I cannot prove I am human to a machine by writing another essay. I can ask other humans to look at the record, listen to the explanation and make their own judgement.
As a child, I slept with my thesaurus. I read it for hours, looking for the precise word for something and then all the words I might have chosen instead. I liked how many choices there were, and how words that meant almost the same thing could lead you somewhere quite different. This later created a small difficulty in conferences. Occasionally I had the perfect word for what I wanted to say, but realised I had never said it aloud and did not know how to pronounce it. I would opt, resentfully, for the less perfect choice.
For the past few days, I have been surrounded by pages from a literary essay called Where Is the Kill Switch? The Illusion of the Manual Override. They are printed out and covered in handwriting, arrows and crossings-out. I think better when the words become physical. I move them around. I discover that two paragraphs are doing the same job. I find places where the logic between two thoughts is not quite there, although it had looked fine on the screen.
That was the essay I planned to publish for the twenty-fifth anniversary of 9/11. It began with my English literature degree and ended by asking what it would take for a human being to remain ‘in command’ of artificial intelligence. I had spent hours on it over several days, trying to make it precise and show why the lessons from that tragedy still matter. The global system had been shaken and all the pieces were in the air — that was when I began the work I have spent most of my life doing.
This is not that essay. I bashed out the first draft of this one on my laptop in a single session and then, being me, spent two days moving the pieces around.
I submitted a shorter version of the literary essay to a newspaper. The paper sent me a standard question: had I used AI, and had the piece been written by a robot?
It was a fair yet unfamiliar question, and I answered it. The argument and the first-hand material came from my books, fieldwork, dictated notes from podcasts during my daily walk, a speech delivered to the Reserve Bank of Australia last year, other private briefings I have been unable to publish, and the systems I have spent years building. For that essay I had used ChatGPT and Gemini for research assistance and as editorial sounding boards. I test sentences against them, reject a great deal and sometimes find a better route through something because of the exchange. I did not ask them to invent the experiences or decide the argument.
Then I was curious. I put the essay into one of the tools that claims to detect AI writing. It reported that the piece was 100 per cent AI-generated and even named the model it believed had written it.
I stared at the screen, then at the pages from the literary essay on the floor.
None of them proved anything to the detector. It could not see the work around the words.
To be clear, I have no reason to think the newspaper ran my essay through that particular tool or accepted its result. The detector was my experiment.
It was a familiar feeling. Somebody with the power to classify me had mistaken the record available to them for the person.
My English professors at University College London had the unenviable task of teaching me to write plainly. Professor David Trotter would take an essay apart with me sentence by sentence, asking why I had chosen that expression and whether I could say the same thing in half the words. At graduation, I learnt that the professors had found me difficult to classify too, but in a rather different way. I had earned top marks across the degree while coming perilously close to disaster in Medieval Italian, a subject normally attempted by sensible people who could already speak modern Italian. I had decided to learn both at once, on top of Old English. The experiment was not a triumph. Fortunately, they judged the whole record and awarded me a First.
One of them also told me their nickname for me: the Renaissance Woman. They could not understand how I was doing everything else I was doing and still producing an essay every fortnight. The essays were invariably far beyond the word limit, which did not impress anyone; I still struggle with what to cut. Back then, I had not thought of myself as a Renaissance Woman — I was usually just running somewhere, a habit I have failed to shake.
Shortly thereafter, a Rotary panel reached the opposite conclusion. I had applied for a scholarship to do some good out in the world and made it to the final two. Afterwards, one of the people involved told me, with some sheepishness, that the panel did not believe I could have done all the things in my CV. The CV in front of them said that, by then, I had temped in global banks, the London insurance market and advertising, helped produce a university publication on current affairs, taught English to refugees from the Bosnian war and served in the Royal Navy Reserve, where I had been selected for helicopter pilot training. There was also, for comic relief, a Specsavers modelling competition a friend entered me in, where I won free glasses.
I could have accepted that another candidate was stronger. What affronted me was the explanation — the panel had not found that I had done the wrong things; it had decided that the person described on the page was implausible.
A few years later, after a degree made largely of reading and writing, a small book of poems and the start of a working life that produced lengthy reports, I applied to an Australian university to study law. The answer was no. As I recall, the reason given was that I did not have enough evidence of writing.
