Nobody is requiring business owners to label AI-assisted content. The problem is that most cannot say what happened inside their own.

It was only supposed to take 10 minutes to do in an afternoon. A writing platform integrated a “AI Dection Scanner” on their platform for users to review if each article was written by AI. The next week, AI companies had started embedding markers in the text their tools generate. I wanted a short line to put at the bottom of what I publish. Fifteen words. Something honest and finished.

Here is what I discovered, and it is not what the current online noise suggests. First it’s important to note – I could find no law that requires a business to disclose AI use in its content. However, the exposure for a business is somewhere else entirely. What I also realized is that most business owners cannot actually say what happened inside their own published work, and the people forming opinions about it will not be waiting for a discovery or explanation.

What actually changed

On August 11, Anthropic confirmed that text generated by its newer models carries an invisible, machine-readable watermark, and that generated files carry signed provenance metadata. Google has watermarked its output since 2024.

The reason it happened is worth understanding, because it explains who the rules are actually aimed at. The EU AI Act’s transparency obligations took effect on August 2, 2026. Article 50 requires providers of AI systems to mark their outputs in a machine-readable format so they can be detected as artificially generated. That obligation sits with the companies building the models. It does not sit with a business publishing a blog post.

There is a separate obligation on deployers, and it is narrow. It covers deepfakes and AI-generated text published on matters of public interest without human review.

The rule everyone is bracing for

In the United States, the story circulating is that the FTC has issued sweeping AI disclosure requirements with enforcement already underway. I went looking for it and I can’t find anything.

What exists on the FTC’s site is a proposed policy statement, published in July and open for public comment through the end of that month. Its subject is AI companies distorting the output of their own systems for undisclosed reasons. It is about model makers. It is not binding, and it is not about whether a business labels its blog posts.

A significant amount of what business owners are reading right now describes a legal obligation that does not exist, and much of it is published by companies selling compliance products. If the case for saying something rests on a rule, and the rule turns out to be imaginary, the case collapses. The case has to rest on something else.

What a detection score is worth

In August 2025, the Federal Trade Commission finalized an order against a company then known as Content at Scale AI, now Workado. The allegation was that it had claimed its AI detector could identify AI-written text with 98 percent accuracy across a wide range of generative tools, and that it could not support the claim.

The order required the company to email its own subscribers and tell them, in its own words, that it had made the claim, that the FTC said it lacked proof, and that it had stopped. It defines the products it covers broadly: anything that detects or purports to detect content generated or altered by AI.

One detail in that file has stayed with me. Two public interest organizations asked the Commission to require the company to submit its detector for independent third-party testing. The Commission declined, on the grounds that the prohibition on unsubstantiated claims was sufficient.

Which means detection tools still grade their own accuracy. There is no independent standard behind the number.

So we now have two separate things that people will discuss as though they are one. There are detectors that estimate, from writing patterns, whether AI was involved. And there are watermarks embedded at the moment of generation. The first is a guess with no verification requirement. The second is more reliable but proves less than people assume. A watermark shows that a model processed the text. It does not show who decided what the text would say. Running your own writing through a tool to tighten it can leave the same mark as generating the whole thing from nothing.

The part most owners have not looked at

This is where my afternoon went sideways, and it is the part I think is genuinely underexamined.

Most established businesses do not produce their own content. They outsource it, which is a sound decision and not one I would talk anyone out of. But outsourcing means the content published under your name, and often under your face, was assembled by people whose process you have never inspected.

If a score appears next to something you published, or a client forwards you an article about watermarking and asks what your policy is, the answer has to already exist. Assembling it under pressure is how businesses end up making claims they cannot support.

The questions are not complicated. They are just ones almost nobody has asked out loud.

• Which AI tools touch our content, at what stage, and who decides that?
• What information about our business goes into those tools, and is any of it confidential?
• Are the tools configured so what we submit is not used to train the underlying models?
• Does a person review the work before it reaches us, and again before it publishes?
• Can the people producing our content put all of this in writing?

That last one is the real test. Anyone doing this carefully will answer without hesitation, because they already know. I put my own answers on a page rather than in an email, because a policy that lives in an inbox is not a policy.

The reader is not going to do this homework

Here is the thing that decides how much of these matters.

Your buyer will never read the Workado order. They will not know that detectors self-certify, or that the FTC declined to require independent testing, or that a watermark records processing rather than authorship. They are not going to sort through what the EU actually requires and of whom.

They arrive with a feeling about AI, formed somewhere else, and they apply it to what is in front of them. You cannot argue with a feeling that formed before you entered the conversation. You can only have already answered.

That is the whole argument for saying something now. Not compliance. Not fear of a rule that has not been written. Sequencing. A position published before anyone asks is a policy. The same words published after someone asks are a defense, and readers can tell the difference without being able to explain why.

What this actually takes

Two or three sentences. What AI does in your process, what it does not decide, and who is accountable for the result. It should be true, it should be specific enough to be useful, and it should avoid any absolute claim you cannot enforce across everyone who works on your behalf.

A promise that no AI touched anything is one staff member away from being false.

The work is not the writing. The work is finding out what is actually true first, which for most business owners means a conversation with whoever produces their content.

I thought that conversation would take five minutes in an afternoon. It did not. But I would rather have had it in August, on my own schedule, than in whatever month someone finally asks.