New AI Transparency Laws Just Kicked In. They Don't Cover a Single Word of Text.

Quick answer

On August 2, 2026, California's AI Transparency Act (SB 942) and Article 50 of the EU AI Act both became enforceable, requiring the biggest AI companies to ship a public tool that lets people check whether an image, video, or audio clip came from their systems. Investigators at Indicator, working with the nonprofit WITNESS, tested 13 of those companies this month. Seven had no detector live at all. Of the seven that did, only two — Google and OpenAI — held up once a file had been edited. Text isn't covered by either law. If you want to check whether something was written by AI, a detector built for that job is still the only option on the table, law or no law.

What the new laws actually require

California's SB 942, sponsored by state senator Josh Becker and signed in 2024, became operative on August 2, 2026. It applies to any generative AI provider with at least one million monthly users in California. Covered companies must embed provenance data in AI-created or AI-edited images, video, and audio, and they must offer a free, publicly accessible detection tool that lets someone check that provenance. The EU AI Act's Article 50 landed the same week with a similar disclosure requirement for the same three media types. Neither law says a word about text.

That's not an oversight so much as a scoping choice: images, video, and audio can carry embedded technical markers (watermarks, C2PA metadata) that survive in the file itself. Text is harder to mark invisibly and trivially easy to strip — copy a sentence into a new document and any visible disclosure is gone. Lawmakers picked the media where a technical fix was closer at hand. That leaves the exact question a classroom, a newsroom, or a hiring manager usually asks — "did a person write this?" — completely outside the new compliance regime.

The compliance test nobody passed cleanly

Indicator, working with WITNESS, ran what looks like the first real-world audit of the new rules. They checked 13 companies covered by the California law: Google, Meta, Microsoft, OpenAI, Adobe, Grok, Midjourney, Mistral, HeyGen, Synthesia, ElevenLabs, Suno, and TikTok. Anthropic wasn't included, since it doesn't generate photorealistic images or video and the law doesn't reach text.

The first finding is blunt. Seven of the 13 didn't have a public detection tool available at all when the investigation ran: Grok, Synthesia, Suno, HeyGen, Midjourney, and Mistral, plus Microsoft initially, before the team credited an inspection feature buried in Microsoft Designer. That's close to half the covered companies sitting outside a law that took effect on a fixed date, according to reporting on the investigation from the Transparency Coalition.

The second finding is the one that matters more. Where a detector did exist, the team ran 243 tests across 85 files — four original and four edited files from each company's own generator, plus files from competing tools, plus six real, non-synthetic files (two deliberately altered to carry forged AI credentials, checking for false positives in the other direction). Most tools could flag their own unedited output without much trouble. Once a file had been cropped, recompressed, or otherwise edited, "results naturally degraded," in the investigators' words. Only Google's and OpenAI's detectors kept correctly identifying their own content after tampering. Under California's law, noncompliance can carry a civil penalty of $5,000 per violation, per day — real money, and still not enough, on this evidence, to get every covered company's own detector working on the one test that matters: does it survive someone actually trying to fool it?

Why this is a text-detection story, not just an image one

It would be easy to read that and think: rough month for image and video provenance tools, glad text isn't my problem. It's the opposite. The investigation shows that even with a legal mandate, a compliance deadline, and public penalties on the table, half the companies required to build a detector hadn't shipped one, and most of the ones that had shipped one buckled the moment a file was edited. That's the outcome under regulatory pressure specifically engineered to make detection easy: the AI company controls the generator, controls the marking, and still only two out of seven got it fully right.

Text detection has never had that kind of regulatory tailwind, and it works from a harder starting position anyway — no legally mandated watermark, no provenance metadata a browser can read, nothing embedded by the model itself for most tools to check against. Anthropic's new Claude watermark, which we covered last week, is the first real move toward that for text from a frontier lab, and even Anthropic's own documentation says the mark signals "Claude touched this," not "a person didn't write this," and it disappears under normal paraphrasing. Nobody has shipped a text-provenance system remotely as mature as what California just found lacking on the image and video side. Independent statistical detection — reading the fingerprints that come from how a model actually generates language, not a marker it was told to embed — is still doing the entire job for text, with zero legal requirement propping it up.

What this changes for anyone checking written content

If you're an educator, an editor, or a hiring manager who assumed "the platforms will handle disclosure now that there's a law" — that assumption never applied to the written word, and this investigation shows it's shaky even for the media types the law does cover. A model card claim ("we support provenance") and a working detector are two different things, and regulators just found the gap between them at scale.

The practical takeaway is the same one we've made in our own guide to what a detection score actually means: a score is an investigative signal, not a government-mandated stamp. No law is going to hand you a text watermark to check against, this year or probably next. A detector trained to read the statistical fingerprint of generated writing — sentence rhythm, word-choice patterns, structural tells that don't require the model's cooperation — is the only tool doing this job for text right now, law or no law.

FAQ

Does the California AI Transparency Act apply to AI-written text? No. SB 942 requires detection tools for AI-generated or AI-altered images, video, and audio from covered providers. Written text isn't covered by the law at all.

Which companies didn't have a public AI detection tool during the test? During the test, Grok did not have a dedicated public AI detection tool. Synthesia likewise lacked such a tool at that time. Neither Suno nor HeyGen offered a public AI detection feature during the evaluation. Midjourney also did not provide a public AI detection tool. Mistral was without a dedicated public AI detection tool as well. Later, Microsoft received credit for an inspection tool integrated within Microsoft Designer.

Which detectors held up best in the test? The benchmark began with Google's detector and OpenAI's counterpart. Despite modifications to the documents, each apparatus still flagged its own synthetic content accurately. A total of seven companies supplied detectors for the assessment.

If there's no law requiring text detection, why would a company or school use one anyway? Because the underlying problem — knowing whether a person wrote something — doesn't go away just because it isn't regulated yet. Schools, publishers, and hiring teams already rely on independent detection for exactly that reason.

Is TheChecker.AI affected by these laws? No, these laws target the companies that generate AI content, not tools that detect it. Try TheChecker.AI's detector to check any piece of writing for AI-generated patterns, free to start.