A Man, a Mind and a New World

I'm going to tell you a story about a boy. A boy born in Florida on November 21, 1998, to a mother and father who both worked within the technology sector. His name was Suchir Balaji, and he grew up in Cupertino, California, home to Apple, the company that made the device you're probably reading this on. You know in the movies when there's a child genius who seems to obtain an unfathomable amount of intelligence and capability at an extremely young age? Well, that was Balaji. He started coding at 11, built his own computer as a kid, and by 14 had written a paper about chip design. And as a genius does, he went on to compete, and win, in programming and machine learning competitions throughout high school before studying computer science at Berkeley. By his early twenties, Balaji was recruited to OpenAI, where one of his first major projects was WebGPT, an early system that helped pave the way for what would eventually become ChatGPT. John Schulman, an OpenAI co-founder who recruited him, later described Balaji's contribution to that project as "essential," saying it would not have succeeded without him. Balaji later worked on organising some of the enormous datasets used to train GPT-4 and eventually worked on ChatGPT itself.

So, just to clarify, at around the age I am now, Balaji was helping build the technology that would become one of the most recognisable products in the world.

By August 2024, after nearly four years at OpenAI, something had changed. Balaji had spent his early twenties helping develop the technology that was now spreading across the world, but the more he thought about how it had been built and what it was becoming, the less comfortable he was with his own part in it. So, at 25, he left. Not for another company, not for a better title, but because he no longer believed he could keep contributing to something he had come to disagree with. As he would later put it, "If you believe what I believe, you have to just leave the company."

Two months later, on October 23, 2024, Balaji went public. The New York Times published an interview with him in which he argued that products like ChatGPT were violating U.S. copyright law, and on the same day his own detailed analysis of the issue appeared online. His concern was essentially this: companies were building enormously valuable products using work created by other people and businesses, only for those products to then become capable of competing with the very people whose work had helped make them possible. This was no longer a private disagreement with a former employer; Balaji had attached his name, his reputation and his own work at OpenAI to the argument. At the same time, he was beginning to work out what came next, planning a nonprofit focused on machine learning and neuroscience and speaking to potential investors about building something of his own. He had left OpenAI only months earlier, had just publicly challenged the foundations of the technology he had helped develop, and was moving forward with whatever came after it.

Then came November. OpenAI and Microsoft were already fighting one of the biggest copyright cases of the AI era, and on November 18, 2024, Balaji's name entered it. Lawyers for The New York Times identified him as one of several people who may hold documents relevant to the case. He had worked with the datasets at the centre of the dispute, had already spoken publicly against OpenAI, and now his knowledge could matter in court. Four days later, on November 22, Balaji's parents heard from their son for the last time. He stopped responding to their messages and, increasingly worried, they eventually asked San Francisco police to conduct a wellbeing check. On November 26, police entered his apartment and found him dead. He was 26 years old.

San Francisco authorities concluded that Balaji had taken his own life. His parents did not accept it. They hired a private investigator, commissioned a second autopsy, and said the findings did not match what they had been told. They pointed to the state of his apartment, to what they believed was evidence of a struggle, and to the absence of a note. They have been publicly arguing ever since that their son was killed. The police and the Office of the Chief Medical Examiner have not moved.

Here is what I think, for whatever it's worth. I find the timing hard to sit with. Four days between the New York Times naming him as someone who might hold relevant documents and the last time his parents heard from him. He had left OpenAI, gone public against it, and was building something new. That isn't the shape of someone winding down.

I also know what argues against me. The gun was purchased in his name months earlier. The second autopsy findings were never released in full, only characterised by the people who paid for them. A private investigator hired by grieving parents is not a neutral finder of fact. People who look like they are moving forward die by suicide all the time, and the fact that it doesn't fit our idea of what that looks like is not evidence of anything. Coincidences that feel unbearable happen constantly, which is the whole problem with coincidences.

So I'm not telling you he was killed. I'm telling you I'm uneasy, that I've looked hard at the reasons I shouldn't be, and that I'm still uneasy. That's as far as I can honestly take it.

What I do know is that his death is what first pulled me deeper into his story, and somewhere along the way I realised I was learning far more about how he died than about what he had been trying to say while he was alive. Had I come across an essay about copyright law, machine learning and low-entropy model outputs, I probably would not have stopped for very long. The story of his death became more interesting than the story he was trying to tell, and the mystery became easier to consume than the argument. I don't think there is anything wrong with being interested in what happened to Suchir Balaji. I clearly am. A 26-year-old dying suddenly, under circumstances his own family continues to question, is always going to command more immediate attention than a technical argument about copyright law. But I think the problem comes when that curiosity becomes the end of the story. Balaji spent the final months of his life deliberately putting an argument into the world, under his own name, because he believed it mattered. His death may be what draws people to him, but I don't think it should be the last thing they learn about him.

