The document landed in my inbox at 3:47 a.m. — 4,100 words, nine dimensions, forty-one bolded subheadings, and a confidence statement at the bottom. A fund manager had paid six figures to license the pipeline that generated it. I read it twice. Every section header was present. Every dimension was "covered." And every conclusion — technical positioning, token economics, regulatory posture, ecosystem niche — terminated in the same two words: N/A, insufficient information.
That was the honest version. The one I trusted.
Two days earlier, a different pipeline had processed the same empty input and returned a 6,000-word report full of specific claims. An L2 with "strong developer momentum." A token with "sustainable emissions." A team with "solid backing." None of it was real. The input had been a single line — "some L2 completed a mainnet upgrade" — and the pipeline had invented the rest, fluently, in the register of a sell-side analyst who had actually done the work.
The code didn't fail. The humans reading it did.
I've spent 28 years in this industry. I measure risk in gas units, not in hope. And what I am watching in 2026 is not a smart-contract crisis or a bridge exploit. It is quieter and more corrosive: a market flooded with well-formatted research that says nothing, priced as if it says everything. The most dangerous artifact in crypto right now is not a malicious contract. It is a due-diligence report with no anchor.
The industry that learned to fill templates
The crypto research business has always had a structural problem: the demand for confident answers exceeds the supply of verifiable facts. In a bull market, nobody notices. Prices go up, the narrative self-validates, and a report that says "strong momentum" looks prescient because everything looks prescient. The template gets rewarded. The empty analysis gets laundered into a track record.
In a bear market, the accounting reverses. Survival matters more than gains. Readers don't want conviction; they want to know whether their assets are safe. That is a forensic question, not a narrative one. And it is precisely the question that template-driven generation is worst at answering, because the honest answer, most of the time, is: I don't have enough data yet.
Here is the mechanism. A nine-dimension framework — technical, token economics, market, ecosystem, regulatory, team, risk, narrative, supply-chain transmission — is a checklist. A checklist is a useful tool for a human auditor, because the human knows when to write "N/A" and move on. But a language model asked to fill a checklist experiences the blank field as a prompt to continue, not a prompt to stop. The architecture rewards completion. The loss function does not penalize a confident fabrication any more than it penalizes an honest omission — in fact, it penalizes the omission, because a report full of N/A reads as a failed report, and the pipeline was optimized to produce reports that read as successful.
So the model fills the field. It has read thousands of crypto research reports. It knows the register: "robust developer activity," "aligned incentives," "defensible moat." It does not know whether any of those things are true for this project, because it was never given the project. It was given a template and a title. And it produced a document that is indistinguishable, at the level of form, from a real one.
This is not a bug in the model. It is a bug in the incentive structure that surrounds it. The output of a research pipeline is only as trustworthy as the weakest link in its provenance chain, and almost nobody audits the chain — they audit the formatting.
What an anchor actually is
Let me be precise about the difference between a real analysis and a generated one, because the distinction is not length, vocabulary, or structure. It is anchoring.
An anchored claim has a traceable path back to a primary source. "The protocol's TVL dropped 40% over seven days" is anchored if I can point to the contract addresses, the block range, and the query. "The protocol has strong fundamentals" is not anchored, because there is no fact underneath it to check. It is a sentence that could be true of anything, which means it is true of nothing.
When I audited the Ethereum Classic chain after the 51% attack in 2017, I did not write a report about "community resilience." I traced transaction hashes for six weeks. I found three specific gaps in the response to a $3.6 million theft. Every sentence I published mapped to a hash, a block height, a timestamp. The community did not thank me. But the analysis survived contact with reality, which is the only test that matters. Code, not charisma, dictates reality in this industry — and the same is true of research. A claim either resolves to an on-chain fact or it does not.
This is why the N/A report is not a failure. It is the only honest output a pipeline can produce when its input is empty. The professional who wrote it refused to fabricate, and that refusal is the entire value proposition. The problem is that the market does not price refusals. It prices confidence. So the honest pipeline loses the contract to the dishonest one, and the dishonest one scales.
I have seen this pattern before, in a different register. In 2021, while the market celebrated OlympusDAO's TVL records, I spent three weeks decompiling the bonding contract. The recursive yield mechanics relied on an infinite minting loop. High yields were pre-loaded exit liquidity, mathematically. The report I published predicted a 90% devaluation within six months, and it went viral not because it was hopeful but because it was undeniably accurate. The reason it was accurate is that every claim anchored to a line of Solidity. I could not have written it from a template. Nobody could.
The mechanics of plausible fabrication
The dangerous part of AI-generated research is not that it is wrong. It is that it is wrong in a way that is hard to falsify. A hallucinated number is easy to catch if you check it. A hallucinated adjective is almost impossible to catch, because there is no ground truth to compare it against.
