Can People Tell Your Content Was Written With AI? They Can Now.

The choice to disclose AI use in your company just changed in a way no one expected. This summer, AI detection went from a tool with a false-positive problem to platform infrastructure. Which changes the way we think about using AI in our day-to-day work. Companies like Substack now let any reader scan a post for AI involvement, a browser extension does the same across LinkedIn and the open web, and the labs themselves have started watermarking their output at the model level. A few weeks ago we wrote about the disclosure tax, the price your people pay for admitting AI helped with their work. That piece assumed admitting was a choice and now there is no hiding if you are using the technology in your work.

What changed in AI detection this summer?

Detection is now a layer built into the places your content lives, and readers can invoke it without asking anyone's permission.

For years the correct position on AI text detectors was distrust. They flagged human writing as machine-made, missed obvious slop, and got laughed out of serious conversations. That era ended faster than most content teams noticed. On July 21, Substack switched on an integration with Pangram, the current leader in detection accuracy, letting readers scan any post, note, or comment over 100 words and see an estimate of how much was human-written, AI-assisted, or fully generated. Pangram's own study of more than a million social posts found that over 40 percent of long-form LinkedIn content flagged as fully AI-generated, the highest share of any platform it measured. And the same company ships a Chrome extension that scans feeds as you browse, which means detection no longer requires the platform's cooperation at all. The reader brings their own scanner.

Think about that last part. Your prospect reading your company's thought leadership may already have a verdict rendered in the corner of their screen before they finish your first paragraph.

What does watermarking change?

It moves origin from something a classifier guesses to something the text carries with it.

Classifier detection is a statistical argument. Watermarking is a receipt. Starting August 2, Anthropic began embedding imperceptible watermarks into text generated by its newest Claude models, worldwide, to comply with the EU AI Act's transparency rules. The mark survives copy and paste, may survive light editing, and applies at the model level across every surface the model touches, including the API and cloud partners. A detection API is on the way. Anthropic is one of roughly 190 signatories to the EU's code, alongside OpenAI, Google, Meta, Microsoft, and Mistral, so treat this as the industry's direction of travel rather than one lab's compliance posture. The regulation is a reported fact; whether it's good policy is not your problem. What the marks do to your content pipeline is.

Two caveats matter for anyone making decisions off this. The marks are probabilistic: short or heavily edited text may carry no detectable signal, and text a model merely proofread may carry one even though a human wrote every idea in it. And absence proves nothing, since older models and determined evasion still produce unmarked text. The systems are estimates. The labels they produce will not be read as estimates. A reader who sees "likely AI-generated" on your founder's essay does not file that under statistical nuance.

Is the answer to stop using AI in your content?

No. The answer is to decide where authorship is part of the product before a label decides for you.

Here's the shift underneath all of this, and it's the reason the disclosure tax piece needed a sequel. Disclosure used to be a decision. Now it's a property of the text. We call this the disclosure flip: the point where "did AI write this" stops being a question you answer and becomes a question your content answers about itself, whether you're in the room or not. Inside your company, the disclosure tax punished people for admitting AI use. Outside it, the flip removes the admitting step entirely. Both problems have the same root: organizations that never decided, on purpose, where AI authorship is acceptable and where it breaks a promise.

Because that's what actually gets damaged when a label surprises a reader. The label itself is survivable. A support macro flagging as AI-generated costs you nothing, since nobody believed a human artisan wrote their password-reset instructions. The damage happens where the flag contradicts an authorship promise you made, explicitly or implicitly. A bylined executive essay. An investor letter. A condolence note from HR. A "personal" founder post that scans as 100 percent machine. In those cases the reader was paying, in money or trust, for contact with a specific human mind, and the label tells them they didn't get it.

Where does provenance actually matter in your content stack?

Anywhere the reader believes a specific person is talking to them; almost nowhere else.

So the move for leaders is a sorting exercise, and you should run it before your platforms run it for you. Take an inventory of everything your company publishes: bylined essays, sales sequences, support content, product copy, social posts, executive communications, recruiting outreach. For each one, name the authorship promise it makes. Some content promises nothing about who wrote it, and AI can produce all of it you want. Some content carries a person's name and trades on their judgment, and that's where a surprise label costs real trust. The high-provenance tier probably needs a human doing the thinking and most of the writing, with a paper trail that matches the claim. The low-provenance tier needs speed and accuracy, and its origin is nobody's business. The expensive failure is publishing high-promise content through a low-promise pipeline and letting a scanner announce the mismatch to your audience.

We believe that you should build content and communication systems for clients with this boundary designed in from the start, deciding at the architecture level which outputs carry a human's name and which carry the company's, because retrofitting provenance after a public flag is reputation repair, and repair always costs more than design.

What should you do about it?

Identify what content makes an authorship promise you can't afford to have contradicted in public. Sort your content stack by the promise each piece makes and set an explicit AI policy per tier. Then make sure the bylines your leaders sign sit on work they'd defend as their own. The flip already happened. The companies that come through it clean will be the ones whose labels, whenever a reader checks, match the promise on the package.

If you want that boundary designed into your content and AI systems rather than bolted on after a bad scan, learn more about our AI Blueprint approach or reach us at contact@theyor.com.