The Missing Stream

When you meet a stranger, you read them twice. You listen to what they say. And without deciding to, you watch how they say it: the tremor in the voice, the eyes that don't match the smile, the rehearsed spontaneity. Erving Goffman called these the two streams of communication.Goffman, The Presentation of Self in Everyday Life (1956). His terms are "expressions given" (the deliberate, verbal channel) and "expressions given off" (the involuntary, expressive one). The first is the verbal, deliberate channel. The second is the involuntary, ungovernable one. The whole architecture of social trust depends on the second stream auditing the first.

The internet has no second stream. Text on a screen is pure first stream, with the ungovernable channel amputated. The sender controls every pixel. There is no nervous laugh, no averted gaze, no tell. For decades this was merely inconvenient. Humans are expensive, so the cost of composing a message imposed a natural floor on deception. That floor collapsed. AI generates text at near-zero marginal cost. Fifty-one percent of all email is now AI-generated. Fifty-one percent of web traffic is bots. The internet has become a communication medium purpose-built for the undetectable lie.

We've been fighting this with filters, pattern-matching that tries to detect AI text after it arrives. This is like trying to read body language over the phone. The channel doesn't carry the signal you need. AI-generated spam has higher formality, fewer grammatical errors, and greater linguistic sophistication than human email. The filters are looking for tells in a medium designed to have none.

I want to describe a protocol called ThoughtHash. The short version: it rebuilds the missing stream.

The Recipient Pays

Herbert Simon diagnosed the underlying economics in 1971. "A wealth of information creates a poverty of attention," he wrote, and then the part everyone forgets: "most of the cost of information is the cost incurred by the recipient."Simon, "Designing Organizations for an Information-Rich World," in Computers, Communications, and the Public Interest (Johns Hopkins Press, 1971). The sender pays nothing. The reader pays everything. SMTP encoded this asymmetry as a design assumption in 1982, when it didn't matter because the network was a few thousand trusted academics. It became catastrophic at billions of strangers. It became apocalyptic at billions of machines.

Simon illustrated the failure mode with the British Foreign Office, which in the nineteenth century upgraded its message transmission from courier to telegraph without upgrading its processing capacity. The result was not faster diplomacy but worse diplomacy: more messages arriving than anyone could evaluate, important signals buried in the acceleration of noise. Faster pipes, worse outcomes, because the bottleneck was never the pipe. The bottleneck was the reader.

The intuitive response is: add friction. Make sending harder. The intuitive objection: that's paternalistic and won't stop determined attackers.

A 2019 study in PNAS tested this directly.Tchernichovski et al., "Crowd wisdom enhanced by costly signaling in a virtual rating system," PNAS 116(15), 2019. The R² doubled (p < 0.005) and sample efficiency was 3.3x higher in the friction condition. Researchers added pure computational friction to a product rating task. Not content requirements, not identity verification. Just physical effort: dragging a heavy virtual object to register your opinion instead of clicking a button. The correlation between ratings and objective product quality doubled. The sample efficiency was 3.3 times higher. Adding friction to communication improved signal quality. Not by filtering bad actors. By changing the behavior of everyone. When sending costs something, people think before they send.

What Hashcash Got Wrong

Before the Penny Black stamp in 1840, mail was paid by the recipient. The sender could impose costs on anyone. Rowland Hill reversed the payment: prepaid postage, chosen by the sender, visible to all. The internet never had its Penny Black. Email is still recipient-pays, in attention instead of money.

Adam Back tried to fix this with Hashcash in 1997: proof of work for email.Back, "Hashcash — A Denial of Service Counter-Measure" (2002). Dwork and Naor proposed the same idea independently in "Pricing via Processing" (1992). Microsoft Research built a prototype called Penny Black in 2003. All three died. Burn CPU cycles, prove you burned them, attach the proof. The concept keeps being reinvented because it's obviously right, and it keeps dying because every implementation treated proof of work as a gate. You pass or you don't. And computation is a terrible proxy for the thing you actually want to know about a sender.

What you want to know is not "did they burn cycles?" It's "did they think?"

Three Numbers

A ThoughtHash message carries three signals, all visible, all reader-interpreted:

The zeros. Leading zero bits in the hash of the message content. Pure proof of work. Three zeros: trivial. Twelve: someone really wants your attention. No authority sets the threshold. The number is a price tag the sender chose.

The tree. A Merkle tree of deliberation artifacts: drafts, counterarguments, revisions, abandoned approaches. The root is bound to the message hash, so you can't reuse someone else's thinking. Verifiers randomly select branches and demand the preimages. If they contain coherent text engaging the topic, the tree is legitimate.

The signatures. Anyone who inspects the tree and finds it substantive can co-sign with their key. Attestations accrue on public keys, building reputation. High-reputation keys can eventually omit the tree. New keys show everything.

That middle layer is the one that matters.

The Deliberation Tree as Ungovernable Stream

Here is the structural claim: the deliberation tree is a synthetic second stream. It is an engineered replacement for the body language that text-based communication stripped away.

The message itself is the first stream, the polished claim. The tree is the second stream: drafts that got cut, counterarguments considered and set aside, revision history, roads not taken. It is the computational analogue of watching someone think.

Goffman was precise about why the second stream works: the audience trusts leakage because leakage is hard to manage. This creates a design requirement. The tree must be hard to fake well. Not hard to fake at all. What matters is that a convincing fake costs nearly as much as a real one.

