A benchmark score is not proof. It is a press release dressed in the costume of evidence.
I want to start there, because of a small item that crossed my desk this week. Crypto Briefing, a publication I read more for its coverage of token markets than for its reporting on artificial intelligence, published a short piece claiming that employees inside Google harbor doubts about the capabilities of Gemini, the company's flagship AI model. No benchmarks. No named sources. No internal documents. Just a claim, attributed to people we cannot see, about a model we cannot inspect, published by a newsroom whose expertise lies elsewhere. Within hours it had been aggregated, quote-tweeted, and folded into the ambient anxiety that now surrounds every large language model release. I read it three times — not because I believed it, but because of what it exposed about the way we have learned to trust machines we cannot audit. The story was never really about Gemini. It was about the verification deficit at the heart of the AI industry — the same deficit that decentralized systems were invented to close.
To understand why that deficit matters, you have to understand what a modern AI model actually is as an artifact of trust. When a lab releases a frontier model, it hands you three things: a weight file you usually cannot see, a set of benchmark numbers you cannot reproduce, and a narrative you are asked to accept. The weights are proprietary. The training data is undisclosed. The evaluation harness is internal. What remains is a claim — "state of the art," "frontier," "best-in-class" — and an invitation to believe it. This is not entirely new. Traditional software vendors have always guarded their source. But conventional software fails in legible ways. A database either returns the record or it does not. A payment API either settles the transaction or throws an error. An AI model fails in ways that are statistical, contextual, and — crucially — invisible from the outside. You cannot diff a hallucination. You cannot unit-test a worldview. And so the entire industry has quietly migrated from verification to faith.
The Gemini item is interesting precisely because it inverts that faith. Anonymous employees, we are told, doubt the model. But stop and notice the structure of the claim: it is exactly as unverifiable as the marketing it undermines. We are asked to trust the skeptics on the same terms we are asked to trust the labs. There is no proof on either side — only assertion, counter-assertion, and the market's appetite for a story.
Crypto learned this lesson the hard way, and it paid for it in real money. I have spent the better part of a decade watching protocols promise things they could not prove, and watching those promises dissolve the moment liquidity and luck ran out. I retreated to a cabin in the Scottish Highlands for six weeks after the collapse of Terra and Celsius, drafting an essay I called "The Burden of Belief," because the betrayal I felt was not really about price. It was about having believed claims that were never verifiable in the first place. That is the emotional cost of a system built on trust rather than proof, and the AI industry is now walking the same road, at ten times the speed, with none of the scar tissue.
Crypto's founding insight was not that money should be digital. It was that trust should be verifiable. Before Bitcoin, we trusted banks because we had no choice; the ledger was theirs, the rules were theirs, and the audit was theirs. Satoshi's contribution was not a currency but a verification mechanism — a way to prove that a transaction happened without asking a gatekeeper to confirm it. Trust is not given; it is verified. That single inversion is the most important idea to come out of the last twenty years of computer science, and the AI industry has, so far, declined to adopt it. Instead it has rebuilt the pre-Bitcoin world: a handful of enormous institutions, each issuing claims about systems the public cannot inspect, each asking for belief on the strength of brand rather than evidence.
If the Crypto Briefing claim has any substance, the doubt is probably not about raw capability. It is almost certainly about one of three things: hallucination rate under adversarial prompting, code-generation correctness, or multimodal consistency. These are the failure modes that benchmarks flatter and users eventually discover. A model can top MMLU and still fabricate a citation; it can pass HumanEval and still introduce a subtle security bug into a production codebase; it can ace a vision benchmark and still misread a chart. Internal teams see the long tail of failures that the leaderboard hides. That is not a scandal — that is normal engineering. The scandal is that none of it is externally verifiable. When a team says it doubts a model, and the only evidence is a whisper, we are not watching a technical disclosure. We are watching a rumor perform the function that an audit should perform.
And this is where I want to slow down, because the benchmark problem is deeper than most people admit. LMSYS Chatbot Arena, MMLU, HumanEval, GPQA — these are not neutral instruments. They are social constructs, and like all social constructs they can be gamed, through data contamination, prompt tuning, or selective reporting. When a lab trains on data that overlaps a test set, the score inflates and no one outside can tell. When a lab reports only the benchmarks it wins and omits the ones it loses, the market sees a fraction of the truth. We built oracles and price feeds precisely to solve this class of problem in finance — to replace self-reported numbers with independently verifiable ones — and yet we accept self-reported intelligence scores as though they were carved in stone. The irony is not subtle. The crypto industry spent a decade learning that self-reported reserves are worthless, and the AI industry is now issuing self-reported intelligence as if the lesson never happened.
I have skin in this game. In 2026 I led a cross-functional team building what we called a Provenance Layer — a system that used cryptographic attestation to mark content as human-created, and to verify that marking at scale. We partnered with ten media houses to test a verification cost of roughly one cent per item. The hardest part was never the cryptography. It was the economics: verification only works when it is cheap, continuous, and boring. The moment verification becomes expensive or ceremonial, people route around it. I think about that constantly when I read AI press coverage, because the industry has built a verification system that is expensive, discontinuous, and theatrical — independent evaluations cost money, they are released on launch days, and they are designed for headlines. That is not verification. That is performance.