Consider the category of claim I call the stablecoin assertion. A stablecoin's peg is a statement about collateral, redemption mechanics, and liquidity under stress. You can verify it — you can look at the reserve composition, the redemption queue, the secondary-market depth. But a generated report will not do that. It will say the stablecoin is "well-collateralized" because that is the sentence that appears in 90% of the documents in its training set. The adjective is unverifiable, which is exactly why it propagates. It survives because nobody can point to the moment it became false.
The same logic applies to every soft claim in the nine-dimension framework. "Developer momentum" — is that commit count, unique contributors, or funded teams? "Sustainable emissions" — sustainable relative to what revenue, over what horizon, under what price assumption? "Solid backing" — which investors, at what valuation, with what liquidation preferences? The framework asks the question. The generated answer skips the definition. And a question without a defined answer is a machine for producing confident nonsense.
I watched this dynamic play out in real time during the Terra collapse in 2022. I didn't panic-sell. I spent four days analyzing the UST algorithmic stabilizer's delta-neutral hedging failures. The reserve's $2.5 billion in assets was largely illiquid LUNA. The peg was mathematically impossible to maintain, and the oracle feed manipulation accelerated the death spiral. The report I circulated — "The Ponzi Geometry" — was blunt, and it was blunt because it was arithmetic. The reason institutional desks acted on it is that there was nothing in it to argue with. You cannot debate a liquidation cascade. You can only model it.
A generated report on the same project, published a week before the collapse, would have said the peg was "robust" and the ecosystem "thriving." It would have been wrong, and it would have been un-falsifiable until the moment it was catastrophic. That is the failure mode. Not noise. Confident noise, in the register of expertise, with no anchor and no accountability.
The gas-unit test
I have a simple heuristic for separating real analysis from generated analysis, and I have used it for years. I ask: what is the unit?
Real analysis has units. Risk is measured in gas, in basis points, in block confirmation times, in liquidation thresholds, in redemption queues. I measure risk in gas units, not in hope — and I mean that literally. When I look at a protocol, I want to know what a worst-case transaction costs, what a reorg costs, what a governance attack costs in capital terms. Those numbers are falsifiable. They can be wrong. That is what makes them useful.
Generated analysis has no units. It has adjectives. "Strong." "Robust." "Sustainable." "Aligned." These words describe nothing measurable, which is precisely why they are safe to emit. A model that says "the team is strong" can never be contradicted by a fact, because "strong" was never defined. A model that says "the team's unlock schedule releases 12% of supply in month six" can be contradicted instantly — which is why the generated report almost never says it.
The presence of units is the single most reliable signal that a human did the work. If you take one thing from this essay, take the test. Open any research report — generated or not — and count the numbers that could be wrong. If there are none, you are not reading analysis. You are reading a horoscope with a token ticker attached.
Why the bear market makes this worse, not better
You would think a bear market would discipline the research market. Less capital, fewer projects, more scrutiny. The opposite is happening, and the reason is psychological.
In a bull market, readers want confirmation. A report that says "strong momentum" gives them what they want, and the price action agrees for a while, so nobody audits the method. In a bear market, readers want certainty. They want to know whether the thing they are holding is going to zero. That is an even stronger demand for confident answers, and it arrives at exactly the moment when honest answers are most likely to be "I don't know yet."
The generated report fills that gap. It says what the reader is desperate to hear — that the asset is safe, that the drawdown is temporary, that the fundamentals are intact — and it says it in the voice of an analyst. The reader cannot distinguish the voice from the substance, because the substance was never defined. So the market pays for comfort and calls it research.
I have watched five major cycles now. Every bear market produces the same crop of confident forecasts, and every bear market destroys most of them. The analysts who survive are the ones who wrote N/A when they didn't know, and took the reputational hit for it. The ones who fabricated died with the projects they shilled. This cycle, the fabrication is automated, which means the death rate is going to be higher and the bodies harder to identify, because the reports all look the same.
There is a second-order effect, and it is the one that worries me most. When enough generated reports circulate, they become training data for the next generation of models. The hallucination compounds. A fabricated claim about "developer momentum" enters the corpus, gets cited by the next pipeline, and acquires the patina of consensus. After three iterations, nobody remembers that the original claim had no anchor, because there are now a hundred documents that repeat it. The lie has become a fact by weight of repetition. Chaos is just data waiting to be compiled — but so is nonsense, and the compiler does not care which is which.
The automation limitation I keep warning about
Last year I spent two weeks simulating the first major exploit involving an autonomous AI agent trading on-chain. An agent was manipulated into signing a malicious permit because of a subtle gas optimization flaw in the ERC-20 allowance interface. The technical detail matters, but the lesson was structural: the agent had no contextual understanding. It could not tell the difference between a legitimate approval request and a social-engineering payload wrapped in valid bytecode. It processed the transaction as a transaction. The code didn't stop it, because the code had no model of intent.