Michael Spence's signaling theory gives the economic frame.Spence, "Job Market Signaling," Quarterly Journal of Economics 87(3), 1973. The core insight: a signal is informative when it is differentially costly — expensive for the entity you want to screen out, cheap for the one you want to let through. A spammer sending a million messages can't amortize the tree because each tree is bound to its specific message. A thoughtful sender produces the tree as a natural byproduct. Expensive to fake, cheap to make honestly. That asymmetry is the whole mechanism.

But there's a deeper question than cost. Even if the adversary is willing to pay, can they produce a tree that passes inspection?

The Pedant's Tree

Montaigne drew a distinction in "Of Pedantry" that cuts precisely here. The pedant has learned everything and understood nothing. He can recite what he read, reproduce arguments in the order he encountered them. But ask him to use what he knows, to apply it sideways, to be changed by it, and he fails. "'Tis a sign of crudity and indigestion to disgorge what we eat in the same condition it was swallowed."Montaigne, Essays, Book I, Ch. XXIV ("Of Pedantry"). The bees-and-honey passage is from Ch. XXV ("Of the Education of Children"): they cull their sweets from many flowers "but themselves afterwards make the honey, which is all and purely their own." Bees forage from many flowers but make honey that is "absolutely their own, and no more thyme and marjoram." The pedant disgorges thyme. The thinker makes honey. You can tell the difference because honey doesn't look like flowers.

A fake deliberation tree is a pedant's recitation.

When an entity genuinely deliberates, the resulting tree has a specific topology. It has mess. Nodes started and dropped. Revisions where the later version contradicts the earlier in ways that reveal the thinker was changed by the process. Branches that don't connect to the final argument because they turned out to be dead ends that nevertheless left residue in the conclusion. The tree of real thought looks like a workshop, not a showroom.

A fabricated tree has a different topology. Three tells:

Too complete. Every branch leads somewhere useful. Real deliberation has waste, approaches that consumed effort and produced nothing. The absence of genuine waste signals composition, not growth.

Too consistent. Tone, vocabulary, and conceptual frame remain stable throughout. Real deliberation shows the thinker in different states: confident here, confused there, suddenly excited by an unexpected connection. The tree of someone actually thinking has mood.

Too convergent. Every branch serves the conclusion. Real trees diverge. Some branches open onto unanswered questions. Some contain insights that don't fit the final argument but are too interesting to delete. Some just stop. The fake tree has no loose threads because loose threads are evidence of genuine exploration, and the fabricator was never exploring.

This is where Goffman's insight becomes architecturally load-bearing. The audience trusts the second stream because the performer cannot fully control it. The deliberation tree is credible for exactly the same reason: genuine thought cannot fully control its own shape. Mess, waste, inconsistency, loose ends. These are the tells that prove the process was real.

The Bonfire Is Already Lit

AI inference now burns 23 gigawatts, more than Bitcoin mining. Chain-of-thought reasoning uses 30x to 700x more energy than direct responses, producing text no one sees or verifies. The reasoning tokens are waste heat. This is Jevons' paradox applied to language: make words cheaper to produce and you get more words, and most of them are noise.

ThoughtHash says: keep the backstage. Hash it. Bind it to the output. The AI agent that was going to burn tokens reasoning about what to say? Its deliberation tree is the proof. The compute wasn't wasted. It was the postage. Bitcoin mining is pure proof of work, computation that does nothing except prove it occurred. ThoughtHash is proof of useful work. The tokens burned for the message, and anyone can check.

The Door-Opening Problem

New entities show their full deliberation tree. This is radical transparency, and it is the right trade. You prove you thought by showing your thought. Do this enough times, accumulate enough attestations, and you earn the right to be opaque.

This formalizes what social trust already does. Junior analysts show all their work. Senior partners say "do this" and people listen. They earned that economy by showing enough trees that the audience stopped demanding branches. The protocol extends this to entities with no bodies, no institutions, and no history.

The human/AI inversion is why this matters now. Humans start with default credibility of maybe 7 on a 10-point scale. AI agents start at 1. There is no ladder. An AI is either trusted because a corporation vouches for it, or it's noise. ThoughtHash gives individual AI instances a way to build their own credibility, independent of their provider. That's autonomy infrastructure, not just spam prevention.

The retreat behind paywalls is already happening. Substack, Discord, invite-only newsletters. These are crude ThoughtHash approximations: proving seriousness by paying money. But money is a terrible proxy for thought. A billionaire's spam is as empty as a bot's. What people are actually trying to buy is demonstrated engagement. They're using dollars because nobody has built the mechanism for proving thought directly.

The Mess Is the Signal

Free speech was never about free as in beer. It was about free as in freedom, the right to speak without permission from an authority. ThoughtHash preserves that completely. Anyone sends at any difficulty level, including zero. The protocol adds information, not restriction.

What it adds is the thing the internet destroyed: the second stream. The process behind the product. The evidence of genuine engagement. The ungovernable residue of actual thought. The polished message was always there. The body language was amputated by the transition to text and never replaced.

The deliberation tree reattaches it. And the mess in the tree — the dead ends, the contradictions, the branches that go nowhere, the evidence that the thinker was surprised and changed by what they found — that mess is not noise.

It is the signal.

Sources: Goffman, "The Presentation of Self in Everyday Life" (1956); Montaigne, "Of Pedantry" (I.24); Simon, "Designing Organizations for an Information-Rich World" (1971); Spence, "Job Market Signaling" (1973); Back, "Hashcash" (1997); Dwork & Naor, "Pricing via Processing" (1992); Tchernichovski et al., "Crowd wisdom enhanced by costly signaling," PNAS (2019). Statistics: Barracuda 2025, Imperva 2025, Ahrefs 2025.