Aaron Horwath's essay in Noema, "Now Entering: The Era of Bleh," opens with a copywriter named Lukas Weber who no longer writes. He manages agents. They write the newsletters, find the prospects, answer the email. Horwath's point is that this is the ideal most companies are chasing, and that the chase ends somewhere nobody wants to be: every company in a category pumping its thinking through the same few models, and every output sliding toward the same generic bleh. He borrows Kyle Chayka's AirSpace, the reclaimed-wood coffee shop that looks the same in every city on earth, and names the knowledge-work version Airbrain.
He is right about the symptom. I read AI output all day. Ours, our clients', and every month the answers eight AI engines give about the businesses we work for. The bleh is real. You can feel it in a proposal by the second paragraph.
Where I part with him is the cause. Horwath places the problem in the tool and in the process, and he prescribes friction and AI-free zones. I think the problem sits in an empty chair, and Heidegger described it long before there was a model to blame.
I. Why the Models Agree
Horwath cites a Duke and Technion study: ask language models the same open question and their answers cluster; ask people and their answers scatter. He reads that as the flaw. I read it as the models doing exactly what they were built to do.
In August I wrote about the Platonic Representation Hypothesis, the finding that every AI model, whatever its architecture and whatever data it trained on, converges on the same underlying structure of meaning. They converge because they all learned from the same object: human language. That finding is why getting recommended by AI is one job and not eight.
It is also Horwath's Airbrain, seen from the other side.
If every model reads from the same map, then every model, asked the same thing with nothing else to go on, walks to the same spot on it. And the spot a model walks to when nobody tells it where to go is the middle. The most probable next word is, by construction, the word most people would have written. The center of a map of everything people have said is the average of everything people have said.
Airbrain is the center of the map, delivered on request.
Drew Thomas Hendricks
The convergence is expected. What should worry us is how many people are asking from nowhere.
II. Heidegger Called It Averageness
In Being and Time, Heidegger describes a way of being that he calls das Man, usually translated as "the they" or "the one." Most of the time, he says, we do not speak as ourselves. We speak as one speaks. We enjoy what one enjoys, read what one reads, judge a book or a painting the way it is generally judged. The they is nobody in particular and everybody at once, and nobody is accountable for it.
He gives its features names that could have been written for this year. Its mode is Durchschnittlichkeit: averageness. Its effect is Einebnung: a leveling down of every possibility to what is already understood. Its language is Gerede, idle talk: passing along what gets said without the grasp of the thing that would ground it. Idle talk, he writes, offers the possibility of "understanding everything without previously making the thing one's own."
Read that line next to a stack of AI-written blog posts and it stops sounding like metaphysics.
Kierkegaard got there first. In The Present Age he described leveling as the work of "the public," an abstraction that nobody is and everybody answers to, which flattens every distinction it touches. He was writing about newspapers. Swap in a chatbot and the description holds.
A language model is trained on what gets said. Ask it with nothing of your own, and it hands you what one says: das Man with a keyboard.
Drew Thomas Hendricks
Ask a model for "a newsletter for a physical therapy practice" and you get the newsletter one writes. It is competent, and it could belong to any practice in the country.
Heidegger did not think the way out was to escape the they. You cannot. It is how all of us are in the world most of the time, and it is how language works. Authenticity, for him, is a modification of the they, not an exit from it: taking up your own situation, what I wrote about in April as thrownness, and choosing from inside it instead of letting the they choose for you.
The same holds for the model. Keep using it, and show up as someone when you do.
III. The Empty Chair
In April I argued that understanding is not stored in a model. It happens between the model and the person using it, in what Gadamer called the fusion of horizons. You bring your history, your knowledge, your questions. The model brings its own. What comes out belongs to neither side alone.
A fusion needs two horizons.
When one side brings nothing, no facts, no customers, no judgment, no history, there is nothing to fuse. The only horizon in the room is the model's, and the model's horizon is the average. Bleh is what a fusion of horizons looks like when one of the chairs is empty.
Bleh is what a fusion of horizons looks like when one of the chairs is empty.
Drew Thomas Hendricks
This is the Cookbook Problem at the scale of a whole economy. Everyone now owns the same three-star cookbook. Lukas Weber is not cooking. He is supervising the cookbook. The trouble is that nobody at his table knows anything the model does not already know.
What goes in the chair is whatever the model cannot have, because it was never written down anywhere the model could read:
- The fact only you know. The real wait for a first appointment. The vintage that sold out in a week. The question a client asked on Tuesday that nobody in your industry has answered.
- The customer as they actually talk, not as a persona document describes them.
- The judgment of what to leave out, which is most of what separates a good piece from a complete one.
- A position. Something you believe that the average does not.
Horwath notes that the em dash became the mark of AI writing. We do not use them at Nimbletoad at all, and a bleh paragraph with every em dash stripped out is still bleh. What gives it away is that nobody in particular could have written it.