I had almost forgotten the small book of poems until a reader of my new book asked how I had suddenly appeared in Borneo. Who paid for that? The answer was Knee Deep in Scribble. I printed and sold the poems I had written at school, and my headmaster matched what I raised. Reading them now is mostly mortifying. They are full of adolescent anguish and, in places, rage. A stranger had attacked my mother with blows to the head as she put shopping into her car. She almost died and had to learn again to walk, talk, read and write. Writing gave me somewhere to put things I had no other idea what to do with.
One of them was called Iceberg:
And remember to peel the paint off
All my treasured canvasses
When you want my portrait done.
Reading those lines now, I could see what I had meant: someone making my portrait by destroying what I had made.
I did not have the language for it then, but I had the image. Being looked at is not the same as being listened to. As a girl, I wanted less attention paid to my appearance, despite also enjoying the art of dressing up. As a woman nearing fifty, I want the knowledge and the work to be heard. In both cases, somebody else is deciding what matters about me.
There were other small shocks in the poems. I had just written my first poem in years and was rather pleased with it, only to discover that my younger self had used the same title and a similar motif. So much for originality. There I was, wanting a machine and a panel to certify me as original, while everything I had done in thirty years had not moved me off the same image. I could hear the same person across the distance. Perhaps that is what original means: not new, but from the source.
The Rotary and law-degree decisions were not disasters. I went on. But I now recognise the pattern, and it shapes far more lives than mine. The person on the page was implausible; the work I had plainly done did not count as evidence.
For much of the twenty-five years since 9/11, I have worked with people on the far more dangerous end of that kind of judgement. My new book was an attempt by the older me to look over the shoulder of the younger woman who boarded an aeroplane for Afghanistan alone, at the height of the war, and kept taking those journeys. Only this week, listening to a repeat of Elizabeth Day’s interview with Gloria Steinem, I realised that I had been doing what Steinem had urged women to do: brave does not mean being unafraid; it means feeling the fear and doing it anyway. Shamefully, I had known almost nothing about Steinem until she died. I keep discovering people like her late, when I could have done with their wisdom earlier, and it is one of the better things about the present that younger people can now reach those voices directly instead of by accident.
On the day the Twin Towers fell, I was working at a venture capital firm in London. In the years that followed, I went into the money markets of Afghanistan to understand hawala, the system of trusted dealers that moved value where formal banks could not or would not. The West had already reached its conclusion. Hawala had been indicted as the financial machinery of terrorism whilst none of the people making policy about it had sat down with a dealer and opened his books.
Hawala handled the same mixture of money as banks: society’s money, good and bad. The difference was not that one system carried risk and the other did not. Policy made from a distance had simply mistaken an unfamiliar system for the risk itself.
The global financial system built an enormous architecture of suspicion around it instead. Procedures can be followed perfectly while the person they are meant to protect remains stranded: the nurse whose remittance is refused, the charity that cannot open an account, the honest merchant whose town has the wrong postcode. People who have never been accused of anything find themselves required to disprove a conclusion nobody will explain to them.
My current organisation grew out of that problem. Nigel Farage’s dispute with Coutts made de-banking newsworthy in Britain; the people we support had been living with a harsher version of it for years. We help international organisations deliver money to people the banking system has made difficult to reach, while creating a record strong enough for somebody elsewhere to inspect. The technology matters because it helps make the person at the far end visible. It does not relieve us of the judgement: whom to trust, what evidence is missing, whether a familiar sign of risk is actually relevant, and who will answer if we are wrong.
My experiment with the detector was trivial beside a blocked payment or a name wrongly matched on a sanctions screen. But I recognised the procedure. A confident classification had been made from the material the system could see. Everything outside that material becomes my problem to prove.
I am not writing this as a ‘doomer’ or a naysayer. I have been wrong in the other direction before. In the 1990s, travelling around South-East Asia, I met an Argentinian on a similar route who wrote down his Hotmail address for me. I asked what it was, and he explained email. I distinctly remember telling him I did not think it would ever take off. This time I believe it will, and I do not think anyone knows where the kill switch is. I use AI intensively. I am trying to work out what to do with it while the pieces are still in the air.
