On 2 August, the EU AI Act's transparency rules came into force. California followed the same week, requiring major AI companies to embed hidden fingerprints in the images, audio and video they generate.
You've probably seen the headlines. AI watermarks. Digital fingerprints. New laws forcing tech companies to brand their output like cattle. And if you've been using Claude or ChatGPT to draft your marketing emails, your first thought might have been: am I about to get caught?
I'd ask a different question.
What the headlines got backwards
The laws are real, and the direction of travel is clear. The big tools are being pushed to leave a trace in what they produce, and that pressure is only going one way.
For images, audio and video, it works. An AI-generated picture has millions of pixels. You can nudge thousands of them by amounts no human eye can see, and that hidden signal survives cropping, compression and reposting. The fingerprint is real, and it lasts.
Words are a different story.
Every word in a sentence is visible and doing a job, so there's nowhere to hide a signal. The only technique that exists works like this. When the AI is torn between near-identical words ("quickly" or "rapidly"), it can be quietly nudged towards one, in a pattern only the AI company can check for later. Over hundreds of words, that pattern becomes faintly detectable. Edit the copy properly and it washes out completely. Short copy can't carry it at all.
And most tools don't even bother. At the time of writing, ChatGPT and Claude don't watermark text. Only Google's Gemini does it by default.
So most AI-written copy is undetectable, and no reliable technology can prove a piece of edited text was written by AI.
The threat you should actually worry about
Because real detection barely works on text, a cottage industry of AI detectors has grown up that works on something far less reliable.
Vibes.
These tools scan writing for patterns that feel machine-made, and they get it wrong all the time. They routinely flag human writing as AI, especially polished, structured, professional copy. Which is exactly what good business writing looks like. Students have been wrongly accused. So have professional copywriters. The tools can't tell the difference, and neither can the people waving them around as evidence.
If a client, a competitor or a platform ever shows you an AI detection score as proof of something, know what it is: a guess dressed up as a verdict. Unless it's the AI company checking its own watermark, it proves nothing. Don't accept it as evidence, and don't use these tools to judge your own suppliers either.
The bigger risk for you is being falsely accused, and not knowing how to push back.
Where the fingerprints are real
For pictures and sound, take this seriously. AI-generated images, video and audio from the major tools now carry embedded data saying who made it, with what system and when. From January 2027, large platforms will start showing that information to users. By 2028, new phones and cameras will embed "this is a real photo" credentials at the point of capture. The infrastructure is being built now.
In plain English? The AI image in your brochure, the AI voiceover on your explainer video and the AI face on your website will increasingly announce themselves.
I wouldn't stop using the tools. I'd use them where nobody's expecting a real photo, like illustrations, concept visuals and backgrounds. And I'd be careful with anything presented as a real person, a real event or a real testimonial. Fingerprints aside, you shouldn't be faking those anyway.
What has always mattered
Keep a person in the loop on everything you publish. The EU rules do treat AI-assisted content that someone has reviewed differently from raw machine output. The better reason is that it produces better work.
Never present AI-generated media as real people or real events. It's dishonest, and your customers will find out eventually.
And judge your writing on the only test that has ever counted. Is it true, is it useful and does it sound like you? Generic output fails that test whether a machine wrote it or a bored human did. The businesses that keep their reputations through all this will be the ones whose content was so clearly grounded in real expertise and real results that the question never came up.
So before your next piece goes out, read it out loud. Does it sound like you talking? If it doesn't, it won't feel like you to the person reading it either.
