Judge, Jury, AI

Every platform now passes a verdict on whether a human wrote this. The tools can’t tell, and origin was never the thing worth judging.

For a while, if you opened almost any writing platform, you could watch the same argument happen a thousand times a day. The em dash is the mark of the machine. Nobody says “delve” any more without indictment. A quiet pile-on around a shifting list of proscribed words, “genuinely,” “curious,” “boasts,” “testament,” each one promoted from ordinary vocabulary to evidence for the prosecution. “This place is turning into AI nonsense”, ran the prevailing complaint, usually written in prose no more obviously human than the thing it was objecting to.

The platforms heard it. And because a platform’s instinct when it hears a complaint is to ship a feature, they did.

So now there is machinery. Substack has partnered with a detection company called Pangram, which scans what you publish and hands the reader a percentage, a number claiming to say how much of your work a machine thinks a machine wrote. LinkedIn has added a button, tucked in the menu on every post, that reads: Seems like AI slop. Instagram attaches a label, sometimes because the writer disclosed it, increasingly because the platform decided for itself. Three of the largest publishing platforms in the world, within about a year of each other, arriving at the same conviction: that the reader is owed a verdict on whether a human made this, and that they will supply it.

And this is where all hell breaks loose, because the machinery does not do what the complaint assumed it could. The entire apparatus rests on one assumption, so widely shared that almost nobody thought to test it: that AI writing can be reliably told apart from human writing.

It cannot. Everything downstream of these buttons and badges and percentages is the consequence of building a courthouse on top of a thing that does not work.

The sentence for the flagged

Set aside for a moment whether the detection works. Assume, generously, that it did. A more interesting question sits underneath it, and it is the one nobody is answering: what happens to a piece of writing once it has been flagged?

Nobody will tell you, because the honest answer is that nobody outside the company knows, and the company has no reason to say. It is reasonable to assume that content marked as likely-AI will be treated differently by the systems that decide who sees what. Not necessarily worse. Differently. Shown to fewer people, or more, or the same number but a different set of them. The point is not that the outcome is bad. The point is that the outcome is unknowable, applied automatically, by a system that has never once had to show its working.

We have just watched this exact dynamic play out on one of these platforms, on a different axis. Late last year a group of women, led by the entrepreneurs Cindy Gallop and Jane Evans, ran a simple experiment. They had men post their content, word for word, and compared the reach. Gallop’s post, to an audience of well over a hundred thousand, reached around eight hundred people. The same words on a man’s account, with a fraction of the following, reached over ten thousand. LinkedIn’s response was that its algorithm does not use gender as a signal, that its own testing found no demographic bias in distribution, and that the differences people were seeing had other explanations. The problem is that this is unfalsifiable from the outside, because the algorithm is a black box and the only party who can see inside it is also the party being accused. The episode did not resolve; it simply dispersed, leaving behind a widespread private conviction and no way to confirm or dismiss it.

That is the model for what is coming with AI detection, and it is worse, because at least gender is a thing you cannot be wrongly assigned. A false AI flag is an accusation you cannot see, cannot appeal, and frequently will not even know you have received. Your reach quietly changes. You theorise about why. You never find out. The verdict was delivered, acted upon, and sealed, all before you knew a trial was underway.

The tool does not work

All of which would be a manageable problem if the detection were accurate. It is not, and this is not a matter of opinion or of waiting for the technology to mature. It is documented, repeatedly, by the people who have actually tested it.

Dr Sam Illingworth, who writes the newsletter Slow AI and has done more careful work on this than almost anyone, keeps arriving at the same finding from different angles: the detectors cannot reliably do the one thing they are sold to do. And when they fail, they do not fail randomly. In a Stanford study, seven detection tools were run across essays written by people who speak English as a second language. On average they flagged 61% of those genuine, human-written essays as machine-made. One tool flagged 97%. Every word had been written by a real student sitting a real English exam.

This is not a quirk of one dataset. Louie Giray, a researcher at Mapua University in Manila, has documented the same pattern in the academic literature and named its victims precisely: false positives fall hardest on non-native English speakers and on writers with distinctive styles, and the accusations that follow do real damage to real careers. The tools mistake carefulness for machinery. A person who learned English deliberately, and therefore writes it cleanly, structurally, a little formally, produces exactly the texture a detector has been trained to read as artificial. So does a neurodivergent writer whose patterns do not match the expected human mess. The writing that gets flagged is disproportionately the writing of people who have already spent their lives being corrected for how they express themselves.

