People in Disguise

crs48 with Claude · History

Before he was anything else, Jaron Lanier was a musician — the kind who collects instruments the way other people collect arguments. He owns hundreds of them, many rare, many ancient, and he has a habit of opening talks not with slides but with something like a khaen or an oud, played properly, while the audience recalibrates its idea of what a Microsoft scientist looks like. Keep the instruments in view. They are not a personality quirk decorating a technology critic; they are where the critique comes from.

In the early 1980s, musicians and engineers settled on a protocol for connecting electronic instruments: MIDI. It was built around the gesture its designers had in front of them — a key on a keyboard, going down and coming up. Note on, note off. It shipped, it spread, and because everything then had to interoperate with everything else, it stuck. But a violinist does not play note-on, note-off. A voice slides between the keys; that is where most of the music lives. A protocol built from one instrument’s idea of music became the box every instrument had to fit through — and forty years later, it still is. Lanier, who could hear exactly what the format dropped, made MIDI his life’s cautionary tale: digital systems freeze their designers’ first guesses into infrastructure, and then the infrastructure quietly teaches everyone that the guess was the truth. He gave the pattern a name — lock-in — and a warning label that doubles as the thesis of everything he has written since: “the most important thing about a technology is how it changes people.”

This essay is about him — the forty-year argument he has been making from inside the machine, and what it means for what people like us should build. We wrote an essay two weeks ago about the interfaces the web froze too early; this is the portrait of the man who heard it happen first, in music, and spent four decades warning the rest of us where it leads.

The forty-year argument

Most technology criticism dates badly, because most of it is about a product. Lanier’s does not, because his subject was never the product — it is what the product does to the person using it. Read his work chronologically and the same claim arrives every decade wearing new clothes.

Start earlier than most retrospectives do. In 1995 — three decades before the current agent boom — he published a short essay called “Agents of Alienation”, aimed at the era’s clippy-ish software agents. The argument was not that agents wouldn’t work. It was worse: that they work on you. An agent, he wrote, is “a way of using a program, in which you have ceded your autonomy” — and since the agent is dumber than you, you meet it in the middle. You simplify your requests, flatten your taste, learn its categories, and in his four bleakest words, “you make yourself dumb” so the agent can look smart. Hold that thought for later; 1995 is not done with us.

By 2000 he had named the ideology he was against — cybernetic totalism, the belief that people are, in the end, nothing but information patterns — and in 2006, in the essay “Digital Maoism”, he attacked its collectivist cousin: the fashionable idea that an anonymous hive produces wisdom, and that authorship is a quaint inefficiency. In 2010 the argument became his manifesto, You Are Not a Gadget: lock-in, MIDI, and the observation that the templated profile — the multiple-choice identity form — is a MIDI for personhood, shrinking what a person is until it fits the database. “You have to be somebody before you can share yourself,” he wrote, and much of the decade since has been an experiment in what happens when the order is reversed.

In 2013, Who Owns the Future? gave the economy of all this a name that stuck: siren servers — the elite computers that lure everyone into interacting for free, harvest the data, keep the proceeds, and push the risk back out to everyone else. His signature comparison: Kodak at its peak employed more than a hundred and forty thousand people; Instagram employed thirteen when it sold for a billion dollars. Economists have real objections to that framing — we will get to them — but the book’s deepest line is not the statistic. It is the ontological one: “digital information is really just people in disguise.” There is no such thing as data that came from nowhere. Every training set, every graph, every recommendation is congealed human effort with the names filed off.

By 2018 he had stopped being polite about the mechanism. Ten Arguments for Deleting Your Social Media Accounts Right Now introduced BUMMER — his acronym for the machine: “Behaviors of Users Modified, and Made into an Empire for Rent.” Advertising, he argued, had quietly become “continuous behavior modification on a titanic scale”: adaptive algorithms running intermittent reinforcement on billions of people, tuned by feedback, rented to whoever pays. And because negative emotion is cheaper to trigger than positive, the tuning has a direction: “social media is biased, not to the Left or the Right, but downward.” Note the precision of the target — it was never the internet, never the phone, never even the algorithm. It is the business model. Services you simply pay for are, in his telling, explicitly exempt.

Then the models arrived, and the man who had spent thirty years being called a pessimist wrote the most level-headed essay of the boom. In “There Is No A.I.” (The New Yorker, 2023) he refuses both available hysterias. The pragmatic position, he argues, is to “think of A.I. as a tool, not a creature” — and to see the tool clearly: a large model is “an innovative form of social collaboration”, a mash-up of millions of people’s writing, drawing, and judgement. “Big-model A.I. is made of people” — and the way to govern it is not mystical alignment but the oldest tool in the librarian’s drawer: provenance. Keep track of where things came from. “When you lose context, you lose control.” He has kept at it since: a 2025 New Yorker essay on AI companions as the next behaviour-modification empire, and a 2026 university lecture arguing that provenance is not just fairness accounting but epistemic infrastructure — the thing that lets a society keep hold of what is real.

