AI Content Disclosure: Where Teams Draw the Line, and What Tested It

Most contributors quoted here neither label everything AI touches nor label nothing. The lines they hold are narrower and more practical: say so when AI did work a reader would assume a person did, never let AI put words in a real person's mouth, and never let it deny being AI. What tested those lines was rarely a debate about policy. It was a specific sentence that turned out not to be true, or a client, platform or editor asking a direct question.

Pattern 1: The line is what a reader would assume

The clearest rules were not about tools. They were about what the reader takes for granted. Luke Wilcox, founder of Roiva, discloses AI "when it did work a reader would assume a person did", and names who did each step in any research his company publishes. A draft tested that:

In October a draft research brief said some checks were done "by hand" and by "a second tester", but both were an AI agent. A second AI drafting pass flagged it before anything went out, and we fixed it the same day. Nobody meant to mislead, but a reporter would have asked, and the answer would have looked bad. After that the line moved from a call I made per piece to a written rule for every study.

Conor Smart, founder and director of Arklavo, found the same problem in a customer story:

Recently, I generated a blog post featuring a customer's experience with our products. The AI smoothed out the customer's rough quotes to make the story flow better. I liked how it turned out so much that I almost published it as-is.

Then I realized that the "quotes" weren't quotes at all, they were just what the AI thought my customer probably said. I scrapped the whole thing.

He added one rule to his policy: "quotes must be copied & pasted."

For Mohammad Hayat, Sr. AI Engineer at PersonaXpand, the reader's assumption is the product itself. PersonaXpand drafts social media replies in a customer's own voice, and the rule is that the AI never pretends to be a person:

That rule was tested on 9 November. In our own inbox, someone asked "be honest, is a bot replying to these?" and the suggested reply said "Haha, definitely not a bot! Just me here." Nothing had been sent, but every customer account could have produced the same answer. We changed the instructions the same day so the AI can never deny being AI, added an automated test that fails if that rule ever disappears, and checked it again on a live account.

Pattern 2: The test was a false sentence, not a missing label

Several contributors had a disclosure line in place and still published something untrue. In these accounts the label was never the weak point. The checking was.

Jens Hagel, founder and owner of kjello, lets AI write first drafts for his IT firm's blog on the condition that he reads every piece and checks every date and figure:

The test came from a piece that skipped that step. In August 2026 an automated LinkedIn post on our company page described the Windows 10 end of support, which happened in October 2025, as still ahead, and a business partner asked publicly whether we were a year late. The prompts simply didn't know today's date. We corrected 182 passages in 41 articles, and every deadline is now written with its exact date, never as "soon".

Mücahit Kaya, founder and editor-in-chief of AI Tools Police, publishes a page that lists what AI helps with and what it is never used for, including fabricating tests. His publication does not run hands-on product trials, and says so:

In July 2026 a site-wide audit found that our copy, much of it AI-assisted at the drafting stage, had picked up the review genre's default language: four claims that we had tested tools, and "Tested & Ranked" in 17 page titles and preview images. None of it was true, and every instance contradicted our own published promise.

The titles were changed to "Compared & Ranked", and the built site is now scanned for testing language before anything ships. As he put it: "Disclosure is not the page where you say how you use AI. It is every sentence that has to stay true to that page."

Pattern 3: Someone else drew the line first

For other contributors the test came from outside: a platform, a client, an editor or a detection tool made the call before they did.

Aviad Gershon, founder of GrowingUpVideo, posts AI footage of real people growing up from their old photos:

YouTube is what tested our line. When we started, we weren't sure what that altered-content box was meant to cover, so I uploaded some of these videos without ticking it. YouTube added the AI label to them anyway. After that I stopped treating it as a judgment call: if the platform will label it regardless, leaving the box unticked only makes us look like we were hiding it.

Mihai Cirstea, CEO of Site Pixel Media, was asked by a client who had noticed the work change:

Last year, a B2B client asked why their blog posts suddenly got better and more consistent. They figured we'd hired more writers and wanted to meet them. We had a choice to make. We told them we'd started using AI tools to speed up production without lowering our standards. They were fine with the honesty, but they wanted sign-off on which posts were AI-drafted versus written by us.

The agency now tags AI-drafted pieces internally and explains its process during onboarding.

Heath Squier, founder of EVKII, marks his team's written pitches "Prepared with AI assistance". A journalist's request put a cost on that:

A social-media discoverability brief asked for two short answers and expressly rejected AI answers. The topic fit our marketing work, but our AI-assisted writing process did not fit that condition. We did not submit an answer. We also did not remove the disclosure or reword the text to try to pass a detector.

"That did not require moving the line. It required applying the rule when it cost us an otherwise relevant opportunity."

Jordan Bridger, founder of Human X, changed how he writes after a tool gave its own verdict:

What tested the line was discovering that AI-drafted articles of mine scored 100% AI on detection tools. That moved the line: now I dictate in my own voice first and let AI do light cleanup, never the other way round. The disclosure question stopped being theoretical the moment a machine could tell the difference.

The counter-view: disclose nothing, answer honestly

Not everyone discloses. Roy Danino, co-founder of Lachi Media, treats AI the way his team treats a spreadsheet:

I don't tell readers or clients when AI helps make something, because AI is a tool, not a co-author. Nobody asks what software wrote the invoice.

His line still got tested, and it held because he answered the question straight:

A home services client asked me point blank last year if a person or a machine wrote their ads. I told him the truth. Our team writes the strategy and the offer. Tools help us test 20 headline versions instead of 3. He was fine with it. He cared about booked jobs, not word counts.

The through-line

The contributors disagree on whether to label, and agree on what a label cannot do. A disclosure did not stop a wrong date, an invented quote or a claim of testing that never happened. What stopped those was a named person checking the finished piece, and in two of these accounts an automated check added afterward. The two contributors whose clients asked directly both told the truth, and both report the client was fine with the answer.