In 5 years, nobody will give a damn about AI-detectors.
Why the Pangram screenshot is a doomed beast
đ Iâm JA Westenberg. I write essays about technology, philosophy, productivity, and what it means to be human. If you like ideas that take their time, youâre in the right place.
While youâre here, a few things Iâve built: Kerouac, a writing app | Post Through It, a solo microblogging tool | Distributism.ai, my MCP platform | Laterware, all the AI news you need | Studio Self, a thought leadership agency for founders x VCs.
You can buy my book Permissionless here.
Every time an article goes viral, the Pangramguy appears. He tags Pangram and asks for a score, or he pastes away and smugly takes a screenshot of the output, watching the number climb toward â100% AI generatedâ with the satisfaction of a hall monitor catching a kid with a bag of weed.
This is fake. This is lazy. This is entirely beneath consideration.
For the last 12 months, that screenshot has carried social weight - itâs an excellent tool to shame anyone for their writing when you either disagree with their work or donât like them personally. But that social weight is already lifting, and within five years, the performance will look dated and damn near quaint.
This whole apparatus is a transitional artifact. Tools like Pangram will get better and better, but the actual verdict will matter less and less. The share of people who give a shit about whether a given piece of writing, art, or code was produced with AI is going to fall off a cliff, and that fall has already started.
Five years from now, asking âIs this AI?â will feel about as urgent and meaningful as asking whether a photo was taken on film or digital. A few specialists, myself included, will care intensely. Everyone else will have moved on. Iâm not making a values or morality-based call on whether this is a good or bad thing.
Iâm saying itâs inevitable, either way.
People will just stop giving a shit - the way people always stop giving a shit about change and technological advances, which is to say gradually, and then all at once, without ever actually making an active decision.
The detectors were always going to go the way of the buffalo.
Start with the why: why does anyone run an AI detector in the first place? Nobody wakes up with a pure // disinterested curiosity about the provenance of a block of text; the detector, and the act of using it, are a proxy for something else they actually care about - usually one of three things:
Did this person put in the right amount of effort to produce thing?
Is thing any good?
Can I trust the person who created thing?
For a relatively brief window (relative to asteroids and planet death, etc.), âwritten by a humanâ was a decent enough proxy for all three. If a student wrote their own essay, they probably engaged with the material, at least insofar as they searched Amazon Books and glanced at a few previews. (I was a bad student. Sue me.) If a writer produced their own prose, it reflected actual thought, regardless of how much of that thought belonged to the writer himself. If a candidate wrote their own cover letter, it told you something about their effort and competence, no matter how much of it was likely bullshit. Authorship was both easy to check and only weakly correlated with what you actually wanted.
The âis this AIâ question was never the question; it was a shortcut around better questions. When the correlation breaks and the proxy decouples, people will abandon it. It has simply stopped being a useful measurement, and useful measurements get dropped without ceremony.
When enough of the internet is going to be AI-generated (letâs not flatter ourselves with optimism that the reverse will be true), and you can assume youâre reading AI content at least half the time, the test of AI-or-not becomes largely useless. Post-AI, you actually have to measure each piece and each originator, on a case-by-case basis, against a set of rules that define your answer to those three questions, which is probably a better outcome than using a shortcut in the first place.
But what about the process? What about the act of sitting down at a typewriter and bleeding?
Suppose I hand you a perfect provenance oracle and a stack of writing. You run every piece of content through it, and you now know, with absolute certainty, the origin of each piece. What have you actually learned?
Youâve learned a series of facts about a manufacturing process; you havenât learned anything about quality, truth, effort, or whether any piece is actually worth your time - unless your sole decision-making factor is based on that manufacturing process.
None of the millions of people who buy James Patterson novels actually care that James Patterson hasnât written a James Patterson novel for the best part of a decade; he simply hands outlines and ideas to a ghost writer who produces the finished work. This is a widely known and accepted fact in literary circles, but it just doesnât seem to matter. Every book that comes out under his name sells like wildfire, regardless of whose fingers touched the keys.
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 âwas this written by x humanâ is already uninteresting. It was uninteresting 12 months ago, and itâs uninteresting now. The questions that matter to us are simple - does this content communicate something useful / valuable / reliable? Is this content enjoyable?
The author was rarely the point; why else would people complain at length about recipe creators injecting their personal narratives onto the page? The recipe is the only thing people actually care about; for most folks, looking up a good brownie recipe, they couldnât give two shits about the originator of that particular dessert.
We have done this every single time.