The institutions closest to this have quietly backed away from it. Vanderbilt University, which submits tens of thousands of student papers a year, disabled Turnitin’s AI detector in 2023 and said something worth quoting, because it is the whole problem in a sentence. Even at the 1% false-positive rate the vendor claimed, that would mean around 750 students a year wrongly flagged. And, they added, they had no insight into how the tool worked. An organisation paying for it could not see inside it either. They did the arithmetic, could not verify the mechanism, and switched it off.

This is not confined to students and homework detection. As I write, it is happening at the top of the publishing industry, in public. In the summer of 2026 a debut thriller that had sold at auction for a reported figure north of two million dollars was withdrawn by its own agents over suspected AI use. The detail worth holding onto is that the agents did not run the manuscript through any detection tool. They said so plainly: the technology was not, in their judgement, reliable enough to trust. They acted on the suspicion regardless. The author, Jerry Falade, denied using AI and observed that he was one of three Black writers to sign major deals that year and then face AI-suspicion cancellations, a pattern he did not think was a coincidence. Around the same time a publisher pulled a horror novel that had already sold thousands of copies, on the strength of reader suspicion aggregated online. The verdicts arrived first. The evidence, such as it was, came after, if at all.

And here is the part that is truly befuddling. In each of these cases, the writing was not the problem. The thriller went to a fourteen-way auction because editors thought it was extraordinary. The novel that got pulled had readers who could not put it down. Nobody withdrew these books because they were bad. They withdrew them because they might have been made with the wrong tool, which is a different objection entirely, and a much stranger one. We have arrived at a place where a piece of writing can be, by common agreement, one of the best things an editor read all year, and still be destroyed on the suspicion of how it came to exist. The quality was never in question. Only the confession was.

Slop

But now we get to the real nail in the coffin. None of this was ever really about detection. Accuracy is beside the point, because the machinery was never built to establish a fact. It was built to license a judgement.

Look again at the word LinkedIn chose. Not “this may be AI-assisted,” not “automated content,” nothing neutral or verifiable. Slop. It is a sneer wearing the uniform of a content-moderation category. And a sneer does not require the underlying claim to be true, which is the entire advantage of it. “Seems like AI slop” is not a statement about how a piece was made. It is a statement about how a piece made the reader feel, dressed up as a statement about its origins, and handed a button.

This matters because slop is not a category that exists in the writing. It is a perception that exists in the reader. The same three paragraphs can be slop to one person and not to another, depending on nothing more than whether they liked it. A carefully worked essay with two em dashes in it can be flagged as slop by someone who has recently decided em dashes are the tell. A genuinely lazy, thoughtless post can sail through untouched because it happens to be written in the rough, misspelt register we have agreed to read as human. The label cannot tell the difference between contempt and detection, because it was designed to let the first pass as the second.

And that is what the button actually distributes. Not accuracy. Permission. Permission to convert “I didn’t like this,” or “I disagree with this,” or “this person writes more cleanly than I trust,” into an accusation that sounds technical and therefore fair. We spend whole paragraphs worrying about whether the machine can tell AI from human, and the honest answer is that it does not need to, because the machine was only ever there to give a human’s suspicion the authority of a verdict.

The people who wrote the complaint that started all this wanted less thoughtless content in their feeds. That is a reasonable thing to want. What they have been given instead is a mechanism for pointing at anyone, on the strength of a feeling, and having the platform nod along.

Everyone has a microphone

There is a fair version of the complaint underneath all this, and it deserves to be met rather than dodged.

Feeds genuinely are fuller of nothing than they used to be. Posts that say what a hundred other posts said that week, in the same shape, reaching for no thought the writer actually had. This is real, and the irritation with it is legitimate. But it is worth being precise about where it comes from, because the current panic has misfiled it under AI, and the filing is wrong.

Start with a number that cuts against the obvious story. Despite a decade of being told that everyone is now a creator, only around 1% of LinkedIn’s active users actually post in any given week. The emptiness in the feed is not the sound of a billion new voices arriving. It is a very small number of people posting a great deal more, and increasingly posting with help. The barrier that fell was never the barrier to entry. It was the barrier to volume.

