Father of VR: The best AI future nobody is talking about

ELI5 / TLDR

Jaron Lanier thinks “artificial intelligence” is a bad way to describe what these systems actually are. An AI model is not a new alien mind — it’s a giant pile of human work (everyone’s writing, code, photos, driving) blended together. If you call it a mind, you start treating it like a god, and the people building it go a little crazy. If you call it what it is — a collaboration of millions of people — then the people whose work feeds it deserve to be seen and paid, and a much nicer future opens up where humans stay valuable instead of becoming “obsolete.”

The Full Story

The word “AI” is a marketing trick, not a description

Lanier’s central move is small but everything follows from it. He says the term “artificial intelligence” was coined in the late 1950s not to describe a discovery, but to win an argument. A group of researchers — including his own mentor Marvin Minsky — wanted funding and prestige, and they wanted to outshine an older scientist named Norbert Wiener.

Wiener’s framing (he called it “cybernetics”) was that a computer is never a thing by itself. It’s always wired into the world: it changes the world, and the world changes it back, in a loop. Think of it like a thermostat — pointless to describe a thermostat without the room it’s heating. The “AI” crowd threw that away. They declared the computer a self-contained box with its own intelligence.

“You declare the computer a box in its own right. You give it its own reality because it’s intelligent rather than just being part of this interactive flow. So it in a way it’s a step backward.”

Lanier argues this framing isn’t just inaccurate — it actively breaks your ability to think clearly about what the machine is doing.

A model is “a box of people”

Here’s the reframe. A large language model — the thing behind ChatGPT — is built entirely from human-made data. So instead of “an alien intelligence,” picture a giant Wikipedia: everybody’s contributions mashed together. Nothing in the box that didn’t come from a person.

“AI is a pile of [nothing] without us. It is nothing. There’s only people.”

He’s careful to say this isn’t a scientific claim — both descriptions are technically identical, you lose no accuracy either way. It’s a choice of framing. But the framing has enormous downstream consequences for how sane the whole project stays.

When the host tries the standard counter — but what if AI becomes recursively self-improving and takes off on its own? — Lanier cuts him off. Wrong question. The right question is whether human collaboration can compound on itself. And obviously it can: we learned to talk, then write, then print, then broadcast, then edit digitally — each one folding back and making the next leap possible. An AI model is just the newest layer of that. The improvement is real. It’s just ours, not the machine’s.

Why the framing makes the tools work better

This isn’t only philosophy. Lanier claims the “box of people” view makes you a better user of the tools.

If you believe the model is a budding super-mind, you expect it to eventually stop hallucinating and fix its own security holes — so you wait, and you forgive failures with “next year’s model will be smarter.” If you understand it as a compressed average of human work, its behavior stops being mysterious. It’s great at common tasks (build me a simple website — done, because the internet is full of those) and shaky on rare ones (unusual projects with little precedent data fail more often). That’s exactly what you’d predict from a thing made by blending what humans have already done.

“It’s not an alien angel that can synthesize every possible world. It’s a combiner of what’s happened.”

He’s genuinely impressed by recent coding gains — he thinks code has “always sucked” and better tools are overdue — but insists none of it would exist without GitHub and the humans who wrote all that code first.

Data dignity, and a clever safety trick

So what do we do about it? Lanier’s proposal is “data dignity”: pay people for the data their work contributes to a model, on a market basis — actual demand for actual use, not a flat handout.

To make the idea concrete he offers a safety example. Imagine a cornered criminal pointing a phone at his kitchen and asking a model how to build a bomb from what’s there. Guardrails catch the obvious version, but not always. Now imagine the model, while answering, reaches back into its training data and asks: which clusters of source material, if removed, would most change this answer? Kitchens and ingredients, sure — but almost certainly also a cluster about bombs. That “trace back to the humans” check becomes a second, independent signal.

He compares it to two-factor login. The reason a code texted to your phone works is that it’s a separate channel — harder to fool. Tracing an output back to its human sources gives you “multifactor AI”: a second model running alongside the first, semantically grounding what’s happening by pointing at the people behind it. Brains, he notes, aren’t one giant lobe — they’re competing parts that check each other. Same principle.

Why he hates UBI

The host raises universal basic income as the usual humane fix. Lanier rejects it, and his reason is structural rather than moral.

“You might start with Bolsheviks, but you’ll end up with Stalinists.”