He makes this argument, note, from the eleventh floor of the machine: he has been at Microsoft for nearly two decades, with the self-mocking title “Prime Unifying Scientist” in the CTO’s office, and a standing disclaimer that he speaks only for himself. You can read that as compromise. We read it the way he seems to: if the argument is that technology should be built differently, the place to make it is where technology gets built. It is also, we should say plainly, a tension — and one we share, further down this page.

The spec hidden in the critique

Here is the thing about Lanier that separates him from most of the genre: underneath the criticism there is a buildable specification. He rarely writes “thou shalt not” without an engineering alternative attached, and if you collect the alternatives across the corpus they form a coherent design brief — one that has been sitting in plain sight, largely unimplemented, for decades.

Keep provenance. His favourite exhibit is Ted Nelson’s Project Xanadu, the 1960s hypertext design the web didn’t choose, in which every link ran two ways: content always knew where it had been used, and use always pointed back to origin. Attribution — and, Nelson hoped, payment — falls out of the architecture for free. The web’s one-way link made publishing gloriously cheap and made context evaporate; for Lanier this is the original sin behind everything from misinformation to the AI training-data fight. Don’t build siren servers — architectures whose economics require one privileged computer to see everything. Don’t do behaviour modification — no adaptive feed whose objective function is your future behaviour. Sell things, not people — his exemption for direct-payment services is an instruction, not a loophole. And keep agents on a leash with a name tag — a tool you direct, never an anonymous creature you defer to.

We did not start from this list. xNet’s design brief came out of its own history — a decade of watching local-first research, and a charter arrived at by auditing our own architecture against the general critique of extractive software. But when we finally read the corpus end to end, the overlap was close enough to be embarrassing. Two independent walks ending at the same place is not influence. It is something better: evidence that the place is real.

Receipts, not vibes

So let us mark the homework honestly, spec line by spec line — because an essay that admires a critic and then gestures vaguely at “alignment with his values” is exactly the kind of writing he has spent forty years puncturing.

Provenance. This is the deep one. In xNet, every edit you make — every keystroke batch, every task ticked, every row changed — becomes a small, permanent record with a precise shape:

// packages/sync/src/change.ts — every edit you make becomes one of
// these. Note what a change cannot be: anonymous, or unsigned.
/** Content-addressed hash of this change */
hash: ContentId
/** Hash of the previous change in the chain (null for first) */
parentHash: ContentId | null
/** DID of the author */
authorDID: DID
/** Ed25519 signature of the hash */
signature: Uint8Array

Read those fields as a Xanadu partisan would. Every atom of data in the system knows who made it (authorDID — an identity you mint yourself, not an account you rent), proves it (signature), and knows where it sits in history (parentHash, a hash chain). When two people edit the same thing, even the tie-break is authorship-aware. There is no anonymous blob at the bottom of the architecture into which individual contribution dissolves — which is, almost word for word, what the “Digital Maoism” essay asked collaborative software to preserve, and what we walked through line-by-line in an earlier essay. Context, kept, at the atom.

No siren servers. A siren server’s defining property is asymmetry: it must see everything, and the value must pool where the seeing happens. xNet’s hubs are built to be the opposite — relays and meeting points, not observers. The master copy of your workspace lives on your own device, in an ordinary SQLite file; a hub you choose, rent, or run on a spare machine forwards signed changes it cannot forge between your devices and your collaborators. There is no behavioural surplus pooling anywhere, because none is collected — and that is not a policy promise, it is a property you can audit: a continuous check in the repository fails the build if anyone wires in a third-party analytics or advertising SDK.

No behaviour modification. The BUMMER machine needs an adaptive feed with an engagement objective. xNet does not have one. Feeds are chronological; notifications follow rules you can read, not a model of your weaknesses; nothing in the interface is optimising you. This is the charter’s Calm axis, and it is enforced the same boring way as the rest — in code review and in checks, where dark patterns are treated as defects.

Agents with name tags. And here 1995 comes back around. The current wave of software agents is about to re-run the experiment Lanier described — people ceding autonomy to processes that act on their behalf — at a scale he could only sketch. xNet’s answer is a mechanism we call the agent passport: an agent operating in a workspace gets its own cryptographic identity, distinct from yours, carrying an explicitly scoped grant of authority. Every change an agent writes is signed by the agent’s key — so the provenance machinery above applies to the machines too. An agent cannot impersonate you, cannot exceed its scope, and cannot act without leaving a signed trail; the riskiest capabilities cannot be waved through from a chat window at all. A tool with a name, on a leash you hold — the 1995 essay, resolved as an access-control design.

