09 Sep Your AI writing has a tell, and your readers are learning it.

Two years ago you could publish anything a language model wrote and nobody would bat an eyelid. Now a single em dash in a LinkedIn post and everyone reading instantly thinks you didn’t write it.

The em dash has been used for ages, but not quite this frequently. It became a tell because the models loved it, people noticed and the association stuck. Writers who have used em dashes their whole careers are now editing them out of their own work to avoid the accusation.

That is the interesting part. The punctuation was just the most obvious part.

The second tell is becoming more obvious now.

I started noticing the next one a few weeks ago, in other people’s copy, then in my own.

It is the construction that qualifies every positive statement with a negative. “We build products, not websites.” “This is a partnership, not a vendor relationship.” “It is a content build, not a marketing one.” The sentence names the thing, then names its shadow, and the pairing creates a feeling of judgement having taken place.

I went looking at a proposal that had just come off my desk. Twenty nine slides. Eighty one instances. Roughly two and a half per slide, in a document I had read three times and thought was clean.

Once you see it you cannot stop seeing it. It is in product pages, LinkedIn posts, agency about-pages and half the newsletters in my inbox. It will get named publicly within the year, and then it will carry the same stink the em dash carries now.

Why models write like this

Both tells come from the same place and understanding this is important.

A language model is trained to produce text that people rate highly. Text that reads as considered, balanced and authoritative gets rated highly. So the model learns the surface features of those qualities and applies them everywhere.

An em dash performs deliberation. It suggests a writer pausing mid-thought to add a qualification they judged necessary. The antithesis construction performs decisiveness. “A hub, not a supermarket” sounds like two options were weighed and one was rejected. In truth the thing has been named twice, once directly and once by its opposite, there was really no deliberationl. 

The rule of three does the same job. “Faster, cleaner and more reliable.” Three items sound complete and considered. Two sound thin. Four sound like a list. So everything arrives in threes, regardless of how many things there actually are.

There are quieter ones. The hedged opener that begins with “it is worth noting that”, so the sentence arrives pre-softened before anyone has objected to it. The closing paragraph that summarises what you have just finished reading, because thoroughness is easy to imitate in that way. The perfectly even paragraph lengths, four lines each, all the way down the page, because nothing in the training data ever rewarded a one line paragraph that lands hard.

These are performances of thinking. That is why they irritate people who cannot yet articulate why. The reader senses effort being signalled where no effort was spent.

What it costs a brand

Here is the commercial version of the argument, and it is the reason I care about this beyond craft.

A reader who clocks one tell does not simply discount that sentence. The suspicion runs backwards over everything they have already read, and forwards over everything that follows. Your case studies, your pricing page, your founder’s note about why the company exists. All of it gets re-read from a more suspicious perspective. 

For a business that has spent real money on being distinctive, this is expensive. You pay a brand studio to find out what makes you different from six competitors selling something similar. You pay a photographer so the imagery looks like you and nobody else. Then the copy that carries all of it is generated by a system whose entire design goal is to produce the most broadly acceptable output available.

Being average is the safe answer for a model. It is a terrible answer for a brand.

Picture the worst case scenario (and I tend to have this recurring nightmare whenever I utilise the robots) : A prospect opens your proposal, hits a sentence that sounds like something they generated themselves and quietly labels your proposal as untrustworthy. The real danger here is that nobody says a word, it is just an idea that sticks from that moment onwards.

What actually works

I use these tools every day and my output would be slower and thinner without them. The argument here is about how much original thought is in your words. Are you reading through it as an editor would? 

Write down your voice as rules, and keep the file. Real constraints that a stranger could follow. Which words are banned. Which spelling convention. Whether headings are sentence case. Whether exclamation marks are allowed in client-facing copy. Mine bans sixteen specific words, including “unlock”, “elevate”, “seamless” and “leverage” as a verb. It also fixes South African spelling, so nothing comes back with an American z in it. It also has any of these lazy writing hacks that make the sentences formulaic.

Ban the tells by name. A model will happily comply with “never use em dashes” once you tell it. The same goes for the antithesis construction, once you describe it clearly enough that it can recognise its own output. Vague instructions like “sound more human” achieve nothing, because the model already believes it sounds human.

Give it something only you know. The strongest defence against generic output is specific input. A real number, a client’s actual objection, the thing that went wrong on the last project, the reason you changed your process in March. A model cannot invent your evidence. It can only style whatever you give it, so give it more.

Read it aloud. Every tell I have described is audible. Antithesis has a see-saw rhythm you can hear. Threes have a cadence. If a sentence sounds like it is being delivered at a conference, cut it and say the plain version.Human editing is important because context is everything!

Keep at least one or two passes that are entirely yours. This is separate from proofreading. It is the pass where you argue with the draft, take a position it was too polite to take, and put the sentence in that only you would write. It is so important that your own voice still shines through the writing! 

The tells will keep moving

Whatever replaces the antithesis construction is already in circulation and already invisible. That is how this works. A pattern spreads, saturates, gets named, and becomes a liability. Then the next one starts.

Chasing individual tells is a losing game. The defence is a voice with a viewpoint that stays recognisable while everyone else’s copy converges. That means a lot of correcting, applying it on every job, and reserving the last pass for a human with an opinion.

The tool is a good drafting partner and a fast one. It is still up to HUMANS to have opinions and viewpoints in this world full of analogue context, endless paths that have generations of depth and meaning. Let’s not forget what it makes us human!