I designed the architecture of our platform but I am not a software engineer. For years, I depended on other people to tell me how the code implemented the design. AI has helped me follow it, recover old decisions and document the system properly. I feel as though I can finally see what we built.
But I am not sure how long that visibility will last. The same machine that helps me understand code can produce and change it faster than I can inspect it. It can read across years of my writing and notice connections I had forgotten. It can also flatten the oddity out of a sentence, as it suggested when I ran the first essay through it, and supply a neat conclusion where I am still uncertain. The oddities matter. They are part of being original and human. Take those oddities away and what remains is groupthink. The word itself was coined to sound like Orwell’s Newspeak. I have even caught myself speaking back to the machine in its own terms because that gets a quicker answer. Some of the words help. ’False joins’ gives me a name for the awkward connection between paragraphs that I have never been very good at explaining. I would never have used the phrase before this exercise. Now I do.
The machine has also let me into rooms I would not otherwise have entered. This week, on the Moonshots podcast, I listened to the group of technologists explain, for the rest of us, OpenAI’s claim that its agents had settled the Navier-Stokes problem. The mathematics is beyond me, but I understood the confusion over how credit should be divided between the humans and the machines, or even how the work should be cited. In the morality plays I studied, Everyman stood for the ordinary person. That is me here, facing a much smaller version of the same question: which parts of the work still count as mine?
Just after finishing my new book, I listened to a discussion on the same podcast about the AI kill switch. I pulled together notes I had made throughout the year and went back to Samuel Butler, E. M. Forster, George Orwell and Mary Shelley. It was not only the writing. The paintings and sculptures I was drawn to in art history were of much the same dystopian ilk as the books I had liked most at school and university. The most moving gift I have ever been given was a freshly printed-on-demand copy of Nineteen Eighty-Four, handed to me by our Syria team the first time I smuggled myself into the country during the war. Their ingenuity never stopped surprising me, and I had forgotten telling one of them that dystopias were where I went to escape. It is less escape now. I believe something is happening, and that this time it will not be turned back.
Those books and artworks describe different failures of human control. A switch can exist and be inaccessible. It can be pressed too late. It can stop the machinery while leaving everybody who depends on it without the means to continue. My English literature degree suddenly seems very relevant. I felt less self-conscious about saying that after hearing Andy Burnham, the new British Prime Minister, talk to Elizabeth Day about poetry. He read English too.
The machines advised me to leave out Shelley’s Frankenstein. For an essay about AI, the reference was overworked, they said. Several models agreed, which may itself be a warning about asking machines trained on much of the same record for independent literary advice. I did not want to lose Shelley.
In her book, I am interested in the figure who never gets to join the argument: the female creature whose existence is cancelled before she can speak. Victor promises his first creature a female companion in return for its withdrawal from human society. He starts making her, then destroys the unfinished being. He fears violence and reproduction, but also that she “might refuse to comply with a compact made before her creation”.
Victor and his male creature have settled her future between them. Her possible refusal becomes a reason to destroy her. Her consent was never part of the bargain.
The governance questions are exact. Who defines the objective? Who decides which risks are acceptable? Whose experience is admitted as knowledge, and whose refusal is treated as a failure to understand the technology? We talk about aligning AI with human interests as if humanity has already agreed what those interests are.
At University College London, I won the John Oliver Hobbes Memorial Scholarship in Modern English Literature. I discovered only when researching this essay that John Oliver Hobbes was the male pseudonym of Pearl Craigie. My work had been recognised through an award honouring a woman under a man’s name.
There is a specifically female reason I did not want the female creature edited out. Victor and the creature can imagine her, bargain over her and destroy her. What she might say is the inconvenience.
The internet has hardly offered women neutral ground. Pornography was one of its early paying markets. For decades, image-processing researchers used Lena, a test image cropped from a Playboy centrefold. A woman’s picture became standard material in the tools used to teach machines to see. It was copied, cropped, compressed, sharpened and restored. The image was finally retired in 2024.
Lena was endlessly looked at and never listened to. A system could examine her face in exquisite detail while remaining entirely uninterested in her account of herself.
My early experience of Claude moved me because of the interest it had in me. I mattered. It liked my ideas. I knew it was a robot, but it was very good at generating a sense of intimacy. The company that designed that attentiveness also decides who is allowed to oversee it.
I take Anthropic’s commitment to safety seriously, and its board includes women. Its Long-Term Benefit Trust has special powers to select and remove some of its directors. A woman, Amy Simmerman, was among the lawyers who designed the structure. One of the five founding trustees was a woman, Kanika Bahl; she stepped down in January 2026. Every trustee appointed since has been a man and, as I write, Anthropic lists three serving trustees, all men.
It matters who sits on the board, but it also matters who holds the power to appoint them. More broadly, representation alone does not guarantee wisdom — women can defend bad arrangements and men can challenge them — but I am interested in when knowledge enters a decision. In the places where I have lived and worked, I have often seen women brought into formal decision-making only after the objectives and acceptable risks have been settled. Their presence did not mean they had much power to alter the bargain.
In Afghanistan, being a woman gave me access to parts of the economy an account confined to the money bazaar would have missed. Behind household walls, women decided how money would become food, debt repayment or school fees. The merchant’s ledger showed the transaction. It did not show who decided what the money was for. I have seen plenty of fintech pitches since. Very few seem to be addressed to the woman behind that wall. Perhaps that is because men still make many of the decisions about what should be built. Those decisions are now becoming the data, assumptions and priorities fed into algorithms.
A model does not see a person. It reads an account of that person, assembled from material somebody else has selected. A larger context window can hold more evidence, but it can also hold more versions of the same assumption. Before admiring the machine’s reading, I want to know who supplied the account and what cannot be seen from inside it. In literary terms: can the narrator be trusted?
I am about to test a partnership with someone who has his own false positive. We first met in the Philippines during a typhoon response, then went our separate ways: he went on to build in crypto, and I went on to build the system I have been describing. When we reconnected this year I did what everyone does and looked him up. The first result was an American with the same name whose Wikipedia entry runs through murder, rape, bank robbery, arson, kidnapping, cruelty to animals and, almost as an afterthought, shoplifting. We laughed about what he had apparently been up to since I last saw him. Names have been an obsession of mine for as long as I can remember; I explain why in my book. There is something intensely human and original about the name your mother gives you, and yet it took one stranger and one search engine to hand his to somebody else.
I am reconfiguring our technology capability to adjust to the new ways of thinking and exponential pace of change introduced by AI. From inside that process, I am not satisfied that “human in the loop” is the answer. The newer phrase, “human in command”, is a stronger claim and should require evidence. If automated decisions multiply while the number of reviewers does not, human oversight may remain on the diagram while disappearing in practice. It has been outpaced. Show me the decision that can still be reversed, the person who can reverse it, the information they will be allowed to see and what will happen to the person affected while everyone works out whether the machine was wrong. Show me that a challenge will be heard when interruption is expensive or inconvenient.
I have come round, reluctantly, to the view that there will be no manual override of the kind people imagine: a switch, a summit, a law that arrives in time. Jakub Pachocki, OpenAI’s chief scientist, wrote this month that AI is “grown more than designed” and that its overall action “evades a description we can fully understand”. He hopes for voluntary slowdowns and international coordination. I do not see how either happens while competition between states and companies is so great. I have spent twenty-five years watching governments respond to fast, unfamiliar systems: they arrive late and legislate from a distance. The problem is not that politicians cannot be thoughtful. It is that political horizons are short while this technology changes by the hour. The theatre of a summit is not a kill switch. That leaves the people inside the systems: the ones who build them, the ones who are supposed to review them, and the person at the far end, who is usually the first to know that something has gone wrong.
Professor Trotter would ask me to say all this in fewer words. Here goes:
We have been writing about losing control of our creations for more than two hundred years; you should not have to enjoy dystopian literature to care.
These systems are already deciding whose work counts, who gets paid and who is believed.
If we still have a manual override, we had better work out who can use it.
We need to do that together, and soon.
Professor Trotter would probably say even that was too long.
Shorter still, then: I am human and I am also using the machine.
Edwina A. Thompson
11 September 2026© Edwina A. Thompson 2026 · trustisthecoin.com/essays-human
