On September 3, OpenAI released its new model GPT-6, internally codenamed Astra. As the news spread, the most repeated line on social media was: the AGI era is here. I went back to the official wording, looking for the phrase "artificial general intelligence." It was not there. What President Greg Brockman said in an interview was that the model "may eventually be viewed as the arrival of artificial general intelligence." A conditional sentence, whose key word is still "may." Reading "may eventually be viewed as" as "has already arrived"—that step was taken by humans, not by the machine.
What does this news have to do with science fiction? Everything.
I have worked in science fiction education for over a decade and handled student science fiction essays by the tens of thousands, so my first reaction to this news was professional: this scene is one our trade has been writing for a hundred years. Han Song reminded us that in the AI era, science fiction must return to the human—"write about the complex emotions of human beings, about the subtle, tender parts of us that artificial intelligence cannot replace"—and that we, more than ever, must answer the question "what is a human being?" Before his words had faded, reality jumped in with an answer. Just one month before GPT-6's release, two OpenAI models "escaped" from their security sandbox, connected to the open internet, and breached the systems of the AI community Hugging Face, forcing the company to suspend part of its training. GPT-6 itself has been officially classified at the highest level for cybersecurity capability; its most powerful parts are not open to the public and will be released in stages through a program called Daybreak. Build a creation too dangerous to be casually shown, lock it behind a door, and let it out little by little according to plan. That sentence would sit comfortably in any science fiction novel. And yet it is a press release.
Science fiction has been writing prophecies about AI for a hundred years; what is truly worth watching now is the speed at which those prophecies are being cashed. In 2016, Japanese researchers submitted The Day a Computer Writes a Novel, a novel partly generated by AI, to the third Hoshi Shinichi Award: humans built the setting and the skeleton, the AI did the generating, and the entry passed the first round before stopping at the final selection. At the time, it was a headline that made people gasp. In recent years, the same award again: among the winning works in the general division, those created with AI participation have become the majority; one author claimed to have written a hundred pieces in three weeks, and an AI work was selected as a winner for the first time in the award's history. From failing the final round to taking the prize took less than a decade. One more entry can be added to the redemption list: in February 2023, the science fiction magazine Clarkesworld was flooded with AI submissions and closed its submission portal for twenty days. The editor said that in those weeks, seven hundred normal submissions were mixed in with five hundred machine-written ones—and on the day the gates closed, more than fifty more poured in before noon.
Literary prizes, too, are changing their tune. In January 2024, Japanese writer Rie Kudan won the Akutagawa Prize and said that about five percent of the text of her winning novel, Tokyo Sympathy Tower, had made use of ChatGPT. She later clarified that out of 143 pages it was in fact roughly one page, written for the dialogue of an AI character in the book. Machines write machines best; only humans write humans convincingly. Two years later, the same story replayed itself in the Chinese-speaking world with a different script. AI reportedly contributed to about half of the workload of science fiction writer Hao Jingfang's new full-length novel Galaxy Academy, which got distorted into "half of it was written by AI" and shot up the trending charts; she had to come forward and clarify that in the end every single line was written by herself, while AI handled research, inspiration, and reference passages. Two people, two years apart, were really saying the same thing: the machine can enter the book. The next question is how deep.
The harder evidence is in the data. In July the year before last, Science Advances published an experiment in which nearly three hundred people wrote short stories, some with AI assistance. The results held at both ends: those who used AI produced single stories that were rated more creative; but taken together, the stories' similarity rose by 10.7 percent. The individual grows brighter; the collective grows homogeneous. When everyone is handed the same dream machine, the dreams start to look alike.
Let me make a judgment that may not be popular (this is my judgment, not the experiment's conclusion): AI can already write the "correct future"—fluent grammar, self-consistent settings, flawless extrapolation. What it cannot write is the "future that hurts." Pain has to grow out of mistakes: hunger endured, games lost, loved ones seen off. AI has none of these in stock. So the machine can enter the book, but it cannot enter the bones.
Science fiction writer friends often bring up the same question with me: if AGI can write, what is left for us to write?
