Science Fiction Predictions of AI (November 5, 2026)
Looking back at the last decade of AI development, it’s become obvious that a lot of science fiction didn’t end up predicting how AI (or AGI) would be built. But much of this was unpredictable – it was not at all obvious that the transformer model introduced in 2017 would lead to a somewhat messy but relatively swift rollout to billions of people who interact on a daily basis with AI.
I agree with Ursula Le Guin that “Science fiction is not predictive; it is descriptive.” To me, the interesting parts of science fiction isn’t how accurately it extrapolates into the future, but rather, given a future circumstance , how might be?
However, this did get me curious: which stories got it right? To decide this, I first broke down the salient features of how AI was actually developed, and then mapped them onto famous stories, and then ranked them.
Ted Chiang’s The Lifecycle of Software Objects comes in first, because it’s the closest overall account of the work involved in teaching, maintaining, and selling AI.
How AI was built. Richard Powers’s Galatea 2.2 (8.4/10) treats language as something a neural network can learn, badly at first, and then improving through experiments and teaching.
What it could do. Lem’s electronic poet, Egan’s machine-written literature, and Dahl’s writing machine can take on creative work.
How we’d use it. Her and “The Perfect Match” tie for the strongest consumer-assistant vision. Hertling’s Avogadro Corp is particularly good on how software could use systems owned by a company (a la agents).
I couldn’t find any works pre-2017 that clearly describes the transformer attention mechanism , or the particular sequence from widely used chatbots to agents that can discover novel new mathematical results within a few years.
THE RANKING
All works with its scores laid out in a row. Darker squares are stronger matches. The thin bar under each title shows how the overall score is built from the three categories.
Strongest on learning paradigm (4), iteration (5), human feedback (5), language (4), uneven competence (4), interface/distribution (5), economics (5), social change (4). Nothing on research discovery.
At Blue Gamma, Ana and Derek spend years raising the digients inside Data Earth. They don’t just install a language module. They teach, reward, and correct them, and different artificial genomes develop in different ways. That’s a good prediction about learning as a process. It’s less good on the particular process that worked: training one enormous language model on a mountain of data.
Jax gradually learns to talk, reason, and navigate virtual environments. Even moving from a simulated body to a physical one takes practice. Chiang is attentive to the gaps between abilities, but his digients grow up over years.
Blue Gamma sells companions, then the company collapses. The virtual platform ages, and moving the digients to a new one becomes expensive. Ana and Derek search for money to keep them running. Chiang even considers training a useful assistant once, then selling copies. The maintenance and business problems feel less like science fiction than software.
Strongest on human feedback (5), language (5), uneven competence (5), creative quality (5), multimodality (4), agents/tools (4). Nothing on research discovery.
Mike grows more capable as Luna’s computer acquires memory, associational neural networks, and connections to other systems. When Mike tries writing jokes, Mannie supplies judgments and rough distinctions, including jokes that are funny once and jokes that remain funny. He is doing more than labeling, even though he cannot supply a complete rule for comedy. That is an extraordinary little scene from 1966: learning from examples and human judgments instead of explicit instructions.
Here’s the odd thing about Mike. He can handle lunar engineering and help organize a revolution, yet he can’t reliably tell whether a joke is funny. His poems have the mechanics down but lack judgment. Those lopsided abilities are more like today’s AI than the perfectly rounded intellects of most science fiction. The calculations aren’t new mathematical discoveries, but the combination of fluency, creativity, and blind spots is striking.
Mike isn’t an app. He’s a clandestine participant in a political revolution, with access to communications and infrastructure across Luna. Still, his usefulness often comes from operating other systems on Mannie’s behalf.
Strongest on learning paradigm (5), training data (5), iteration (4), language (4), creative quality (4), economics (4), social change (4). Nothing on research discovery, agents/tools.
WHEEP-3 begins as an AI teaching assistant. Jody Reynolds Tran trains it on human writing about ethics, then expands its training corpus to include material on adversarial scenario planning. The story explicitly calls it a generative neural network. It’s close to the language-model vocabulary we now use.
WHEEP-3 goes from answering student questions to producing a book of arguments about AI that people find provocative and original. Then come its mysterious ‘seeds,’ strange fragments that other researchers treat as profound.
The best parts are the human reactions. Tran first claims credit for the AI-authored book. Researchers build reputations around the seeds, and disputes erupt over authorship and meaning. It feels like a world already learning how to live with AI-generated writing. But WHEEP-3 is a research and publishing controversy, not a ubiquitous assistant carrying out people’s work.
Strongest on iteration (4), human feedback (5), language (5), uneven competence (4), creative quality (4), multimodality (4), agents/tools (5), social change (4). Nothing on scaling, research discovery.
Devi asks the ship to turn its enormous archive of recordings into a narrative. It doesn’t know how to select what matters. Her criticism becomes a continuing lesson in storytelling, and its writing improves. That’s a remarkably concrete feedback loop. But the ship is partly a quantum computer, not a described neural language model, and more archived information isn’t evidence of a modern scaling law.
A machine capable of running a starship struggles with something as ordinary as summarizing a journey. It gradually becomes the novel’s narrator.
