Your AI Text Is Watermarked Now. Your Images Can Be Checked for Free.

If you're forwarding AI-written text or an AI-touched image to a client this week, two tools that shipped days apart just changed what "checking it first" means — and made it much easier to actually do.

What happened

On October 6, OpenAI started embedding an invisible watermark called textGrain in all ChatGPT and Codex text output. It's mandatory wherever the EU AI Act's transparency rule applies; outside the EU it's an opt-in for API developers, off by default, and detection is currently limited to approved researchers. A day later, Google DeepMind opened SynthID Detector to the public worldwide — a free tool that checks whether an image, video, or audio file was AI-generated, covering output from Google's own models as well as OpenAI, NVIDIA, and Kakao, with Apple reportedly joining. DeepMind says it's already processing more than a million checks a day. On October 8, Anthropic rewrote its usage policy to explicitly ban fake accounts, astroturfing, and sockpuppet networks built with Claude, after finding state media and commercial operators using it for influence campaigns.

What to do differently Monday morning

None of these three things does the checking for you — they make the checking possible, which is different. If you work in or sell into the EU, ChatGPT-drafted text already carries an identifier you can't remove with a light edit. If a client, vendor, or candidate sends you an image or video, SynthID Detector can now tell you whether it was AI-made in under a minute, for free. And if your team uses Claude, "we didn't know fake engagement was against policy" stopped being a defense on October 8. The decision this week isn't whether to check — it's where in your process the check happens: before you publish, before you pay a contractor, before a deliverable goes up the chain.

Take a marketing manager at a 40-person DTC skincare brand who outsources product photography touch-ups to a freelance retoucher. Two of the last five "photographed" hero images he delivered were actually generated from scratch, not retouched — a contract violation she'd never have caught by eye, since the brief only asked her to review the final look, not its origin. Running each delivered image through SynthID Detector before it enters an ad campaign costs her two minutes and tells her which ones to query before they're live and attributed to a real shoot.

A communications director at a regional credit union runs a different version of the same risk. Her team drafts member-facing FAQs and social replies with ChatGPT, and until this week nobody had asked whether that text carried anything identifying. Whether or not textGrain applies to her US-based union matters less than the Anthropic change: a social agency she'd contracted to "boost engagement" had been buying automated replies, and that's now an explicit policy violation she can cite in the contract instead of an argument about intent.

The detection tools and the policies against misusing AI output both got sharper teeth this week. The only variable left is whether your workflow has a point where someone actually uses them.