His argument: any system that funnels everyone’s income through one mechanism creates a central control point — and central points get seized, eventually by the worst available actors (an autocrat, a theocracy, a criminal org). Even “decentralized” tech ends up propping up a Google, because network effects always grow a few super-powerful nodes. A market — messy, distributed, combined with other social layers like the Scandinavian and American models do in different mixes — keeps power “confused,” which he says is the only way to keep power from turning toxic. Paying people per use of their data spreads the spoils instead of pooling them.

He’s lived this tension: as both an Author’s Guild board member and a Microsoft scientist, he got deposed by both sides of the Authors-vs-OpenAI/Microsoft lawsuit and had to disagree with both. The class-action remedy (everyone gets a tiny equal slice) is, to him, a sneaky step toward UBI.

The future nobody pitches

The standard AI futures are binary: it kills us, or it keeps us as pets. Lanier distrusts any scenario with no wiggle room — brittle predictions rarely come true. His alternative: an ever-expanding sphere of new creative roles we can’t yet imagine, where people invent new things to do, their contributions feed future models, and they get paid for it. The beauty is it doesn’t need to fully succeed — even at 10% or 30%, it’s still good.

He wants the future radical and continuous — great-grandchildren morphing their bodies and crossing the stars, but with an unbroken line of memory back to now.

“I don’t want singularity. A singularity by definition is dementia.”

He closes on grounded optimism: humanity survived fascism, communism, plagues, and tech disasters. The trend line, he thinks, bends toward fewer people in misery. And he flips the optimist/pessimist label — he’s the optimist (letting his car drive him, betting on human creativity); the people warning AI might kill everyone are the pessimists who’ve simply redefined “optimism” as “we’re about to build a god.”

Key Takeaways

The term “artificial intelligence” was coined in the late 1950s as a funding-and-prestige play against Norbert Wiener’s “cybernetics,” not as a neutral description.
Wiener’s view: a computer is never a standalone box — it’s locked in a feedback loop with the world, steering it and being steered.
A large language model contains only human-made data, so it can be described equally accurately as “a collaboration of people” rather than “an alien intelligence.” Both framings are technically identical; only the consequences differ.
The “box of people” framing predicts model behavior correctly: strong on common tasks with lots of precedent data, weak on rare ones — because it’s a combiner of what humans already did, not a general world-synthesizer.
“Data dignity” = paying people on a market basis for the data their work contributes to models, per actual use.
A safety mechanism falls out of this: trace each output back to the training clusters most responsible for it (a bomb query would surface a “bombs” cluster). This is a separate, concurrent channel — “multifactor AI” — analogous to two-factor authentication.
Lanier opposes UBI as politically unstable: any single distribution mechanism becomes a central point that gets seized; “you might start with Bolsheviks but end up with Stalinists.” Markets keep power distributed.
He distrusts the binary kill-us-or-pet-us futures because brittle, no-wiggle-room scenarios rarely happen.
His preferred future: ever-expanding new classes of creative (not obsolete) people, paid for novel contributions — valuable even if only partially realized.
He wants radical change with continuity of memory; a singularity, by erasing the line back to now, is “dementia.”
He flips optimism/pessimism: building toward a “god” got rebranded as optimism, while betting on human creativity got rebranded as pessimism.

Claude’s Take

This is Lanier doing what Lanier does best: one stubborn reframe, pushed relentlessly until the room looks different. The core claim — that “AI is just a recombination of human work, and calling it a separate mind is a choice with consequences” — is genuinely clarifying, and his point that the framing changes how well you use the tool is the most practically useful thing in here. The training-data-clusters safety idea (essentially influence-functions / attribution research dressed up as “multifactor AI”) is real, ongoing work, not hand-waving.

Where I’d push back: he leans on “both interpretations are technically identical” to win arguments, but then smuggles in strong claims — that AI “does nothing,” that emergent capabilities don’t really exist — that go beyond that careful symmetry. Whether models only ever interpolate within their training data or can genuinely extrapolate is an open empirical question he treats as settled. And his repeated armchair psychiatry of AI leaders (“they went crazy,” “their personalities degraded”) is rhetorically fun but evidence-free. The economics of data dignity also stay frustratingly abstract — he waves off the “it’d just be pennies” objection as a mere “cultural attitude” without showing the mechanism that makes the numbers work.

Score 8: a sharp, original, well-argued perspective from someone with real standing, and the rare AI conversation that’s neither doom nor hype. Docked for the moments where confident assertion stands in for argument, and for leaving the central economic proposal more vision than blueprint.

Father of VR: The best AI future nobody is talking about | Jaron Lanier