The fork in data dignity

Now the part where we disagree with him — because the portrait is worthless without it.

In 2018, with the economist E. Glen Weyl, Lanier turned the siren-server critique into a programme with a name: data dignity. On StarTalk he compressed its premise into ten words: “data only comes from people. Data doesn’t come from angels.” The programme has two halves. The first is the provenance half: contribution should remain attributed — traceable, inalienable, never laundered into an anonymous aggregate. The second is the market half: that attributed data should generate royalties, flowing through collective-bargaining organisations they call MIDs — mediators of individual data, fiduciary unions for your digital exhaust — so that the people a model is made of become a paid creative class rather than an unpaid substrate.

We built the first half. We refused the second. And the refusal is not a shortcut — it is a position, and it can be argued largely in his own terms. The strongest critiques of the market half have been public since the week the book shipped: for most people the royalties would be pennies, dressed in the costume of empowerment; the accounting apparatus would be a surveillance system of its own; and — the sharpest one — putting a price on harvested data does not end the harvesting, it legitimises it, converting a violation into a transaction. A market in your data still requires the mine. It just promises you a small cut of the ore.

xNet’s wager is that dignity comes before the invoice: collect nothing, attribute everything, and let the money flow where Lanier himself pointed in his exemption — direct payment for things people choose to buy. In our corner of the world that means you pay for hosting, or run your own; AI is metered by what you actually use, not subsidised by what can be extracted from you; and people who build things on the platform price their own work. Attribution without a data market; commerce without a mine. He might not accept the trade — the middle-class economics of Who Owns the Future? wanted royalties at society scale, and our answer only covers the software we ship. But between a data economy priced by siren servers and one with no prices at all, there is a third door, and we think it is the one his own exemption points at.

People are the answer

There is a sentence near the end of the 2023 essay that reads like a summary of the man: “people are the answer to the problems of bits.” It is easy to hear that as sentiment. It is not. It is an engineering claim — the same one he has been making since the MIDI wars: that every system which forgets it is made of people degrades the people it is made of, and every system that keeps the people visible — named, attributed, paid or at least asked — stays correctable. The claim scales down as well as up. It is true of a feed, and of a training set, and of a change log four fields wide.

The instruments were the point all along. A khaen does not disguise the player; it amplifies them — you can hear the breath in it. Software can be built like that. The note the protocol couldn’t hear was never lost because it was unplayable; it was lost because a format shipped without it, and everyone adjusted. Forty years on, the adjustment is optional. The formats are still being written — some of them by us, in the open, with the author’s name signed on every note.

If you want to play one: use the app — free, offline, private. Read the commitments and the receipts behind them, or build on the protocol yourself. And read Lanier — start with the New Yorker essay, then the books. He has been holding this note a long time. It deserves the chord.

Sources

• The conversation this essay links: StarTalk Plus — Neil deGrasse Tyson and Jaron Lanier on the AI Illusion (YouTube). The “data doesn’t come from angels” line is from this conversation, via the episode’s published transcript.
• The corpus: You Are Not a Gadget (2010); Who Owns the Future? (2013); Ten Arguments for Deleting Your Social Media Accounts Right Now (2018); “There Is No A.I.” — The New Yorker, 2023; “Agents of Alienation” (1995); “One Half a Manifesto” — Edge, 2000; “Digital Maoism” — Edge, 2006.
• Data dignity: Lanier & Weyl — “A Blueprint for a Better Digital Society”, Harvard Business Review, 2018 and Arrieta-Ibarra et al. — “Should We Treat Data as Labor?”, AEA Papers & Proceedings, 2018. Provenance as epistemic infrastructure: Brown University — the 2026 Cooper Lecture.
• The critics engaged above: Slate — Will Oremus on micropayments, Forbes — Tim Worstall on the Kodak/Instagram comparison, OneZero — Joshua Adams on why paying for data could entrench surveillance, and Institute of Network Cultures — responses to “Digital Maoism”.
• The architecture and the receipts: xNet — Why and the Humane Charter. Companion essays: Clutch Power, The Loom You Can Read, and The Vault and the View.

This is an independent essay about Jaron Lanier’s published work. He has no affiliation with xNet and has not reviewed or endorsed this essay or this project; quotations are brief, attributed, and used as commentary and criticism. Where a quotation comes from an episode transcript rather than a text he wrote, the sources above say so. All artwork here is original, and this page loads nothing third-party.