mark that only TikTok can read. It admits labels can disappear when a video is reuploaded or edited elsewhere.

Adobe and Microsoft rely on metadata. Adobe adds it to everything made in Firefly. Microsoft adds it to images from OpenAI models in its Azure cloud. OpenAI’s help pages note the weak spot: metadata can be lost after an ordinary screenshot.

Instagram labelled 15 of 105 AI images in a March test

Labels that people see are a separate problem. In March, Indicator posted more than 200 images and videos from Google, Meta and OpenAI tools on five platforms. LinkedIn and Pinterest labelled 67% of them. YouTube labelled about half and TikTok about a third. Instagram labelled 15 of 105 images.

Some platforms have moved since. In May YouTube began adding labels itself when it detects undisclosed realistic AI video. On 1 October it said its likeness detection would soon cover voices as well as faces, without giving a date.

Text is harder still. In OpenAI’s own tests, textGrain was found in about 95% of 400-token passages. When one word in ten was swapped for a synonym, detection fell from about 92% to 66%. With one word in four swapped, it fell to 17%. OpenAI adds that a missing watermark does not prove a person wrote the text. Google, for its part, already tells sites to check AI content by hand.

Italy, Spain and France add rules of their own

Some governments are not waiting for the code to settle. Italy’s Department for Information and Publishing issued a circular on 16 July for public-service spots that ministries air in free slots on RAI. Spots made with AI must show “Realizzato con contributo IA”, or “made with AI contribution”, for at least seven seconds. AI voices and AI actors are banned in those spots.

Spain’s government approved an AI bill on 26 May, and it is now in parliament. Since 2023 France has required influencers to mark AI-made images as “images virtuelles”. In April its government told the Senate it would rather enforce Article 50 at EU level than write new national rules.

In Poland, a law published on 27 July set up a new AI watchdog, the Commission for the Development and Safety of AI. According to the news agency PAP, its powers to inspect and fine start on 28 October. The Polish ad body IAB Polska published a guide on 5 August. It says the label belongs to whoever works the AI tool, usually the creative agency. Polish agencies, advertisers and lawyers had already agreed on who adds the label, and that only content that looks real needs one.

Under Article 50, advertisers and agencies carry the label

Article 50 of the AI Act has applied since 2 August. AI providers must mark output so that software can detect it. Anyone who uses AI to make a realistic image, sound or video must tell the audience. Fines go up to €15 million or 3% of worldwide turnover, whichever is higher. For small and medium firms, the lower amount applies.

The ad platforms help only in part. Since July, Google Ads adds an automatic note to ads made with its own AI tools. It sits in a panel called “How this ad was made”. For ads made elsewhere, the advertiser must tick a box. Since June, Meta has been collecting AI notes in “About this ad”. It says it will detect ads made with other companies’ AI tools, but it has not said which signals it reads. A file from another tool that carries no mark, with nothing ticked, gets no label.

The European Advertising Standards Alliance, which represents ad self-regulators, welcomed the Commission’s Article 50 guidelines in July. It also warned that they stretch the idea of a deepfake beyond the wording of the AI Act. In its view, the ad industry needs its own labelling standards.

“Detectors won’t do this job for agencies, because each one reads only some of the marks,” says Grzegorz Kubicki, CEO and founder of iMarketing. “For a campaign that runs in several EU countries, the simplest route is one label added in production that meets the strictest local rule.”

Two dates come next. AI tools that were on the market before 2 August must mark their output by 2 December 2026. Under the code, detection tools should work together by 2 February 2027. Until then, an image that passes one checker can still be AI-made: it may come from Meta, or it may have lost its metadata in a screenshot.