6 min read
When OpenAI tested its new invisible watermark for ChatGPT text, it caught the AI writing about 92% of the time. Then the testers swapped out a quarter of the words in a 400-token passage (roughly 300 words), and the detection rate fell to 17%. Those two numbers, reported by PPC Land from OpenAI’s own figures, explain almost everything a small business owner needs to know about the news.
Here is the short version. On October 5, OpenAI announced that text from ChatGPT and Codex would begin carrying an invisible watermark for users in the European Union, on every plan, rolling out over the following weeks. The reason is Article 50 of the EU AI Act, which requires providers of AI systems that generate text, images, audio, or video to mark their output in a machine-readable way “where technically feasible.” Outside the EU, ChatGPT users see no change, and in the API (the developer connection that other apps use to tap into OpenAI’s models) watermarking stays off unless a customer opts in, according to ITdaily.
A watermark here is not a visible stamp. It is a statistical pattern: the model nudges its word choices in a way that a special detector can recognize later, while a human reader notices nothing. That subtlety is exactly why the topic is attracting so many confident and wrong takes. Let’s test five of them.
Myth 1: “Clients can now paste my copy into a checker and catch me using ChatGPT”
Verdict: false, for now. OpenAI has not released its detector to the public. At launch, access is limited to approved researchers and expert organizations. Your client cannot open a website, paste your newsletter, and get a yes or no from OpenAI’s watermark system.
That could change. Regulators may eventually want public verification tools, and other companies may build their own. But today, the watermark is a compliance mechanism aimed at researchers and authorities, not a client-side lie detector.
Myth 2: “I’m in the US, so this has nothing to do with me”
Verdict: mostly true for your own writing, but not the whole story. If you write in ChatGPT from Texas or Ontario, your text is not being watermarked under this rollout. Still, three situations can touch a solo business that never leaves North America:
- Your contractors may be in the EU. A virtual assistant in Portugal or a freelance copywriter in Germany who drafts with ChatGPT will hand you watermarked text.
- The apps you use may opt in. Because the API setting is a choice for each business that builds on OpenAI, a writing tool you subscribe to could turn watermarking on for its own reasons.
- Rules travel. Several US states are writing their own AI disclosure laws. Our plain-English guide to what the new state AI laws mean for a solo business covers where that stands.
The practical takeaway: you don’t need to change anything today, but you should stop assuming AI text is permanently untraceable.
Myth 3: “A quick edit scrubs the watermark, so the whole thing is pointless”
Verdict: half true. The 92% to 17% drop is real, and it comes from OpenAI’s own testing. Rewrite a quarter of the words and the signal mostly disappears. Translation and heavy paraphrasing likely weaken it further.
But “easy to remove” is not the same as “pointless.” The watermark is designed to catch raw, untouched output at scale, such as spam farms, fake reviews, and mass-produced propaganda. It was never meant to police a bakery owner who drafts a newsletter with AI and then rewrites half of it.
And here is the quiet lesson for small businesses: if your content would survive a 25% rewrite unchanged in meaning, you should be doing that rewrite anyway. Raw AI text tends to sound like everyone else’s raw AI text. The edit that adds your examples, your prices, your local details, and your opinions is the same edit that makes your writing worth reading.
Myth 4: “AI detectors can already prove who wrote something”
Verdict: false, and OpenAI says so itself. In its announcement, the company was blunt: “The absence of a detected watermark does not prove human authorship.” A clean result tells you nothing. A positive result tells you that, at some point, a watermarking model produced some of those words.
Third-party AI detectors (tools that guess from writing style rather than reading a watermark) have an even shakier record. OpenAI shut down its own AI text classifier in 2023 because of its low accuracy, and freelancers have been wrongly accused by clients who trusted detector scores. If a client ever confronts you with a detector percentage, you are on solid ground asking what tool produced it and how it handles false positives, meaning human writing that gets flagged as AI.
The flip side: don’t use these tools to judge your own contractors either. Judge their work on accuracy, voice, and results.
Myth 5: “The safest move is to never admit I use AI”
Verdict: false, and getting more false. Hiding AI use worked when nobody asked. Clients now ask. Contracts now include clauses about it. Regulators in the EU and in several US states are writing disclosure rules. A business that has been quietly passing off AI drafts as fully handmade work is building a small reputational time bomb.
The safer move is boring and honest: decide what you use AI for, write it down, and tell clients in a sentence. Most clients care about two things, that the facts are right and that their confidential information stays confidential. Address those and the AI question usually ends there.
If you have not written that policy yet, our guide to writing a one-page AI policy takes about an hour.
What a sensible solo business does this month
None of this calls for panic, but it is a good prompt to tidy up three things:
- Add a disclosure line where it matters. Something like: “We use AI tools to help draft and edit. Every piece is reviewed, fact-checked, and finalized by a person.” Put it on your services page or in your proposal template.
- Check your client contracts. If you deliver writing, design, or translation, add a clause stating whether AI tools may be used and who reviews the output. This protects you when a client later claims they were promised “100% human” work. Ownership questions sit right next to this, and our post on who owns the AI content your business publishes walks through them.
- Edit like you mean it. Treat every AI draft as a first draft. Add specifics only you know. Read it out loud. If it sounds like a brochure for nobody in particular, it is not done.
If you work with EU clients or contractors, also ask your contractors which AI tools they use and whether they disclose it to you. It is a two-minute conversation that avoids a much longer one later.
The bottom line on ChatGPT watermarks
The EU watermark is real, limited, and fragile. It will not let your clients catch you, it does not reach US users today, and a determined editor can wash it out. What it does signal is the direction of travel: AI-generated text is moving from invisible to labeled, one regulation at a time. The businesses that will be comfortable in that world are the ones that already use AI openly, edit it heavily, and can explain their process in one sentence.
So here is a question worth answering before your next client asks it: if someone wanted to know exactly how AI fits into your work, could you tell them in thirty seconds?
Related reading
- Only 23% of Small Businesses Have an AI Policy. 9 Quick Wins to Write Yours on One Page
- 5 Myths About Who Owns the AI Content Your Small Business Publishes
- What Do the New State AI Laws Mean for a Solo Business? A Plain-English Starter Guide
- AI Translation Is Good Enough for Your Website’s First Draft. Never Publish It Unreviewed