IV. Which Bends in the River
Horwath's best image is the Kissimmee River. In the 1960s the Army Corps of Engineers straightened more than a hundred miles of meandering Florida river into a canal to stop the flooding. It worked, and the wetlands collapsed, and the state has spent roughly a billion dollars putting the bends back. His argument is that the slow, meandering path an idea used to take through an organization, draft to review to revision to approval, was doing work nobody saw, and AI is straightening it out.
I agree the bends mattered. I disagree about which ones.
A lot of the old meander was not insight. It was waiting. A draft sat in somebody's inbox for a week, came back with a comma moved, and went to the next inbox. Nobody misses that bend, and nobody should.
The bends that mattered were the ones where the draft ran into someone who knew something. The salesperson who said customers never use that word. The compliance lead who flagged a claim the practice cannot make. The owner who read it and said, "That's not us." Each of those is a horizon entering the dialogue. Keep every one of them. Straighten the rest.
I would change the question Horwath leaves his readers with, from "where should there be friction?" to "where does knowledge enter the work?" Every place a person who knows something touches the draft is a bend worth protecting. Everything else can run as straight as the tools allow.
An AI-free zone is a blunt instrument for this, and it can miss in both directions. A room with the tools and an empty chair produces bleh. So does a room with no tools and five people repeating what the industry says. People invented bleh. Every "full-service agency passionate about results" homepage was written by a human, long before anyone could blame a model. The AirSpace coffee shops were designed by people scrolling the same feeds. The tools made averageness instant and free.
V. The Doorman Is Holding the Context
Where I agree with Horwath most is Rory Sutherland's doorman fallacy. Define a doorman by the task on his job description, opening the door, and you will replace him with an automatic door. Then you discover he was also security, the taxi line, the person who knew every resident's name, and the reason the building felt like a place worth the rent.
The version of this I see most often is the "human in the loop" who checks AI output. A reviewer who does not know the business can tell you whether a paragraph is well-formed. They cannot tell you whether it is true. That is the prisoner in Plato's cave grading shadows: expert at the shape, blind to the object. Checking AI output for grammar and tone is easy and getting cheaper. Checking whether it is true of this business, this month, for these customers, requires someone who knows the business. That was always the doorman's real job.
Horwath, drawing on a Cambridge paper by McGurk and Khachaturov, predicts a barbell: value collecting at near-free AI production on one end and premium human craft on the other, with the middle hollowing out. I think the middle splits instead. Middle roles defined by tasks will go. Middle roles defined by context will stay, because context is the one input the models cannot converge on. It is not in the training data. It lives in the relationship, the history, the phone call last Tuesday.
A small agency lives in that middle, and I have made our bet: own the context, not the tasks. It is the same thing I wrote in April about accountability. One person who knows the whole situation, who can tell a true paragraph from a plausible one, and who answers for the result.
VI. The Second Reader
Horwath leaves out a second cost. I see it every month in the AI answers we read for clients.
Bleh used to cost you readers. A person skimmed your generic post, recognized nothing in it, and left. That was the whole price. Now there is a second reader. When someone asks ChatGPT, Perplexity or Google's AI Mode who to hire, the engine reads the web and assembles an answer. And an engine does not need your page for the average answer. It already has the average answer. It was built from it.
A page that restates the center of the map gives the engine nothing to attribute to you. What an engine can use a source for is what it cannot produce by itself: a specific fact, a verified number, a clear position, a named thing that only you did. Meaning is relational. A business is a region on that shared map, and a region needs edges. Bleh has no edges.
An engine does not need to cite you for the average. It is the average. It needs you for the part only you could have written.
Drew Thomas Hendricks
That is why, when we write for a client whose name we want the engines to say, the first question is never the keyword. It is: what does this business know that the average does not? It is also why we built Eidomai, to find out month by month whether the engines are naming a business at all, and whether what they say about it is true.
Horwath expects the market to eventually reward distinctiveness again, the way coffee shops drifted back from AirSpace once sameness stopped paying. I think that correction is already underway, and it will move faster in AI answers than it did in coffee shops. A customer might visit two identical cafés. An engine has no use for two copies of the same answer.
VII. Never Ask From Nowhere
The era of bleh is real, and old. Kierkegaard called it leveling. Heidegger called it averageness. What changed is the price: averageness used to take effort, and now it takes a few seconds and costs nothing.
The way out is the one Heidegger described. You do not leave the they, and you do not need to leave the tools. You take up your own situation from inside them. Every time you sit down with a model, you bring what is yours: your facts, your customers, your judgment, a position you are willing to be wrong about. Then the fusion has two sides, and what comes out belongs to someone.
The model will always hand you the average. The only question is whether anyone at the table knows anything else.