Substack's AI Detector Is Judging the Wrong Thing
AI is a tool, not a substitute for judgement. As Substack introduces AI detection, CEOs should focus less on how content is created and more on whether its ideas can withstand scrutiny.
Substack has introduced an AI detection feature that estimates whether a post was written by a human or with AI assistance. Predictably, opinions are divided. Some see it as an important step towards transparency, while others see it as another imperfect attempt to solve a problem that’s moving faster than the technology trying to detect it.
My first reaction wasn’t to ask whether the detector works. It was to wonder whether we’re asking the right question.
I’ve always thought AI was just another tool. Used well, it puts your thinking into overdrive. Used poorly, it accelerates rubbish. The difference has never been the technology. It’s the person behind it. If you have taste, judgement and a clear point of view, AI can help you articulate ideas more quickly and more clearly. If you don’t, all you’ve done is produce polished nonsense at scale.
That’s why I think AI detectors risk measuring the wrong thing.
Imagine hiring a CEO.
You wouldn’t read their CV, glance at a score and make an offer. You’d interview them. You’d ask why they made certain decisions. You’d explore the trade-offs they considered. You’d push back on their thinking. Within a surprisingly short conversation, you’d know whether they genuinely understood the business they’d led or whether they were simply repeating a well-rehearsed story.
Writing is heading in the same direction.
A detector can estimate how a piece of text was produced. It can’t tell you whether the person behind it actually understands the argument they’re making. Those are two completely different questions.
I’ve had people accuse me of using AI to write before. They’re asking the wrong question. The real question is whether I can explain and defend every idea I’ve published. If I can’t, then the criticism is fair. If I’ve outsourced my thinking, my judgement or my convictions, I’ve crossed a line. But if I’ve done the work—built the argument, challenged my own assumptions and arrived at a conclusion through experience—then I have no issue using AI to help organise those thoughts into something coherent.
That’s a distinction I think many people miss.
There’s another uncomfortable truth here. Most people believe AI-generated content is slop.
They’re right.
The internet is filling up with generic articles that say very little, and AI has dramatically reduced the cost of producing them. The unintended consequence isn’t simply more low-quality content; it’s that genuinely thoughtful work becomes harder to discover because it’s competing with an endless stream of competent-looking mediocrity. Some of that work may have been written entirely by humans. Some of it may have been written with AI sitting quietly alongside the author. The provenance matters less to me than whether the thinking is original.
Over the past year, one observation has become impossible to ignore. You can’t fake human interaction. You can’t fake lived experience. You can’t fake the lessons learned from difficult conversations, failed launches, unhappy customers or hard decisions. Large language models are remarkable at synthesising information, but they don’t accumulate wisdom in the way people do. They don’t sit in board meetings. They don’t negotiate with customers. They don’t carry responsibility for outcomes.
For CEOs, I think that’s the opportunity.
As AI becomes embedded in every business function, we’ll spend less time judging the artefact and more time judging the person behind it. I’d encourage more face-to-face conversations with the people creating your strategies, presentations, proposals and recommendations. Ask them why they chose a particular approach. Challenge their assumptions. Explore what they rejected along the way. That’s where genuine expertise reveals itself. The conversation is the test, not the document.
I’m also cautious about putting too much faith in any detector. At the end of the day, it’s software making a probabilistic assessment. Time will tell whether Substack’s approach proves valuable, but any system that labels a piece of writing carries the risk of getting it wrong. If a genuinely original human author is falsely identified, that creates questions that extend well beyond technology and into trust and reputation.
So if I had to choose between an algorithm’s opinion and a conversation with the author, I’ll choose the conversation every time.
Software can estimate where the words came from.
Only people can demonstrate where the judgement came from.
P.S. I promised I’d disclose this: this article was AI-assisted from start to finish, with every idea, argument and edit reviewed by me. Substack’s detector classified it as “100% AI-assisted”. Whether that’s right or wrong is almost beside the point—the software can’t interview me about the ideas. You can.