This post was written by AI

This might be my final post on Substack.

Substack has integrated Pangram, an AI detector that allows readers to scan writing on the platform and receive a judgment about whether it was written using AI.

I don’t want my work read that way.

A little while ago, I downloaded the Pangram Chrome extension because I wanted to understand what using an AI detector actually felt like. It was horrible.

Before I had even read a LinkedIn post, Pangram had already placed a judgment beside it. Instead of encountering someone’s words and deciding for myself what I thought about them, I was being told how suspicious I ought to be. I could feel it colouring my response. The detector wasn’t helping me exercise judgment. It was getting there first.

We talk constantly about the importance of retaining human judgment in an AI-mediated world. Apparently, that means embedding an automated verdict into the reading experience before the human has had a chance to read.

Even Substack’s own announcement repeatedly hedges the claims being made for this technology. Pangram cannot tell whether great human care went into producing a piece of writing. It cannot tell whether the ideas, experiences, expertise or judgment behind it came from a person. It is, we are assured, “not perfect”.

So why install it as an arbiter of authenticity?

Writers can also run Pangram on their drafts before publication. Presumably, that means anyone who wants to conceal their AI use can repeatedly edit a piece until it passes. Writers can then report and remove scans they believe are mistaken.

So what exactly is being achieved here?

This is the same technological arms race we have watched unfold in education: detect, evade, improve the detector, improve the evasion. The people acting deceptively learn how to circumvent the system, while everyone else has their work subjected to an unreliable automated judgment.

The detector therefore cannot reliably establish what matters, can be gamed by anyone motivated to game it, and may be wrong often enough that Substack has already built in a mechanism for removing its judgments.

But sure. Let’s make it part of the reading experience.

I have never believed that the presence of AI assistance tells us whether a piece of writing contains human thought, care, expertise or intention. I don’t believe typing every word yourself is what makes an idea yours. I don’t believe readers are entitled to an audit trail of another person’s cognitive and linguistic processes before deciding whether their work has value.

As I have argued here before: maybe it’s nunya.

Substack says people should be free to choose which tools they use to express themselves. It also says that Substack itself uses AI all the time—to write software, conduct research and build features.

“But people should know what they’re getting.” Do they though?

I’m not really interested in publishing on a platform that encourages readers to assess the supposed production method before engaging with the ideas. I don’t want my writing accompanied by a machine-generated suspicion score. And I don’t want to support the normalisation of automated systems passing judgment on people’s words while pretending they are merely providing “transparency”.

So, bye.

And because this decision is, fundamentally, about my right to decide how I produce my own writing, I have deliberately used AI to write this entire post. Normally, I would edit an AI-assisted draft extensively until every sentence felt precisely like mine.

This time, I haven’t changed a word. But it is everything I want to say.

Pangram will probably tell you it was written by AI.

Congratulations to it.

I’ve just told you the same thing myself.

---

EDITED TO ADD BECAUSE JFC ARE YOU KIDDING ME: