A while ago I read a draft I had written and did not recognise the person who wrote it.It was not bad. That was the problem.The sentences were clean. The structure held. The research sat in the right places. Every paragraph did its job and moved on.And it sounded like nobody.I had used AI for the heavy lifting, the way I always do now. Research. Structure. Sorting the mess into an order. What came back was smooth, competent and completely unhaunted.There was no scar in it. No moment where a real person had been wrong about something and had to work out why.I deleted most of it and started again at 4:30 the next morning.What I did not understand at the time was that this was not my failure as a writer. It was a measurable effect, and researchers have now put a number on it.The tool lifts the floor and lowers the ceilingIn 2024, Anil Doshi of University College London and Oliver Hauser of the University of Exeter ran a controlled experiment and published it in Science Advances. They asked people to write short stories. Some wrote alone. Some were offered story ideas from GPT-4. Then evaluators rated the results, not knowing which was which. You can read the study here.The stories written with AI help were rated as more creative, better written and more enjoyable.The gain was biggest for the least creative writers. The tool lifted the people who needed lifting. That is a good thing and I want to say it plainly before I say the rest.Here is the rest.The AI-assisted stories were more similar to one another than the human-only stories.Everyone got better. Everyone got closer together.The researchers called it a social dilemma. Each writer is better off. The group produces a narrower range of what exists.Individually better. Collectively smaller. That is not a trade-off anyone agreed to make.Then they counted the ideasA second team went looking for the same effect at scale. Kibum Moon, Adam Green and Kostadin Kushlev analysed 2,200 college admissions essays across three preregistered studies and published the results.They built a measure they called the diversity growth rate. It asks a simple question. When you add one more piece of writing to a pile, how many genuinely new ideas does it bring?Every extra human essay added two to eight times more new ideas than every extra GPT-4 essay.Two things about that finding matter more than the number itself.The gap grew wider as more essays were added. The homogenising effect compounds. The bigger the pile, the more the machine repeats itself.And they tried to fix it. They changed the prompts. They changed the parameters. They pushed the model toward variety in every way they could think of.The gap survived all of it.Data: Moon, Green and Kushlev, Journal of CreativityIt is not only writingYou could argue that essays and short stories are a narrow test. So look at raw idea generation, before anyone writes a word. Barrett Anderson, Jash Shah and Max Kreminski ran a study with 36 participants who produced 1,271 ideas between them, and published it at the Creativity and Cognition conference.Half used ChatGPT. Half used a different creativity tool with no AI in it at all.The ChatGPT group produced more ideas, and the ideas had more detail. On any individual scorecard they won.But compared across users, their ideas were less distinct from one another. Each person felt they were exploring. Collectively they were converging.Then the researchers found the thing that has stayed with me since I read it.The ChatGPT users felt less responsible for the ideas they had produced.Not less satisfied. Less responsible. Something about the way the idea arrived loosened their grip on owning it.An idea you do not feel responsible for is an idea you will not defend, refine, or stake anything on.The real worldBoth of those are studies. Controlled, careful, and small.So the same group went and looked at the real world.In a preprint released in early 2026, Moon, Kushlev, Green and colleagues examined the personal statements of 372,793 real college applicants. They compared essays written after ChatGPT was released in late 2022 with essays written before it. You can read the preprint here, with the usual caution that it has not yet completed peer review.What they found is the strangest result in this whole field.After ChatGPT arrived, the words in those essays became more varied. The surface got richer. The vocabulary widened.And underneath, the ideas became more alike. Sentence by sentence and essay by essay, the concepts converged.They gave it a name. Semantic disjunction.We are sounding more different from each other than ever, while saying the same thing.Think about what that does to a reader trying to tell people apart.The old signals of effort are gone. Polished prose used to mean somebody cared enough to work at it. Now polish is free and arrives in four seconds.So the surface stops carrying information. And almost nobody notices, because each individual piece looks better than what that person could have made alone.You cannot detect a collective problem from inside your own document.It also happens inside your headThe last study is the one that unsettled me most, because it does not measure the writing at all. It measures the writer. A team at the MIT Media Lab led by Nataliya Kosmyna wired 54 people to an EEG and had them write essays under three conditions: with an AI assistant, with a search engine, or with nothing at all. The paper is called Your Brain on ChatGPT.Brain connectivity scaled down in step with the amount of help. The people writing alone showed the strongest and widest networks. Search engine users sat in the middle. The AI group showed the weakest coupling of the three.Their sense of owning what they had written followed the same ladder. Lowest in the AI group. Highest in the group with no tools.And then a small, terrible detail. Minutes after finishing, most of the AI group could not quote a line from the essay that carried their name.They had produced it. They had not been through it.The researchers called what accumulates cognitive debt. You get the output now and pay for it later, in the thinking you did not do.I want to be careful here. This is a preprint, the sample is small, and other researchers have published a comment asking for the results to be read more conservatively. Treat it as a signal, not a verdict.But note what else that team found. Inside each group, the essays converged. Same names, same phrases, same topics. The homogenising showed up again, in a different lab, measuring something else entirely.Data: Kosmyna et al., MIT Media Lab, arXiv 2025 (preprint)Why nobody noticesThe most uncomfortable finding in this research is not the homogenising. It is that the people doing it could not feel it happening.Writers using the tool reported being satisfied with their work. They were right to be. Judged on its own, each piece was better than what that person would have produced alone.The question that would have caught it is a different question. Not “is this good?” but “is this different from what everyone else is about to publish?”Almost nobody asks the second one, because there is no reason to. Nothing prompts you. The tool does not warn you. Your draft does not look like anybody else’s draft, because you cannot see anybody else’s draft.So the loss happens at a level no individual can observe, to people who are each doing their job well.That is what makes it a system problem rather than a discipline problem. Telling people to try harder will not touch it.Five independent research groups, five methods, one direction. Sources listed at the end.Orwell described this eighty years earlyIn 1946, George Orwell wrote an essay called Politics and the English Language. His argument was not about grammar.It was that bad language corrupts thought, and corrupted thought produces worse language, and the two feed each other in a loop.He listed the enemies. Dead metaphors nobody actually pictures. Long phrases doing the work of a single verb. Pretentious words used as camouflage. The passive voice, which hides who did the thing.And words worn so smooth by use that people can agree on them without agreeing on anything.Then he gave his rules. Never use a long word where a short one will do. Cut any word you can cut. Use the active voice. Prefer plain English to jargon.Now read that list again and think about what a language model produces when you ask it to write something for you.Grammatically perfect. Confident. Full of phrases you have read ten thousand times before. Frictionless, and empty.Orwell described the failure before the machine existed that would mass-produce it.What this does to the price of thingsHere is where it stops being a writing problem and becomes an economic one.For twenty years the rule of being found was simple. Blend in. Match the format. Use the words people search for. Look like the other results so the system knows where to file you.I built an entire business on that rule.In 2009 I started writing and the product was information. Explain a thing well, structure it clearly, and people arrived from all over the world because the explanation was scarce and I had one.That trade is finished. Not fading. Finished.The explanation is now free, instant, personalised and infinite. Nobody needs my clear explanation of anything, and it took me a while to say that out loud.So the rule flipped, and most people have not noticed.The machine now produces the average for free, in unlimited quantity, faster than anyone can read it. Competence has stopped being scarce.And a thing that is not scarce has no price.If everyone is being lifted to the same competent middle, then the middle is where value goes to die. What survives is whatever could not have come from the tool.Not better writing. Different writing. Writing with a person inside it who has been somewhere and paid for the trip.When the machine makes the same for free, the only asset left is the thing it cannot copy.What I keep and what I hand overI am not writing this as somebody who refuses the tool. I use it every week and it has made me faster.But that morning taught me where the line sits, and I have not moved it since.What I hand over: the research. The structure. The charts. The hunt for a study I half remember. The tedious work of turning a mess into an order.What I keep, always:The headline. I write ten before I choose one. The first three are obvious, the next four are worse, and somewhere around eight something honest shows up.The story. Every piece needs a human in it. Mine, or somebody else’s, or a metaphor carrying the weight of one.The point of view. What I actually think, including the parts I might be wrong about in public.The editing. Five to ten passes, cutting the lines I liked most. Stephen King called it killing your babies, and he was not exaggerating.That last one matters more than it sounds. The passes are not decoration. Writing is how I work out what I believe, and if I hand the sentences over I hand over the thinking with them.There is a story about Sartre going blind. People suggested he dictate into a tape recorder instead of writing.He refused. He said that with a machine he would always be either lagging behind it or running ahead of it.He was not being difficult. He knew the machine would set the pace, and that the pace was where the thinking lived.The one testSo here is the diagnostic I now run on everything before it goes out.Could a machine have written this?If yes, delete it. If no, you are winning.It sounds harsh. It is the cheapest quality control available, and it takes four seconds.Look for the parts of a piece that only you could contribute. A decision you regret. A number from your own life. A thing you believed for years and no longer do. A moment you were wrong in front of people who mattered.The machine has read every story ever written. It has never lived one.It has no scars, no bad Tuesday, no company that failed, no room that went quiet when it finished speaking.You have all of those. They used to be the cost of a career. They are now the only part of your work that cannot be generated.The verdictThe research says the same thing three times from three directions.Individual work improves. Collective range narrows. Words diversify while ideas converge. And the effect gets stronger as more people join in.So we are heading toward a world of well-written, well-structured, well-researched pieces that all quietly agree with each other.Nobody chose this. There is no villain in a room somewhere deciding it. There is only a system that pays for volume, so everyone reaches for the same tool, and the tool has a centre of gravity.Which leaves one question worth carrying around.What is one true thing about you that no machine could ever write?Find that. Then put it in the work.It is not a style. It is the only thing you own.This is the ground Zyrro is built on. It reads your own history back to you — the decisions, the failures, the questions you keep returning to — so the thing only you can say stops being invisible to you. The early list is open.SourcesAnil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances, 2024Kibum Moon, Adam E. Green and Kostadin Kushlev, “Homogenizing effect of large language models (LLMs) on creative diversity”Moon, Kushlev, Green et al., “The Link Between Diverse Words and Original Ideas Is Weakening in the AI-Era College Admissions” (preprint, 2026)Barrett R. Anderson, Jash Hemant Shah and Max Kreminski, “Homogenization Effects of Large Language Models on Human Creative Ideation,” Creativity & Cognition, 2024Nataliya Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt,” MIT Media Lab, arXiv, 2025 (preprint)Miloš Stanković et al., comment on “Your Brain on ChatGPT” (2026)George Orwell, “Politics and the English Language,” Horizon, 1946.Janet Emig, “Writing as a Mode of Learning,” College Composition and Communication, 1977.The post AI Made Everyone Better. Originality Became Scarce. appeared first on jeffbullas.com.