Social media handed a microphone to everyone, which was the entire promise of it, and the promise contained a cost nobody quite priced in: a platform is not the same as something to say, and we built a culture that treated them as identical. Having the account became the qualification. The result, long before anyone had heard of a large language model, was a growing pile of content produced by people who had access to an audience and no particular point of view to bring to it, repeating the ambient consensus back to itself, because repetition is what you do when you have the means to speak and nothing pressing to add.

AI did not cause this. What AI did was remove the last bit of friction that used to force a minimum of effort. Before, a person with nothing to say at least had to type the nothing themselves, and the labour of it filtered a little of it out. Now the nothing can be generated in full, instantly, at length, and posted before the absence of a thought has had the chance to become inconvenient. The floor did not lower. It fell away.

Analogue, Automate, Augment, Delegate

The reason a binary verdict of AI or NOT AI is so hollow is that it cannot see any of the actual work. Human or AI, on or off, real or fake: none of these describe how a thinking person uses these tools, because the real answer is not a yes or a no. It is a series of decisions about which parts of a task carry your judgement and which parts do not.

At lifestack, the business I am building with my co-founder, this is the model we work from, and I will name it plainly because it is the clearest argument I have against the slop button. Four words. Analogue, automate, augment, delegate. They are not stages you graduate through. They are a description of where the thinking lives.

Analogue is the part only you can do. The idea, the point of view, the judgement about what is worth saying and what it means. Nothing touches this. It is the thing the empty post is missing and the thing no tool supplies, because it is not a task, it is a mind. Automate is the machinery underneath: the processes and structures that mean the work can happen at all, the way I plan and hold a body of writing so I am never starting from a blank page for the wrong reasons. Augment is where you codify your own judgement so a tool can extend it rather than replace it, feeding it your back catalogue, your standards, your prior thinking, so that what comes back is more you rather than less. And delegate is the handing-off, to a person or a machine, of the work that does not need your hand on it, so that your capacity returns to the part that does.

I can show you this because you are reading the output of it. Every idea in this essay is mine. The argument, the structure, the position it takes, the order it takes it in: analogue, all of it, and not negotiable. I used AI to research, to check, to find the study I half-remembered and the figure I could not place. I used it to edit, to argue back at a sentence that was not working, to hold my own past writing up against the draft so I could sharpen myself against myself. At no point did it decide what I think. At every point it extended how far I could get with what I think.

A detector, run over this piece, would return a number. That number would be an answer to a question so crude it is barely a question at all. It could not tell you which sentences are the ones I lay awake with and which are the ones a tool helped me tighten. It could not distinguish the analogue from the augmented, because it was never built to see that anything but the binary exists. And the reader who reached for the slop button would be doing the same thing: collapsing a whole spectrum of judgement into a single verdict, and mistaking the collapse for discernment.

The lost art of reading

There is an older way to know whether a person is behind a piece of writing, and we have not lost it. We have just stopped trusting it, in favour of a button.

You read the thing. You notice whether it has a point of view, whether it argues something a person would bother to argue, whether there is a mind in it doing work you could not have predicted. And if you want to know how it was made, you ask the person who made it, and they tell you, and you believe them until they give you a reason not to. This is not naive. It is the entire basis on which writing and reading have ever worked. The trust ran between two people, and it was extended, and occasionally it was betrayed, and the betrayal was survivable because it was rare and because a person could be held to account for it.

Here is what I suspect will actually happen, though. Enormous quantities of attention will now be spent litigating whether these detection tools work, whether the percentages are fair, whether the button is being abused, when none of that is the question that matters. The question that matters is whether a given piece of writing, made by a human or a machine or some honest combination of the two, is adding anything to the conversation or taking something from it. And on that measure the origin is irrelevant. Plenty of entirely human output detracts, actively and on purpose. Andrew Tate did not need a language model to make the internet worse.

So here is the older technology I would reach for instead, and my mother taught it to me on a playground. If someone is saying something you do not like, go and talk to someone else. Do not give their version of events your airtime. If you have decided a piece is AI, and that this makes it not worth your time, then the remedy was never a public verdict on the writer. It is the quietest thing available to you. Unfollow. Unsubscribe. Mute. Move.

The one thing all of us can be certain of is that there are more people saying interesting things than any of us will ever have time to read. That was true before the machines and it is true now. Go and find them. Spend your attention on what deserves it, and let the rest go unremarked, which is a fate far worse for a bad piece of writing than a flag could ever be.