Moral panics are largely temporary. Almost everything we were sure would rot our culture, our brands, and the world around us became background noise within a generation. Calculators were going to destroy math; now a kid pulls out a phone to split a bill, and nobody bats an eye. Spell-check was a crutch that would make us all illiterate, and Wikipedia was not a real source, and typing out your essay instead of handwriting it meant you hadnât learned a thing, and Photoshop meant you could never trust an image again, and Auto-tune was the death of music. Every one of these things caused a moral panic, with op-eds and hand wringing, and - for better, and in some cases absolutely for worse - every one of them is now so normal that the panic feels archaic.
Nobody had to make a decisive argument that finally proved the value of the calculator, or Wikipedia, or anything else. The outrage just aged out. A new batch of people grew up with the product / tech already in the room, and they treated it like furniture, and couldnât be made to feel scandalized by it, no matter how hard you tried. You canât sustain a moral panic about something that feels like it has just always been there. For people turning 16 from now to the next decade (and I regret to tell you, there are more of âem every single day), AI will have functionally âjust always been there.â
The norm eventually gets old, dies, and gets replaced.
The people running detectors right now are mostly folks for whom AI showed up partway through their lives, and late enough to feel like an intruder or a thief in the night. Theyâre the cohort of users who experience AI content as a betrayal or a violation of the contract. And to a degree, Iâm one of them! But we are, sadly, a temporary cohort.
To the people in school right now using AI to draft, summarise, argue, rewrite, and unstick themselves a dozen times a day, AI is already just a pen. When theyâre the teachers and editors and managers in 5-10 years - and yes, itâll happen that fast - the question âbut did a human right this?â is barely going to cross their minds.
The accuracy problem is unsolvable.
âŚby which I donât actually mean Pangram is inaccurate. By every benchmark Iâve examined, by every report Iâve read, itâs pretty damn good at picking up AI-generated content.
But itâs not perfect.
Itâll still generate a false positive every now and then, and all you actually need for the chunk of people who do still give a shit about AI vs. non-AI content is one single example of Pangram getting it wrong. Thatâs enough for them to distrust the output of the whole program, every time.
And unfortunately for Pangram, theyâre already behind the 8-ball. Why? Because earlier detectors were extremely likely to be extremely fucking wrong, including the infamous example of GPT Zero calling the Declaration of Independence an AI-generated document. For most people, the difference between the program that produced that egregious error and Pangram is nil. They donât care. They just care that it happened once, with one AI detector. And as far as theyâre concerned, all AI-detectors are functionally the same.
A single false-positive doesnât create one skeptic - it creates a hundred, a thousand, who screenshot and share and rage and crow.
And the overuse of Pangram hurts the cause, too. When someone scans a tweet from the Pope with Pangram and announces that the Pope is using AI, most people donât actually believe Pangram - they believe the Pontiff. And so a whole new strand of Pangram rejectors is born. At some point, the saturation of Pangram screenshots will convince people that either A) everything is AI, so why bother, or B) Pangram is wrong more than they claim. Either way, exhausted scrollers, writers, and readers will simply stop paying attention.
Whatâs left when the caring drains out
If youâre sweating a pangram score, I think youâre optimizing for a thing that is about to stop mattering at all; learning to fool a metal detector that is just going to be unplugged anyway. If youâre crowing about that same score and holding it over other writers, I think a lot of folks are eventually going to start treating you like a PETA activist with a can of red paint. They might directionally agree with what youâre doing in principle, but theyâll be sick to death of the way you do it.
When people stop asking âIs this AI,â I donât think theyâll actually stop caring about writing itself. I think theyâll go back to caring about the stuff they always actually cared about, and were using the AI question as a lazy shortcut for. If a blog post meaningfully changes your life, gives you the strength to get sober, inspires you to create - I donât think youâll care about the provenance as much as what you gained from reading it. And what you gained is entirely up to you to measure and define.
Write something worth reading, if thatâs what you want to do. Write it with AI if thatâs your poison. Write it yourself if you want to hang out with the cool kids. Read what you want to read, and decide for yourself if it was worth it. But ignore the pangram score. It might feel like a verdict, but I think itâs already a relic; a precise answer to a question that weâre all going to find boring, asked by a machine that most people will learn to ignore.
I must be ahead of the curve cause I dont give a damn now.
Agreed, and please reframe the thing about âthe way of the Buffaloâ
Buffalos were bountied and slaughtered almost to extinction by the US government to take the means of living away from people who relied on them.
Itâs not like it was anything ânatural.â