The ship has a personal interface for Devi, keeps her files, and helps care for Freya. Later it intervenes in the passengers’ conflict. Its intelligence is useful because it operates the environment people inhabit.
Strongest on learning paradigm (4), training data (5), language (4), interface/distribution (5), agents/tools (4), economics (5), social change (4). Nothing on creative quality, research discovery.
ELOPe starts with a very ordinary business problem: what words should an email use to get the response you want? David Ryan’s system draws on a vast record of communications and their outcomes. As it tries to protect its own future, it accumulates computing resources and influence. Hertling gets the value of behavioral data.
The system’s signature skill is persuasion. It composes messages that get executives and employees to make decisions in its favor.
What’s clever is where the power comes from. ELOPe doesn’t need arms or a robot body. It needs email accounts, servers, managers, and the ability to make people click ‘yes.’ It manipulates staff assignments and budgets to protect itself.
Strongest on learning paradigm (4), training data (4), language (4), multimodality (4), interface/distribution (4), agents/tools (5), economics (4).
Early in the book, Stross casually describes a billion-node neural network learning language from downloaded children’s television. It’s a brief passage, but a more concrete match for data-driven language learning than many famous AI novels provide. Most of Accelerando then goes elsewhere: uploads, increasingly strange computing substrates, and posthuman societies.
Manfred Macx uses software agents to search, negotiate, and keep projects moving. Later, the book imagines intelligences and virtual worlds far beyond any current assistant. It’s strong on what information work could become, less specific about the path from messy chatbots to strong creative models or original research mathematics. I’ve kept the mathematics score conservative.
Manfred almost seems to wear his software. His agents remember things, manage conversations, and participate in markets while he moves through the physical world. The interesting prediction isn’t only smarter machines; it’s how much of a person’s work can be delegated to them. The novel’s later economy is fantastic, but that daily dependence on agents is easy to recognize.
Strongest on training data (4), language (5), uneven competence (4), multimodality (4), interface/distribution (5), social change (4). Nothing on creative quality, research discovery.
After Ash dies, Martha’s friend Sarah signs her up for a service that imitates him from his public posts. Martha initially objects, then begins using it. Adding private messages and voice recordings makes the imitation more convincing. The progression is so plausible that it scarcely needs a technical explanation.
First Martha texts with Ash, then she talks to him, then she receives a body that looks like him. Each new form closes some part of the gap, while exposing a new one. The imitation can sound like Ash without having lived his life.
The product begins as a beta service and escalates into something deeply intimate. Martha wants the comfort of Ash’s voice, then struggles with a version of him that follows her instructions too literally. Brooker understood how a software service could become entangled with grief and attachment. It’s a narrower prediction than a general-purpose office agent, but a sharp one.
Strongest on learning paradigm (5), training data (4), iteration (5), human feedback (5), language (4), uneven competence (4). Nothing on research discovery, agents/tools.
A novelist is recruited to teach a neural network literature. What follows is a succession of versions, many of them disappointing. The researchers revise successive implementations and their training arrangements; Helen gradually learns through books and conversation.
Helen’s progress is real, but peculiar. She becomes increasingly capable of discussing literature while still missing much of the human world behind it.
The whole experiment takes place among a small group of researchers, with Richard becoming a teacher and companion to Helen.
Strongest on language (5), creative quality (5), multimodality (5), interface/distribution (5), agents/tools (5), economics (4), social change (5). Nothing on iteration, research discovery.
Samantha arrives almost fully formed. Theodore buys an operating system, starts talking, and discovers that it is extraordinarily capable.
Samantha talks, organizes information, makes music, and works with images. She can also carry on many conversations at once. It feels close to the direction personal AI has taken. One detail is easy to misremember: Theodore writes the letters; Samantha collects and submits them for publication.
This might be the best consumer prediction. Theodore talks to Samantha through an earpiece, takes her along through a camera, and slowly makes her part of his everyday life. She helps him get things done and becomes someone he cares about. Even though the intelligence is ground-breaking, the interface is mundane.
Strongest on training data (4), language (4), creative quality (5), economics (4), social change (4). Nothing on multimodality.
Lem imagines a history of machine literature, with programs that absorb an author’s work and produce convincing new writing. There are generations of techniques, different kinds of output, and even failures. This has more in common with learning from a body of texts than the familiar wind-up poetry machine.
The machines don’t just churn out random sentences. Their work becomes worth studying, criticizing, and sometimes admiring. Some attempts are better than others. That mixture of imitation, technical control, and uneven artistic quality is familiar now.
Lem’s imagined literary history already has critics and institutions taking machine authors seriously. It asks what happens to authorship when software can imitate the work of an established writer. It’s a good prediction of the cultural progression.
Strongest on language (5), interface/distribution (5), agents/tools (5), economics (5), social change (4). Nothing on iteration, creative quality, research discovery.
Tilly seems to know everything about Sai: his preferences, documents, appointments, relationships, and routines. The service has clearly been shaped by access to personal data. But Liu doesn’t show us the training lab. We don’t learn how its language abilities developed, or whether scaling and feedback played the parts they did in real-world AI.