There is a real technical path here, and it runs through the same primitives we use in decentralized systems. Zero-knowledge proofs of inference would let a model prove that a specific output was produced by a specific computation over specific weights, without revealing the weights. Trusted execution environments offer a weaker but cheaper version of the same guarantee. Cryptographic attestation of training-data provenance — a signed manifest of what went into the model — would let outsiders audit the inputs without exposing the proprietary recipe. None of these are science fiction. They are engineering problems with known shapes, and the academic literature is already dense with prototypes. What is missing is not capability. It is incentive.
And incentives, in the enterprise, are misaligned in a familiar way. I spent part of 2024 helping a major UK pension fund draft an investment thesis for Bitcoin, and I watched how institutions actually evaluate trust. They did not want an open, permissionless chain. They wanted auditable controls, clear SLAs, and a counterparty they could call. This is the same reason the RWA narrative has stalled for three years: traditional institutions do not need your public chain. They have their own trust rails, and those rails are private, contractual, and boring — and boring is exactly what they want. The same will be true of AI verification. Enterprises will not wait for an open protocol. They will demand attestation from their vendors, and they will get it — but only if we make verification a public good before the private version calcifies into a standard no one can inspect. Freedom arrives when the gatekeepers go dark — but only if someone has already lit the alternative.
There is a second parallel that worries me more. Over the past few years we have watched dozens of Layer 2 networks compete for the same small pool of users and the same thin liquidity, slicing an already-scarce resource into ever-finer fragments and calling it scaling. The AI industry is doing something structurally similar. Every week brings a new "frontier" model, a new leaderboard, a new claim of superiority — and yet the underlying base of verifiable evidence barely grows. We are not expanding the frontier of proof. We are fragmenting the appearance of it. A hundred models that cannot be independently compared are not a hundred steps forward. They are one step forward, repeated until it looks like progress. And in a sideways market, where capital is patient and attention is scarce, that distinction matters more than it does in a bull run. Stillness reveals the signal beneath the noise — but only if there is a signal to find.
I have seen this movie before, in a market I know too well. The "blue chip" NFT label was a trap — a story that held only as long as liquidity held. When the bid dried up, the floor collapsed, and what remained was a JPEG and a promise. The "frontier model" label is the same kind of artifact. It is a social consensus, not a technical fact, and social consensus is a terrible foundation for infrastructure. The protocol remembers what the market forgets. Benchmarks are forgotten the moment the next release lands. Verifiable computation, if we build it, is not.
For those of us who work in decentralized systems, the Gemini story is not a spectator sport. It is a preview of the demand curve we should be serving. The world is about to be flooded with synthetic content, synthetic agents, and synthetic claims, and the only durable answer to synthetic trust is cryptographic trust. This is the thesis behind human-centric cryptography — the idea that the moral imperative of the next decade is preserving verifiable truth. When I started building provenance tooling, people asked why a blockchain person cared about AI. The answer is that they are the same problem wearing different clothes. Both ask the same question: how do you trust a system you cannot see?
There is one more layer, and it is the one the skeptics rarely name. If employees inside a major lab are genuinely uneasy, the unease may not be about accuracy at all. It may be about the alignment tax — the performance penalty a lab pays for making a model safer, and the pressure to reduce that tax in order to win a benchmark war. When the scoreboard rewards raw capability and hides safety cost, the rational move is to under-invest in safety and over-invest in the leaderboard. This is a structural incentive, not a moral failing. And it is invisible to anyone outside the building, because safety evaluations are exactly the numbers nobody publishes. The same dynamic runs through the infrastructure stack: if internal confidence is shaky, capital expenditure on training clusters and custom silicon can quietly slow, and the delay shows up not as a press release but as a gap in a quarterly filing. We build in silence so the network can speak — but the network can only speak if someone, somewhere, is allowed to verify what was built.
Now let me offer the angle I have not seen anyone take, because it cuts against both sides of this story.
The instinctive reaction to the Gemini report is to pick a team. Either the employees are right and Google is in trouble, or the employees are wrong and the story is noise. I think both readings miss the point. The important fact is not whether Gemini is good. It is that after a decade of AI hype, we still cannot answer that question from the outside, and the market has learned to substitute narrative for evidence because it has no choice. The skeptics and the marketers are playing the same game — assertion without proof — and we are the audience, trained to applaud whichever assertion flatters our priors.
There is a harder, more uncomfortable contrarian claim here: transparency is not the answer, and demanding it is a trap. If Google published every benchmark and every internal doubt, we would still be trusting Google's interpretation of its own numbers. The answer is not more disclosure. The answer is verifiable computation — proof that does not require anyone's good faith. That is a much harder ask, and it is why the industry keeps settling for stories instead of proofs. Patience is the validator of true intent, and we have not been patient. We have been credulous. We have treated the absence of evidence as evidence, on both sides, and we have mistaken the volume of the claim for the weight of it.
So what do we watch? Not the next rumor, and not the next launch-day leaderboard. Watch whether any lab moves from self-reported claims to cryptographic attestation — whether the first serious proof-of-inference ships, and whether anyone builds the boring, cheap, continuous verification layer that makes it matter. Watch whether enterprises demand auditable inference SLAs the way they demanded auditable reserves after 2008. The Gemini skeptics, whoever they are, have done us an accidental favor: they have shown that even the most powerful company in AI cannot escape the verification deficit. Code is the only permission we truly need. Liberation is not a promise; it is a state. And it begins the moment we stop asking to be believed, and start asking to be proven.