So rather than leave his story as the mystery surrounding his death, I wanted to understand what he was actually trying to say. What had made him walk away from the work he had helped build, and what did he think the rest of us needed to understand? In honour of him, I want to explain that to you.

His argument, underneath all the technical language, was actually fairly easy to understand. The internet is full of things made by people: journalism, books, photography, research, commentary, code, art, businesses, forums, websites and millions of people explaining something they know to someone who does not. Generative AI became possible by learning from an extraordinary amount of that material, and the legal question is whether companies are entitled to use copyrighted work in that process under fair use.

But the part of his argument that actually mattered, and the part almost nobody repeated, was about what comes out the other end. Everyone was arguing about whether it's legal to train on someone's work. Balaji was more interested in the output. When you ask ChatGPT what happened in an investigation the Times spent months on, it doesn't invent an answer. It reconstructs one. And the more that reconstruction is pinned down by what the model read, the fewer ways there are to say the thing correctly, the more of the original ends up in the answer. Balaji's term for this was low entropy. Low entropy means the model isn't really composing, it's retrieving with extra steps. His point was that at some level of closeness, you stop being able to call that a new work. It's the old one, wearing a different sentence.

Which is where it starts to matter for the rest of us. If I ask ChatGPT a question instead of clicking the journalist's article, reading the researcher's paper, visiting the website or paying the person who originally produced the information, the convenience comes with a bigger economic question underneath it: what happens to the conditions that allowed that information to exist in the first place?

The paradox is: the better the AI becomes at substituting for the source, the less reason someone may have to go to the source; but those sources are part of what allowed the AI to become useful in the first place.

That question is still unresolved, in 2026. Nearly two years after Balaji's death, The New York Times and OpenAI are in front of Judge Sidney Stein on summary judgment, arguing the fair use question directly. The law has moved, though not as cleanly as the headlines suggested. In June 2025, two federal judges in California found within two days of each other that training a model on copyrighted books was transformative enough to count as fair use. But both rulings were narrower than they looked. One judge split his decision, finding that downloading millions of pirated books to build the training library was not fair use at all; that case settled months later. The other told the authors they had lost because they had argued the wrong thing, and then spent pages explaining the argument that would have worked — that these systems flood the market with substitutes, crowding out lesser-known writers, and that this alone would likely decide future cases. He was effectively handing the next plaintiffs a map. This September, the Justice Department filed a statement urging the court in the Times case to treat training-stage copying as fair use. What remains genuinely open is the narrower thing Balaji was pointing at: not whether you can read the work, but what you're allowed to hand back.

I cannot tell you what Suchir Balaji would have wanted people to think about his death, but I can tell you what he wanted people to think about while he was alive because he wrote it down. What I find most interesting about him is not simply that he criticised OpenAI, but that he changed his mind about something he had personally helped build. By his own account, he had entered the industry with an assumption that was common within it: that publicly available material on the internet could be used to train these systems. But after ChatGPT launched, and he saw what happened when the technology moved from a research problem into something used at enormous scale, he came to believe that assumption was wrong. It is easy to recognise a problem you had no hand in creating, and much harder to recognise one when doing so forces you to look at your own work and possibly implicate yourself. Balaji did. And rather than secretly keeping the discomfort to himself, he left the company, put his name behind his argument and opened that argument to everyone else to pull apart.

Suchir Balaji was undeniably gifted. He had the kind of intelligence that revealed itself early; he could understand things most people could not, and he could build things most people never could. He didn't let his intelligence or achievements make him rigid or without principle. He turned the same eye he had for complex systems onto his own work, and I think that was the truest mark of his intellect. Because intelligence is not only the ability to know more; sometimes, its truest form is the ability to question what you already believe, to look honestly at your own work, and to change your mind without seeing that change as weakness. Balaji did that.

At only 26, Suchir Balaji had been part of developing some of the most consequential technology in the history of humanity. At only 26, he also achieved something far greater: character. Intelligence may get you to the top, or close to it; character decides who you are once you get there, and who you are when you leave it.

In memory of Suchir Balaji, 1998–2024.