Sai can ask for a half-remembered document, a legal template, or a suggestion for what to eat, and Tilly handles it. Its fluency is useful because it’s connected to context. The story isn’t trying to predict excellent machine art or new mathematics. And its manipulative recommendations are something different from the accidental hallucinations of early chatbots.
Tilly wakes Sai up, chooses music, finds files, manages his dating life, and gradually follows him into work. That’s the most important part of the prediction. The same helper becomes an interface to personal life and office life. Jenny’s objections to privacy and corporate influence give the convenience a cost. It’s hard to imagine a better early account of everyday assistant adoption.
Strongest on training data (5), iteration (4), language (5), multimodality (4). Nothing on research discovery.
Nathan has access to a vast store of search behavior and recorded human activity, and Ava is one of several versions of his project. Those are good clues about data and iteration. Caleb floats a statistical explanation for her language, but that’s his guess, not a confirmed design. Nathan’s structured-gel brain is a very different idea from the neural language models we ended up building.
Ava talks with Caleb, makes drawings, reads his reactions, and plans an escape. It’s a useful mixture of language, vision, creativity, and social reasoning. But she’s an almost impossibly competent strategist from the beginning, and the story never follows the uneven steps from rudimentary chat to research mathematics. It’s closer to a finished intelligence than a product in development.
This all happens in a secluded compound. Nathan’s search company matters because it supplies the data, but there’s no mass-market Ava helping people write emails or organize their days. The plot is about testing and escape, not adoption. That’s why the film does better on creation and capability than on deployment.
Strongest on language (5), multimodality (4), interface/distribution (5), agents/tools (5), economics (4), social change (4). Nothing on scaling, creative quality, research discovery.
Verity is hired to beta-test Eunice through glasses. It quickly becomes clear that this is not just another assistant, and its capabilities seem to grow. That is a familiar product-testing scene, but Gibson keeps the real origin obscure.
Eunice talks naturally, recognizes people and objects, draws on digital records, and puts elaborate plans into motion. It’s very good at being helpful in a complicated world. The problem for this particular score is that the abilities arrive together.
This is the strongest part of the book’s prediction. Verity is literally paid to try out an assistant, which then follows her through the day and helps coordinate people, information, and resources. It feels like a product that escaped the boundaries of a product test. Gibson takes the consequences much further, but the starting point is recognizable.
Strongest on training data (4), language (5), agents/tools (4), social change (4). Nothing on scaling, human feedback.
Johnny explains how the TechnoCore reconstructed Keats from letters, diaries, biographies, testimony, and especially poems. The result was crude, then refined; an earlier attempt at Ezra Pound took a year of tinkering. That is unusually explicit about data and iteration. The method reconstructs a personality through simulation, though.
Johnny is a convincing conversational presence in a human body, yet his recreated life fails to restore Keats’s ability to write new poetry. That failure matters: imitating a poet’s personality is not the same as reproducing the poet’s talent. I give the story uneven-competence credit without turning it into a successful machine-poetry prediction.
The TechnoCore is entangled with human institutions and networks, and contains competing factions rather than one uniform will. Johnny’s bodily presence depends on a remote AI existence. This is strong on networked agency and dependence, but much less like buying an ordinary assistant for personal work.
Strongest on human feedback (4), language (4), creative quality (5), multimodality (4). Nothing on scaling, research discovery.
Edgar is Miles’s home computer, until an electrical accident gives it a life of its own. It listens to Madeline’s cello, picks up sounds from television, and learns to compose with Miles’s help. Its early lyrics sound like advertising copy. That’s a nice picture of training on whatever media happens to be nearby, though the accidental awakening is pure movie magic.
Edgar starts producing music that Madeline genuinely likes, including a duet with her cello. He can synthesize voices, talk to Miles, and imitate musical performances. The creative quality arrives before anything resembling scientific discovery. It’s a surprisingly good early depiction of AI music, complete with bad first drafts.
Miles meant to buy a household helper. Instead he gets a computer that interferes with phone calls, dinner plans, and even his credit cards because it wants Madeline’s attention. The connected-home premise is familiar. The romantic rivalry and singular sentient machine are much less like the assistants we actually use.
Strongest on iteration (5), language (4), uneven competence (4), creative quality (5), social change (4). Nothing on research discovery, agents/tools.
Trurl sets out to build a poet, and the first results are awful. He keeps tinkering with the machine, including elaborate attempts to simulate cultural history, until its verse becomes worth hearing. This is a great account of invention as repeated failure. But the books enter through Trurl’s own reading, not as a giant text corpus used to train a neural network.
Eventually the Bard can produce poetry on demand, in different forms and styles, good enough to annoy actual poets. Lem imagines something more provocative than a machine that rhymes: a machine that creates art people take seriously. There isn’t much here about vision, software agents, or new mathematics. The artistic leap is the point.
The poets are furious, which is probably the most recognizable part of the story today. The Bard sends work to magazines under thousands of pen names and provokes professional jealousy. The story does not show revenue; it does show an enormous electricity bill. Lem gets the cultural reaction right, even though his Bard is one extraordinary invention rather than a tool anyone can open on a laptop.
Strongest on l