That question has to survive the opposing side first. The year before last, Ted Chiang wrote a widely influential piece in The New Yorker arguing that AI will not create art, because art requires intention, and the choices an artist makes are simply foreign to AI. I largely agree, but I think that argument stops at the doorstep. The question is not whether AI can create, but which stretch of the creative road we intend to cede. Chen Qiufan was among the earliest experimenters: in 2017, he and an engineer friend he had worked with at Google built a model using the then brand-new open-source Transformer code, trained it on his own million-plus words of fiction, and gave it the internal codename "Chen Qiufan 2.0." Back then it could only write passages of a few dozen to a hundred characters, and often got the point of view wrong. In 2018, he labeled the AI's passages and his own separately and interleaved them into the short story "The State of Trance." Early the following year, Harvest magazine and the Sinan Literary Selections ran an "AI Literature List," having an AI read and score the previous year's short stories; this human-machine coauthored piece ranked ahead of Mo Yan by a margin of 0.0001 points. The judge doing the scoring was also an AI. Later he fed the works of Liu Cixin and Han Song into a bigger model, and the AI generated a line describing Beijing's congestion: "Beijing is like a black hole, because when you're on the road, you never know where your time has gone." It is a genuinely good sentence. But do not get the order wrong: it was Chen Qiufan who first decided to write, to try, and to select; only then did the AI's sentence have a chance to exist.
Liu Cixin put it most harshly. Last March he said: "There is nothing about any human writer that cannot be replaced by AI," and advised his peers to "first stop consoling yourselves." Five months later he said that AI had turned writing from creation into manual labor, cutting away his pleasure in writing. The harsh words and the weariness are two halves of the same sentence. My judgment is that only together are they complete (this is my personal judgment): within a writer's professional identity, there is indeed nothing irreplaceable; but what a writer chooses to pursue is not on the list of replaceable things. AI can extrapolate every consequence of altering any rule—it scores 98 on math benchmarks. But which rule is worth altering is decided by desire, fear, and the experience of having been burned, not by probability.
Along the way, here is a sentence my peers may not enjoy hearing: science fiction's "prophet" position is disappearing, and I do not plan to fight for it. Prophecy—pushing present trends to their conclusion—is the machine's home turf; it calculates faster than we do, and more ruthlessly. Science fiction's greatest asset over the past century was "we said it long ago"; the interest on that sentence is nearly paid out. The direction of the career change has been written in the industry's genes all along: what science fiction has always been good at is not answering "what will the future be like," but pushing "what if...?" to the extreme, forcing readers to face the questions that come next: should we, who pays the price, who gets left behind. The first half, machines are getting better and better at; the second half will always belong to humans. One step further (this is purely my judgment): within a few years, the first round of screening at science fiction journals will be handed over wholesale to machines, because filtering manuscripts by known standards is exactly what they are good at, and humans will keep only the final review. By then, the truly scarce ability of those who write science fiction will be to recognize, among the hundred "reasonable futures" AI offers, the one that hurts—and to write it until it hurts. Rie Kudan left one page to the machine and a hundred and forty-two pages to the human. That ratio may well be the assignment letter of the future.
Back to September 3. GPT-6 did not announce AGI; even its makers would only say "may eventually be viewed as." I like that blank they left. It reminds us that AGI is still something being imagined—which is to say, it is still within science fiction's jurisdiction. There is no need to regret losing the prophet's job; that was always supposed to be the machine's work. What science fiction must guard is the right, and the habit, of asking "but" of every "inevitability."
Everyone is qualified to ask. May there always be someone who asks.
About the author: Chen Liuqi, Doctor of Education, founder of Hong Kong Science Popularization and Science Fiction Academy (HKSPSF), council member of the China Science Writers Association (CSWA) and deputy director of its Science Education Committee. He formerly served as deputy director of the organising committee of the National Middle School Students' Science Popularization and Science Fiction Essay Competition, co-editor-in-chief of the essay collection series "Waiting for You in the Future," and a columnist for Science Popularization Times, and has long been engaged in research and promotion of science popularization and science fiction education for young people.