AI: The Reputation Signal · Issue 6
Em dash and proud
Elif Güvençer ·
Sixteen law professors sat down to grade short answers to contract-law questions. They graded blind. Nobody told them which answers came from a colleague and which from a large language model.
Professors rated LLMs far higher than their peers (average win rate = 75.33%), with models performing similarly to the best instructor. LLM responses were also rarely flagged as harmful (3.53%, vs 12.06% for professors).
Within the AI conversation, judgement is what we, as humans, say will remain ours, and if I side with the AI doom narrative, it is our advantage over a synthetic rival that is coming for our intelligence, our capabilities, our jobs…Law is a discipline based on judgement, and the machine outperformed the humans in this test.
The projection for Communications should be clear to you by now: communication is a discipline of judgement, second-order risk thinking. What happens to this moat the function relies on so much when, in a blind test, the machine wins? This uncomfortable question is parked here for another post.
On this one, I will double-click on the moment we find out it was the machine that produced the work.
AI label has a disclosure penalty
According to a study by Columbia Business School, art labelled "AI-generated" was valued 62% lower than identical art labelled human-made — and the same penalty hit work labelled "human + AI collaboration."
In a separate study spanning 16 experiments, readers rated the same poems and stories lower when told that a machine had written them.
These are examples where the label takes away from the (perceived) value of the work.
AI disclosure also takes away from the (perceived) trust in the person who has done the work, with AI.
In a series of thirteen experiments, people who disclosed that they'd used AI were trusted less than people who didn't — and the effect was remarkably stubborn. Clients trusted a designer 20% less after the designer came clean. It held even among people who use AI every single day.
The researchers traced it to a quiet judgment about legitimacy: admit the machine helped, and you're read as having cut a corner.
We have two penalties, both fired by the same trigger. Reveal AI in the work, and the work loses its worth. Reveal AI in yourself, and you lose your standing.
We have been creating mediocre work with our human brains long before AI has come along. AI slop – the low-quality work AI produces- was and is sitting alongside human slop. Paul Smith has an interesting perspective on this.
The standard we'd never survive
Most complaint about AI insinuates a comparison to a flawless human who has never existed. It's biased. It flattens your thinking. It gets things confidently wrong. It takes away your creativity. All true. But hold it against the actual alternative — a person.
Imagine you are writing a book and working with an editor. Nobody would judge you because you work with an editor. Let's assume your editor is human. Someone with a life, problems of their own, their own traumas, biases.
You sent them your draft. Maybe your editor reads your draft through a fight he had with his wife that morning, and is not in a good mood AT ALL.
You cannot tell him to regulate it out of him and look at your copy more objectively. Even if he tells himself that, chances are he can't.
We are complex, emotional beings, and so much of what we do, despite the best efforts of psychology, sits with our subconscious. Human bias is real and difficult to identify and train our brains out of.
AI's biases are also real. The difference is that they are more inspectable. You might not be able to identify the exact root cause of your own decision — and, to be fair, the model can't fully explain itself either. But its behaviour is testable in a way yours isn't: you can run it a thousand times, see exactly where it skews, put guardrails around it, tell it what to ignore, and then check whether that actually worked. Not perfectly — some of what a model absorbs in training you can't fully prompt out for example. But it's measurable, repeatable, and partly steerable.
A clarification here before you throw stones at me for seeming to diss AI's shortcomings. I am not. I am not underestimating the consequences of unchecked AI-generated information running wild in the public domain, cognitive offloading, and other undesirable impacts/risks of AI either.
I am also not ignoring the negative impact it has on the job market, especially creative work.
What I am arguing is that we should be less quick to condemn AI's biases and less forgetful of our own.
The person who is AI-shaming you is using AI
One of my favourite quotes attributed to Rumi (the 13th-century Persian poet and mystic) is: Either appear as you are, or be as you appear.
This call to authenticity should ring true for all of us. People who are harshly judging others for using AI are also using AI.
Do you use auto-correct? You are using AI. Does that mean the sentence Grammarly auto-corrected is not yours? Where is the line?
Needless to say, you should not pass off something you did not make as your own. There need to be ethical boundaries, but is that line cut-and-dried? And is it being drawn out of ethics or fear?
The negative posture toward AI-generated writing has hardened into a tic. The em dash — a perfectly good 250-year-old piece of punctuation — has been recast as a villain. People now strip them from their own writing, deliberately altering their copies to avoid being seen as using AI.
In November 2025, after a steady drumbeat of complaints that the em dash was a dead giveaway, OpenAI's Sam Altman said ChatGPT would finally obey an instruction to stop using them. So the bias even reached back and changed the tool.
Social media is full of AI detectors that don't even work properly — one flagged the Declaration of Independence as 98% AI-generated — and people are using them to prove their work is not AI-generated.
The legal pressure will make it worse
On the 2nd of August 2026, the EU AI Act's transparency rules take effect. Text generated by AI and published to inform the public on matters of public interest will have to be disclosed. Deepfakes, flagged. Synthetic outputs, marked so a machine can detect them.
We are about to require disclosure. Except we punish disclosure when it comes.
The challenging thing is likely because I am surrounded by this increasing stigma around AI usage and fear-based avoidance behaviours, I also find myself doing the same thing. The small voice in the back of my mind says, If AI has done it, is it really that good?
We started to see a divide between people who use AI in their work and those who don't, with the latter wearing it as a badge of honour.
So this bias is being imprinted on us from all sides; we are being conditioned to judge work done with AI and to doubt the ability and character of the person using it. I don't think this bias is going away. I think it's something we're going to have to learn to work with, the way we've tried to work with the others — unconscious-bias training and the rest, with mixed results. Perhaps that's coming next: not training to spot AI-generated work, but training to notice what the word "AI-generated" does to us.
So here we are: The tools get sanded down to hide their own fingerprints, to act and sound less like AI, and we all keep deleting our em dashes in silent shame, hoping nobody notices we ever used them.
This is a catch-22 for Communications. Content creation can be automated and is useful for Comms when done correctly. This efficiency is also required in a discipline that is increasingly being redefined by AI. So the function is pressured to use it and is judged for doing so.
Em dash and proud.
Since this is a piece about disclosure, here's mine about my own writing.
I use AI mostly as an interviewer and then an editor. I ask AI to interview me on a topic, help me flesh out my thoughts, and stress-test ideas. Help me define the skeleton of what I want to write. It can also suggest angles based on my past writing and what it knows about me.
I use it to tear my claims apart, poke holes in my thinking before anyone else does. I have a reviewer set of agents. To name a few, a CCO that reviews whether what I write resonates. I have an AI expert who ensures I remain technically accurate.
I am a trilingual, non-native English speaker. I use AI to ensure my writing is grammatically correct and reads well, and I believe it benefits my writing.
I use AI for image and video creation.
And yes — I know the costs. It can dull my thinking. The intellectual struggle I am experiencing while writing might be short-lived and therefore cost me creativity, originality, and more. I take them seriously and take precautions.
But I don't light my office with candles because the white light of electricity is harsh on the eyes. I wear glasses to cut the glare and keep the light. That's roughly how I treat AI: eyes open to the cost, keen on reaping the gains.
Em dash, and proud.
The last thing I leave you with is a small test: now that you know AI has been involved in one form or another in this newsletter, did it take away from the reading experience? Do you find yourself silently underweighting my ideas, criticising sentences in hindsight? If so, that might be the bias doing its work! First move towards working with it is awareness…