I'm So Tired of "Is This AI?"
The Substack witch hunt begins
AI detection is a futile exercise, and one will that will not matter in a few years time.
The share of AI-generated articles has already surpassed human-created, and whether you like it or not, you are interacting with LLM-influenced work all the time. We’ve reached the critical mass, and there’s no going back unless the entire ecosystem collapses (in fairness, can’t rule that out yet).
AI detectors are unreliable, easily gamed and produce far too many false positives. The biggest irony here is that they are trained on mountains of human-written content, and now use that as a measuring stick to decide what’s human or not. You’ve written like the humans I studied, therefore this must not be written by a human!
Not everyone cares. If you’re reading this, you may be a writer or an avid reader, but you’re in a bubble, especially on Substack, a microcosm of people who are — or want to believe they are — keeping the art of human creation alive. Most people outside of these walls couldn’t give a shit how something was written or by what, as long as what they engaged with was insightful, entertaining or moving.
To borrow from JA Westenberg’s great write up, In 5 years, nobody will give a damn about AI-detectors,
For the vast majority of people, and the overwhelming majority of the writing they actually consume, the manufacturing process is not actually a property they care about for its own sake. An airport thriller, a recipe, an assembly manual, a product description, a weather summary, etc. - the question “was this written by x human” is already uninteresting. It was uninteresting 12 months ago, and it’s uninteresting now. The questions that matter to us are simple - does this content communicate something useful / valuable / reliable? Is this content enjoyable?
That final line is the key point.
My stance on AI is this: If AI is part of your process, I don’t really care. If AI is your process, you can serve that slop to someone else.
It’s about the intention behind the use. That’s why most AI art leaves a bad taste in my mouth, because no matter how much prompting or post-editing, the software did the heavy lifting. The continued stories of “revolutionary entrepreneurs” who want to create businesses that pump out thousands of podcasts or books a week? Yuck. Writing is more of a grey area. Did somebody write a touching, deeply vulnerable essay, but use AI to help them workshop a title, or to fine-tune the intro so it hit harder, or to edit any parts that got repetitive? Fine by me. Did somebody type “write me a personal essay about loss in the style of my favourite blogger” and then just copy-pasta-publish the result? That’s a hard pass from me. (And trust me, you don’t need an AI detector to spot these bad actors). Intention is hard to quantify, and to what level those intentions remain genuine will differ from person to person.
That brings us to Substack, and the announcement that it’s partnering with Pangram — an AI detector tool that claims to be almost 100% accurate with a false positive rate of 1 in 10,000 — to allow this pointless activity to be performed at the click of a button on every post, comment and note. The rationale is that
“…people should know what they’re getting. For example, you might be fine reading someone’s sharp analysis in which they leaned heavily on AI to present their findings but only want to spend time engaging with replies written by a human hand. Either way, you should have a choice in the matter.’
I can see why it’s come to this. I think the point about wanting to know what you’re giving your time to makes sense, and if you are a staunch anti-AIer, you want the agency to spot the slop before you spend a second reading it. But I don’t believe this is the right way of going about that. Again, it comes down to intentions. Will users use this tool to check before they read, or will they use it to hunt down any minute trace of AI on the platform? You know the answer already. The result is going to be public lynchings. A bloodbath on the notes feed of people calling out their once favourite writers because their use of “not X, but Y” phrasing, the rule of three — a writing practice that goes back to ancient times — or em dashes upset Pangram. People will be crawling through a writer’s back catalogue (from this point on), every post, every comment, every note, before finally getting a “This is only 98% human written” result, and patting themselves on the back as they sound the morale alarm.
It might seem harmless to post a screenshot of a “bad” score, but in this current climate it could be damning, because there will certainly be a pile on afterwards, full of over-the-top rage and righteousness. And, to say it again, it might not be accurate! A new form of cancel culture awaits.
Sorry sir, you’ve been Pangramed.
I don’t think the question should be if people use AI or not.
It should be about how they use it.
And if that output still gave you value, or still delivers value to that person’s audience, I don’t see what the problem is. If a media company fired its writers and now pushes out sludge that offers nothing, stop reading. If a writer can’t be bothered to use their brain to reply to your comment, stop engaging. Here on Substack, if somebody clearly just generate posts — strange behaviour; if you don’t have the time or the passion, just don’t bother? — don’t read.
But at the end of the day, each to their own.
We’re in a new frontier of the Internet, and it’s weird, it’s confusing and it’s unsettling. But it can’t, and shouldn’t, be neatly wrapped up and weaponised with AI detectors.
They can measure whether AI was involved. But they can’t measure whether thinking, time, effort, ideation or editing was. We’re fixated on building for the metric that is easy instead of the one that is more important.
