Readers mark down anything that sounds machine-written
Customers trust copy less once they suspect AI wrote it, even when a person did. Generic output now has a sales cost. What helps, and what does not.
3 min read Provecta Group
Last month LinkedIn started letting members flag a post as looking like AI slop, according to reporting in Tubefilter on 10 August. A social network now has a button for the problem, which tells you how familiar it has become.
The button is the visible part. The quieter part showed up a year earlier, in a study by the ad network Raptive. It surveyed 3,000 US adults in July 2025. When people believed a piece of content was written by AI, they rated it roughly 48% less trustworthy, and their stated likelihood of buying the product in it fell by 14%. The detail worth reading twice: the penalty applied even when a person had written the piece. Suspicion alone was enough.
For a firm that sells through proposals, emails and a website, that changes the arithmetic. AI-drafted copy saves the writer time. If it reads as AI-drafted, it costs the reader's trust, and the reader is the customer.
Gartner raised the same concern from the marketer's side in March 2025. Most marketing teams were exploring generative AI, fewer than half saw a significant benefit, and Gartner warned that output was drifting away from brand voice.
Why AI copy sounds like AI copy
Partly it is surface habits: certain punctuation, certain words, a fondness for groups of three and for sentences that set up a contrast and then resolve it. Readers have learned these, rightly or wrongly. Partly it is shape: every paragraph the same length, every email ending on the same upbeat line, every proposal opening by restating the client's question.
Mostly it is emptiness. A model with nothing specific to say about your customer says something general about customers in your sector, smoothly. People recognise smooth and general faster than they recognise any punctuation mark.
What a written house memory can do
We are not neutral on this. Our company memory holds a written communication standard that our AI sessions read before they draft: how to adapt tone to the person reading, what never appears in public material, how our brand names are spelled, no emojis in anything we design. It also holds facts, including what we may say about our work and what we may not.
For a firm starting out, two kinds of rule do most of the work.
The first is a banned list, written down with the reason each item is on it. A banned list is crude, and it is useful because it removes the habits readers spot first. It needs to sit in the shared memory rather than in one person's prompt, or it applies only when that person is drafting.
The second is real material. Approved examples of our own past writing. Verified facts about the client and the job. A model with three good examples and the actual details of this customer writes something only this firm could have written. A model with "write in a friendly professional tone" writes what everyone else's model writes.
Then a person reads it before it goes out, and the question they ask is simple: could a competitor send this exact paragraph with their name on it? If yes, it goes back.
Where we are less sure
Rules about surface habits age quickly. The punctuation readers associate with AI this year may be different next year, and a model trained to avoid one tell picks up another. A banned list needs owning and revising, or it becomes its own kind of template.
And no memory supplies an opinion. The copy that reads most clearly as human says something the writer actually thinks, about something they actually saw. That still has to come from a person who has looked at the customer's problem. We do not know of a way to automate that part, and we are wary of anyone who says they have.
There is a fair objection to all of this: readers are guessing, and they often guess wrong. Raptive's own finding, that human writing is penalised when it is suspected of being AI, says as much. It would be convenient to dismiss the whole thing as prejudice. It does not matter much whether the reader is right. They decide whether to trust you on the strength of the guess.
Try it on your last five emails
Your team's last five emails to prospects are enough for this. Highlight every sentence a competitor could send unchanged. If more than half of each email is highlighted, your problem is less about AI than about specifics, and the fix starts with writing down the specifics you have.
Book 20 minutes at pgco.world/book.