As we went around the room at a recent AI workshop I was facilitating, one participant told us that online arguments are now getting dismissed the moment they smell machine-written. The point may be sound, but once the response feels like somebody’s AI debating on their behalf, the other person stops engaging. It made sense to me. What surprised me was the research published soon after: across a series of experiments, readers rated AI-generated stories higher for quality and absorption, yet still gave an advantage to stories they believed were written by a human.
We appear to want the benefits of AI writing while distrusting the moment its presence becomes obvious. That contradiction creates an Authenticity Tax. This article looks at why AI writing can be genuinely pleasant to read, where the backlash against it is justified, where “this sounds like AI” becomes a shortcut around a good argument, and how to use AI without letting its default voice replace your judgement, experience or authorship.
Why AI writing can read well
Participants rated the AI-generated stories higher for quality and absorption, while stories they believed came from a person received an authorship advantage. In two follow-up studies, participants performed no better than chance at telling the stories apart. That separates readability from valued authorship without proving AI is broadly better at storytelling.
Long before generative AI, I used ProWritingAid to watch the Flesch Reading Ease score, aim for roughly a Grade 8 reading level, remove sticky sentences and fix run-ons. My assumption, which the researchers did not test, is that AI now produces many of those human-pleasing patterns by default. AI writing can simply be good writing.
The Authenticity Tax
The Authenticity Tax is the discount readers apply when they suspect the named writer did not meaningfully shape, review or stand behind the work. It appears when readers cannot see enough evidence that a person chose the argument, supplied the experience, tested the reasoning and accepts responsibility for the result.
The cost can show up as cheapened perceived value, then less engagement and sharing, reduced platform reach, reputational erosion and less trust in future work.
If your professional value comes from judgement, the threshold is higher. I am often introduced to audiences as the “AI expert.” Not my first choice. That makes the article part of my proof that I know how to use the technology. If it is verbose, generic or obviously under-supervised, readers should reasonably wonder whether the work I advise others to do is any better.
The useful distinction is authored versus unowned. AI-assisted work can still be authored when the writer can explain the evidence, defend the judgement and accept responsibility. Unowned work has a name attached without enough of that person inside it.
Why the backlash makes sense
The anti-AI crowd has legitimate complaints. Low-effort output consumes finite attention and transfers work to the recipient.
Another client described YouTube creators who seem to generate a script and read it into the camera. The human face remains, but the language sounds supplied by the same invisible writer. Carry that far enough and one person’s AI is debating another person’s AI while both humans stand behind them.
I have other things to do.
Uniform paragraphs, repeated rhythms, canned contrasts and examples with no lived detail can all contribute to the AI smell. They are editing signals, not proof. “This sounds like AI” becomes lazy when it replaces engagement with a sound argument.
When workslop moves downstream
An organization once invited me onto a volunteer committee likely to use skills I normally charge for. Seriously interested and needing to weigh the responsibility, I asked about the mandate, decision rights and commitment. The lengthy reply left the decision unresolved.
I can’t definitively say AI wrote it. I can say how it landed.
I have always been a slow reader. When paragraphs fail to answer a practical question, I can spend more time extracting the decision than the sender spent making it clear. The sender may save time. The slower reader inherits the unfinished thinking. The lack of brevity and respect for time also made me less inclined to volunteer.
That is one reason people dislike professional AI slop. It moves effort downstream.
Where do you draw the line? What makes writing feel unowned or AI-generated to you before you know how it was made?
The pendulum is already swinging
LinkedIn shows how quickly the pendulum has swung. It added AI tools for first-draft posts, profile-writing suggestions and résumé feedback. Now it says generic AI-looking content without clear perspective is less likely to travel beyond a writer’s immediate network, and it recently added a “Seems like AI slop” reporting option.
The platform hasn’t rejected AI assistance outright. But it is drawing a harder line around polished output that no longer feels owned by the person using it.
Substack has moved in a similar direction. Since July, readers have been able to scan posts for an estimate of how much appears human-written or AI-assisted, while publishers can add a “How I make this” statement explaining how they actually work.
I prefer that approach to pretending the presence of AI settles the authorship question. The more useful evidence is what the writer can tell you about the decisions, sources, process and responsibility behind the finished work.
Legacy publishing has supplied harsher examples. Hachette cancelled the US release of Shy Girl after AI allegations, which Mia Ballard denied. A reported offer above US$2 million for Jerry Falade’s Call Me, I’ll Hide the Body disappeared after his agents said they could no longer authenticate how the manuscript evolved. Falade denied using AI and argued that racial bias shaped the reaction.
What troubles me is that we are still in a “gotcha” stage, where suspicion can eclipse the work before we establish what role AI played. AI is a skill, not automatically a shortcut. Used well, difficult work can look easier from the outside. That is different from proving that a particular writer used it.
How I keep myself in the room
Rather than give AI a topic and place the answer under my name, I use a chain of editorial decisions, with my approval at each stage. Each approved decision becomes context for the next.
I start with the idea, sources and rough thesis, then ask AI to interview me with high-signal questions, often through several cycles. Where permission and confidentiality allow, I connect it to transcripts, Gemini notes and other records to recover the language of real conversations instead of inventing a cleaner memory.
Then I choose the reader personas and governing audience. Against those decisions I develop the title, subtitle, featured-image concept, opening and outline, then the body and conclusion. I repair transitions where needed, run anti-slop and claims reviews, verify sources, edit and handle post-production SEO.
I also give the system a linguistic signature drawn from earlier articles, emails, transcripts and pre-AI writing. That helps preserve how I sound. My experiences, source choices, rejected ideas and conclusions preserve why I have something to say.
Provenance can matter when sharing AI-assisted work too. I might say I asked AI to critique an emergency press release from the perspective of an experienced nonprofit crisis-communications specialist, then explain which recommendations I accept and reject.
I want to use AI openly without letting the model’s house style replace the person who had something worth saying.
Before publishing this article, I ran Substack’s new Pangram scan on the finished draft. It classified 100 per cent of the text as AI-generated.
I winced. Then I realized I could hardly have designed a better test of the argument.
The detector can analyse the language on the page. It cannot see the hours spent recovering source material, answering questions, choosing and rejecting arguments, restoring ideas the model dropped, checking claims or deciding what I was willing to publish under my name.
Parkinson’s Law says that work expands to fill the time available. AI has not meaningfully shortened the time it takes me to write an article I am proud of. Sometimes it takes longer. This article is closing in on 10 hours over the last week because I care about getting the argument right. The tool lets me retrieve more evidence, question the argument from more angles and make more deliberate editorial decisions. It can also misunderstand what I need, lose important context or silently drop an approved idea as the draft evolves.
This article did all of that. AI helped me recover workshop moments, test the thesis, choose the audience and develop the title, image and structure. It also repeatedly lost important beats that I had to notice and restore. The sentences were easy. Preserving the argument was the work. As AI writing becomes better and harder to identify, authorship will depend less on whether a model was involved and more on whether a person remained present in the decisions, the evidence and the conclusions published under their name.
If this is the kind of AI conversation you want more of, subscribe to Context-First Notes. I write about making AI more useful without losing the context, judgement and human responsibility behind the work.