Nobody asked me to take the semicolon out
Substack’s AI detector, its bestseller-only AI perk, and something I found out about broken promises in 1998
The sentence was fine. It had a semicolon in it, doing the ordinary work semicolons do, and I looked at that semicolon for a while and then took it out, because I could no longer be certain a scanner would read it as mine.
Nothing had happened to me. No flag, no accusation, no email, no reader saying anything at all. I had simply adjusted myself in advance, for an audience of one machine.
The feeling underneath that is not outrage. It is flatter and older than outrage. A kind of tired watchfulness, the state of a man editing himself so that he will read as acceptable.
What actually happened
On 21 July, Substack switched on an AI detector built with Pangram. Any reader can now scan a post, note, reply or comment past a certain length and get a percentage estimate of how much of it a person wrote. Chris Best announced it with a coinage for the con it is meant to catch: Claudefishing, the business of taking someone’s attention for writing that had nobody behind it.
The problem is real. There is a great deal of text about at the moment that nobody actually meant, and readers are right to resent paying attention to it. Substack also shipped a genuinely good idea in the same announcement, a place for writers to describe in their own words how their work gets made. After the first wave of complaints the company loosened the rules so that you can switch scanning off without having to run the scan first.
Then there is the accuracy question. Pangram’s current model advertises a false positive rate of one in ten thousand. Detectors have advertised numbers in that neighbourhood before and independent testing has tended to find them optimistic. Dr Sam Illingworth ran his own essay through and got a result of entirely machine, then entirely human, on the same argument, having passed the text through one more free tool in between. The score moved. The author never did.
Two weeks earlier, and much more quietly, Substack had shipped an MCP server. Connect your publication to Claude or ChatGPT and ask it about your subscriber count, your retention curve, which traffic source carried your last post. There is one eligibility condition. You have to be a Bestseller.
Velinwood Court’s note, which I restacked, reads this as AI being a purity problem when it touches your paragraphs and a productivity feature when it touches the dashboard of somebody who earns Substack money. The reading is right. The mechanism is a little gentler than the note suggests: the server is read-only, it cannot publish, it cannot send a note, and it cannot even see your notes activity. Whatever drafting happens, happens in your assistant, and you could do it today with copy and paste.
So the tool is unremarkable. The eligibility line is the thing I keep looking at. Substack has built one relationship with the writers who make it money and a different one with everybody else, and the AI question is where the seam shows.
What I learned about this in 1998
In 1998 I published a paper with Lynne Millward at Surrey on the psychological contract. Lynne died some years ago. The scale is still being cited, most recently last year, which is a strange sort of afterlife for thirty-seven statements and a factor analysis.
Denise Rousseau’s construct is simple enough to explain at a kitchen table. The psychological contract is what you believe you were promised in return for what you give. Her phrase for it is ‘paid-for promises’ (Rousseau, 1995). It appears in no terms of service. Nobody signs it, nobody can point to it, and it governs almost everything.
Which means breach is not a clause being broken. Breach is the moment you find out the promise you had been keeping was only ever yours. Robinson and Rousseau put it plainly in 1994: violation is closer to the norm than the exception. Rousseau’s later phase model describes what follows, a relationship shunted out of maintenance and into repair, where every subsequent signal gets read through the breach.
Our study ran the scale across 476 employees in four service multinationals. Relational orientation, the belief that you are in something ongoing and mutual, turned out to be the single significant predictor of how many unpaid extra hours people worked. Sixteen per cent of the variance, holding age, gender, job type, contract length, tenure and both kinds of commitment steady. Nothing else in the model came close.
Substack runs on unpaid extra hours. Every free post, every note, every reply written at eleven at night, every afternoon spent on a piece that seven people will read. There is a real business underneath it for a small number of writers. For the rest of us the arrangement is: do the work now, on the belief that it becomes something later. That is a relational contract holding up a platform.
The question everyone is arguing about is the wrong one
Whether the detector is accurate is a measurement problem, and it will be settled by measurement people with better corpora than mine. The question sitting underneath it is what a scanner does to a relational contract on the days it is right. You cannot audit somebody for suspected inauthenticity and keep the thing that made them work for free.
The second finding from 1998 is the one we did not expect, and it is the one I would put in front of Substack. We measured commitment to the organisation and commitment to the job, assuming the organisation would carry the relationship. It did not. Job commitment did nearly all the work. Take job commitment out of the equation and the organisational relationship almost vanished. The organisation turned out to be distal and abstract. The job was the thing people could actually feel.
Writers are not relationally committed to Substack. They are committed to the writing. The platform is the room the writing happens in. This is worse news than an exodus, because it predicts something much quieter. Nobody storms out. People simply stop doing the twenty-first note.
Who the false positives land on
Detectors do not fail at random. Liang and colleagues at Stanford put non-native English essays through a set of GPT detectors and better than six in ten came back misclassified as machine-written. What gets flagged is writing that is structured, literal, repetitive and precise. That is also a fair description of how a great many autistic and ADHD people write when they are being careful, and the emerging work on detector false positives points the same way.
I have spent ages telling AuDHD readers that the way they write is not a defect. A detector will tell some of them it reads as fake. The feeling that arrives with that is not new to anybody who has had it before. It is the old one, the sense that your natural setting is wrong and that the remedy is to perform a version of yourself that passes.
Which is exactly what I was doing with the semicolon, at some cost to the sentence.
We are a species that invented writing so that a thought could outlive the person holding it, and we have arrived at a machine whose job is to check whether a thought was in there. Somewhere in that sentence the last five thousand years does a small embarrassed shuffle.
The part nobody wants to say
March and Simon worked out in 1958 that people stay while the inducements outweigh the contributions, and move when they do not. Half my feed is talking about leaving. The other half is pointing out that there is nowhere to go.
A relationship you remain in because the alternatives are worse has a name, and it is not loyalty.
That is what makes the slow decay of a platform possible. It never has to be trusted. It only has to be less trouble than moving. Ghost, Buttondown, beehiiv and a plain mailing list on my own domains are all sitting there right now being slightly harder work, and the honest reason I am still here is that they are slightly harder work.
I have not left. I am not sure I will. But the semicolon is still out, and the sentence is still worse for it, and nobody made me do that.
Before the scan, before the argument about accuracy, that is what I would go looking for in your own drafts. The place where you have already started adjusting.
References
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779
March, J. G., & Simon, H. A. (1958). Organizations. Wiley.
Millward, L. J., & Hopkins, L. J. (1998). Psychological contracts, organizational and job commitment. Journal of Applied Social Psychology, 28(16), 1530-1556. https://doi.org/10.1111/j.1559-1816.1998.tb01689.x
Morrison, E. W., & Robinson, S. L. (1997). When employees feel betrayed: A model of how psychological contract violation develops. Academy of Management Review, 22(1), 226-256. https://doi.org/10.2307/259230
Robinson, S. L., & Rousseau, D. M. (1994). Violating the psychological contract: Not the exception but the norm. Journal of Organizational Behavior, 15(3), 245-259. https://doi.org/10.1002/job.4030150306
Rousseau, D. M. (1995). Psychological contracts in organizations: Understanding written and unwritten agreements. Sage.
Rousseau, D. M., Hansen, S. D., & Tomprou, M. (2018). A dynamic phase model of psychological contract processes. Journal of Organizational Behavior, 39(9), 1081-1098. https://doi.org/10.1002/job.2284