The fear is specific and I hear it in almost every conversation with an owner. If I let anything help me write, my customers will know, and the relationship will be worse for it.
It is a reasonable fear and the research says it is aimed at the wrong target.
People are not detecting the machine
Researchers at Cornell and Stanford ran six experiments with 4,600 participants across 7,600 self-presentation texts, in three settings where trust genuinely matters: job applications, online dating profiles, and Airbnb host profiles. Detection accuracy landed between 50% and 52%.
That is a coin flip. Across thousands of people, in exactly the situations where they had every reason to be careful.
The interesting part is not the accuracy. It is what people used to decide. Readers treated first-person pronouns, contractions, and mentions of family or personal history as evidence of a human author. Those are stylistic habits, not fingerprints, and the researchers demonstrated it by writing AI text that leaned on those cues and having it rated as human more reliably than genuinely human text was.
So the detector people believe they have is a style preference wearing a costume.
What they are actually detecting
Here is the reframe that matters for anyone running a business, whether that is a firm, a clinic, an agency or a coaching practice.
Readers cannot identify a machine. They can absolutely identify writing that could have been addressed to anyone. Every cue in that study points the same direction: specificity reads as human, and its absence reads as suspect.
Which means the risk in your business was never the tool. It is the register. A message that says "I hope this finds you well, I wanted to reconnect and see if there might be an opportunity to explore working together" would have been damaging in 1998. It reads worse now only because there is a new thing to blame it on.
The corollary is uncomfortable and worth sitting with. If your own writing is already generic, it already carries the penalty, and nobody has been telling you.
Why suspicion alone is enough to do harm
There is a second finding that closes the loop and it is the one that should actually shape your behavior.
In an earlier experiment, participants rated Airbnb host profiles that were, without exception, written by real humans. When told the set was mixed and they would have to guess, the profiles they suspected of being AI received lower trustworthiness ratings. The profiles were human. The suspicion did the damage on its own.
That is why "nobody can tell" is not a licence. In an environment where readers assume some proportion of what they receive is machine-written, a message that pattern-matches to machine-written gets punished whether or not it is. You are not defending against detection. You are defending against a vibe, and the only defense against a vibe is being unmistakably specific.
This is an Engage problem more than a writing problem. It decides how much of someone's interest survives the moment they read what you sent.
What specific actually means
Not longer. Not warmer in a general way. Particular.
Something only you would know. Something only they would recognise. The detail from the last conversation that you would have to have been present to remember. A sentence that would be wrong if sent to anybody else on your list.
That is the standard, and it is also, conveniently, the thing that makes a message worth receiving in the first place. The defense against sounding automated and the definition of a good message turn out to be the same object. If your team writes on your behalf, it is also why a tone of voice document earns its keep: it tells everyone what particular sounds like in your business.
It also sets the practical rule for using any tool inside a business. The machine can carry timing, sequencing, retrieval and structure. It cannot supply the particular, because the particular lives in what you noticed about a person. If a system is producing messages that could go to anyone, the problem is not that it is a machine. It is that it has been given nothing to work with. I have written separately about where I let AI near the business and where I do not.
Questions owners ask
Will AI detection software catch me?
Those tools are unreliable in both directions and have been shown to flag human writing as machine-written, particularly from non-native English speakers. Do not organize your business around passing a test that is itself inaccurate. Organize it around whether the message is specific enough to be worth sending.
Should I tell clients if a message was drafted with AI?
That question has a genuinely complicated answer, because research on voluntary disclosure found it lowers trust rather than raising it, across thirteen preregistered experiments. It gets its own treatment rather than a line here. The short version: decide a policy you can state plainly, and make sure being discovered would not embarrass you.
Does this apply to my newsletter and my website too?
More so. Those are read by people who have no relationship with you yet, so there is no accumulated trust to cushion a generic impression. A website is the place where sounding like everybody else costs the most and is noticed the least by its owner.
If people cannot tell, does any of this matter?
Yes, and the Airbnb result is why. People do not need to be right about which messages were machine-written for their suspicion to cost you. Being unmistakably yourself is the only thing that survives that, and it was worth doing anyway.
Sources: M. Jakesch, J. T. Hancock and M. Naaman, Human heuristics for AI-generated language are flawed, PNAS 120(11), 2023, six experiments, 4,600 participants. M. Jakesch et al., AI-Mediated Communication, CHI 2019, reported by Cornell. The disclosure finding is O. Schilke and M. Reimann, The transparency dilemma, OBHDP 188, 2025.