A reader wrote to me this morning. Kevin, a thoughtful correspondent, wanted to know whether I actually write these essays. The volume of what I publish made him suspicious. So did the typos. Could I assure him that a person, not a program, was behind The Hames Report?
I could and I did. But his question bothered me, because ultimately it’s not really a question about me. It’s a symptom of something much scarier and deserves a proper airing as a consequence.
Some months ago my tech team at the Asian Foresight Institute conducted a small test. We had built our own language model. It was trained on everything I have written and every piece of case work I have done as a strategist going back over four decades. Out of curiosity they fed it an entire chapter of The Management Myth, my first book, published in 1994, when the most sophisticated piece of software on my desk was a spell checker that couldn’t tell “their” from “there”. Our model declared, with total confidence, that it had written the chapter itself.
They tried a second, widely available model. It judged with 65 per cent certainty that the chapter had been produced “with the aid of AI”. Only a third system, a beta model belonging to one of our clients and not yet released to the public, declined to claim any hand in it.
I wasn’t quite sure whether to be flattered or offended by the first result, although it was oddly touching. A machine raised entirely on my vocabulary and sentence construction recognised its parent and mistook the resemblance for identity. Children do this. The second result is the one that should worry all of us. A general-purpose AI-detector looked at prose written before the World Wide Web had a browser and concluded that a machine had probably helped. On what grounds? Fluency, presumably. Structure? A certain evenness of tone that comes from having done this for a very long time?
Just mull over what that implies. Over thirty years of practice as a writer, of refining my individual voice and introducing the lyricism for which I am often credited, clearly has made my work virtually indistinguishable, to the machines, from work produced by machines. Competence itself has become evidence against the competent.
This isn’t a personal grievance. It’s a documented failure. In 2023 researchers at Stanford ran essays by non-native English speakers through seven commercial AI detectors. More than 60 per cent of the essays, all of them written by humans, were flagged as machine-generated. Ninety-seven per cent were flagged by at least one detector. The reason is instructive. The detectors score “perplexity”, a measure of how surprising the word choices are. Plain, orderly prose scores low, and low scores look artificial. So a Chinese student writing careful, correct English is convicted, while a native speaker who writes loosely walks free.
That same year OpenAI quietly shut down its own detection tool. The company that built the most famous language model in the world could identify its own output correctly only about a quarter of the time, and falsely accused human writers roughly one time in eleven. It withdrew the tool citing its “low rate of accuracy” and has not replaced it.
So when Kevin’s suspicion is put to the machines, the machines are no more reliable than his intuition. Possibly less. We seem to have built an accusation without a court.
Let me be plain about my method, since it’s a main source of the confusion.
I have arthritis in my fingers. Typing, for me, is slow and painful and getting worse. Some years ago I began composing by voice, using a dictation app that turns speech into text. I am a professional speaker; I have thought aloud in front of audiences for fifty years. Before that I wrote and published music. So composing an essay or a book this way is second nature. And it’s quick. Editing it is not. Editing takes forever, and the errors that survive are the ones a spoken sentence leaves behind when it’s written down: a word that sounded right but is not the word I meant, a homophone, a dropped connective. Those are the “inappropriate word uses” Kevin noticed. They are not machine artefacts. They are the voiceprints from a mouth, not fingerprints from a keyboard.
For a while I did use Quillbot to check both spelling and phraseology. I stopped, because it introduced more errors than it caught. It would “correct” my Australian English, flatten a deliberately unusual phrase, or invent a word I hadn’t said. The typos have actually declined since I removed it.
As for volume: I have something like 130 articles queued for The Hames Report, many of them written years ago, before the pandemic. The daily posts for The Virtual Activist are scheduled into next year. Dancing with the Future, published this year, was begun in 1996. What looks like an inhuman output rate is mostly a very human backlog, the accumulated sediment of a long working life, finally being released because voice has given me back a productivity that my hands had taken away.
I cannot claim that any of this is original or even novel. Milton dictated Paradise Lost to his daughters because he was blind. Henry James, his wrist ruined by years of longhand, dictated his late novels to a typist, and critics have argued ever since about whether the famous convolutions of his last period were the sound of a man speaking rather than writing. Nietzsche, his eyesight failing, bought one of the first typewriters and remarked to a friend that “our writing tools are also working on our thoughts”. He was right. But it didn’t make his thoughts any less his own.
Every generation of writers has used the prosthetics available to it and every generation of readers has grown briefly anxious about them. The anxiety passes once the tool becomes invisible. Nobody today asks whether an author “really” wrote a book because it was produced on a word processor rather than by using a quill. But we are at the uncomfortable moment when the new tool is still visible, and, more to the point, when a tool that genuinely can produce text unaided has arrived alongside tools that just transcribe it. The public can’t yet tell them apart, and I don’t blame them for that.
Here is the strategic problem as I see it, and it’s a foresight problem, not a technical one.
For the whole of human history, the existence of a text was proof that a mind had produced it. That was an assumption so basic we never thought to state it openly. That assumption has now collapsed. Text is cheap. Fluent text is cheap. And the response of many institutions has been to search for a new test of authenticity, some machine that will tell us who is real. There is no such machine. It’s highly improbable that there will be one. Because whatever a detector learns to spot, the next model will learn to hide. The arms race is unwinnable by design.
That leaves us with a choice between two futures. In one, suspicion becomes the default mode. Every prolific writer, every clear thinker, every non-native speaker who has laboured to write well is presumed guilty until proven human, and the proof is never quite available. Trust drains out of the written word altogether. We already see universities running students through detectors that the detectors’ own makers have disowned, and ruining reputations on the strength of a percentage.
In the alternative future, we stop asking where a piece of writing came from and go back to asking whether it’s any good. Is the argument sound? Does it resonate or tell you something you didn’t know? Does it earn your time? These questions have always been the only ones that mattered, and they have the great virtue of being answerable by a reader without any equipment at all.
I know which future I would prefer, and it’s not because I have anything to hide. It’s because I take pride in my writing and the alternative punishes the wrong people. It punishes the fluent, the practised, the disabled who need assistive tools, the foreigner writing in a second language, and eventually anyone who writes well enough to be mistaken for a machine that was itself trained to imitate people who write well. There’s a grim circularity in that, and we should refuse it.
So, yes. I write these essays. I speak them, actually, and then I spend far longer than I would like turning speech into prose, and I miss things, and my friends tell me about the mistakes when it’s too late to fix them. I am 81 years old. I have been writing about philosophy and the future of our civilisation for most of my adult life, and I am now living in a piece of it that I didn’t entirely foresee: a world in which the burden of proof has shifted onto the human.
I don’t resent Kevin’s question. I am grateful for it. Indeed I wrote to tell him that. It was asked honestly and it made me think. That’s what good readers do. But I would asksubscribers of The Hames Report to hold onto one principle. As machines get better and the detectors do not, judge the work. The typos are mine. So are the ideas. The former I will keep trying to fix. The latter I will never invite a machine to fix for me, and I doubt that it could.