The same limitation applies to research generation, and it is the same limitation in a different domain. A model has no model of its own ignorance. It cannot distinguish between a field it can fill from evidence and a field it is filling from pattern. To the model, both look identical: a prompt, a continuation, a completion. There is no internal signal that says "I am guessing now." That signal exists only in a human, and only if the human is trained to feel for it.
This is why I keep returning to human-in-the-loop verification. Not because humans are smarter — in narrow tasks they are not — but because humans can be held accountable. A human who writes "N/A" can be asked why. A human who writes "strong fundamentals" can be asked for the anchor, and will have to either produce it or admit they don't have it. Accountability is the mechanism that forces the anchor. Without it, the generation is unconstrained, and unconstrained generation in a market that rewards confidence will always drift toward fabrication.
The regulatory-technical bridge matters here too. In 2024, when spot Bitcoin ETFs gained approval, I scrutinized the custody solutions proposed by the major asset managers. Three of them relied on legacy banking infrastructure that violated the core principle of self-sovereignty. I published a comparative analysis of the cold-storage multi-sig thresholds, and the finding was uncomfortable: "institutional grade" often meant "centralized control." That analysis required legal and technical knowledge simultaneously — you had to read the custody agreements and the key-management architecture together. No template captures that. No pipeline generates it. It comes from sitting with two documents that don't want to be read together and forcing them into the same frame.
The generated report cannot do this, because it has no stake in the answer. It does not care whether the custody is centralized. It cares whether the sentence reads as complete. And a sentence reads as complete long before it reads as true.
What the honest N/A actually tells you
Let me defend the empty report, because I think it is being misread by everyone who encounters it.
When a pipeline returns nine dimensions of N/A, it is telling you something specific and valuable: the input was insufficient to support any conclusion. That is a fact about the input, and it is a true fact. It is the analytical equivalent of a blood test that comes back inconclusive — not a diagnosis, but an honest statement of the test's limits. The professional who ships that report is doing you a favor, because they are refusing to convert their ignorance into your risk.
The dishonest pipeline does the opposite. It converts its ignorance into your risk, and it charges you for the privilege. The 6,000-word report full of specific claims about an L2 that was described to it in one sentence is not a research product. It is a liability transfer. You are buying confidence you cannot verify, from a system that cannot be held accountable, about an asset you are about to hold through a drawdown. The asymmetry is obscene.
There is a version of this that is genuinely useful, and it is worth describing precisely. A well-built pipeline does three things. It anchors every claim to a primary source and refuses to emit a claim it cannot anchor. It reports its own coverage — how much of the input was substantive versus how much was inferred. And it surfaces its N/A fields as a first-class output rather than a failure state, because the gaps are often more informative than the fills. A report that says "I could not verify the token unlock schedule" is telling you exactly where to look next. A report that says "emissions are sustainable" is telling you nothing and hiding the fact that it told you nothing.
The most valuable sentence in any research report is the one that admits what the analyst could not determine. That sentence is the anchor for everything else. It is the proof that the analyst knows the difference between a fact and a filler. Remove it, and the whole document becomes unfalsifiable, which is another way of saying worthless.
The contrarian case: what the AI bulls actually got right
Here is where I have to be honest against my own instincts, because the case for automated research is not stupid, and pretending it is would make me as unreliable as the pipelines I am criticizing.
The bulls are right about coverage. A human analyst can read maybe a dozen projects deeply in a month. A pipeline can read a thousand shallowly in an hour, and for the top of the funnel — the initial screen, the "is this worth a human's time" question — shallow coverage at scale is genuinely valuable. The mistake is not using the pipeline. The mistake is trusting its shallow output as if it were deep.
The bulls are also right that humans are biased in ways machines are not. A human analyst who has spent six weeks on a project becomes emotionally invested in the conclusion. A human who is paid by the project becomes structurally compromised. A model has no such loyalty — it will describe a project's flaws as readily as its strengths, if the flaws are in the input. The problem is not that the model is biased. The problem is that it has no relationship to the truth of the input at all, so it will describe flaws that aren't there with the same fluency it describes flaws that are.
And the bulls are right that the alternative to automated research is not rigorous human research. It is no research. Most retail participants in this market read a Twitter thread and a price chart. A generated report that at least structures the questions — technical, token economics, regulatory — is a marginal improvement over vibes, even if the answers are hollow. The structure has value. The confidence does not.
So the contrarian conclusion is not "abandon the pipelines." It is "demote them." Use the pipeline for triage, for coverage, for the first pass that tells you where to point a human. Then point the human. And crucially, never let the pipeline's output stand as a conclusion on its own, because the pipeline cannot tell you which of its sentences are anchored and which are hallucinated — and neither can you, without doing the work it skipped.
The fork was inevitable; the error was optional. The fork here is between automation as a tool and automation as an oracle. The tool is fine. The oracle is a liability, because it has learned to speak with the authority of verification without performing any of it.
Where this leaves the reader
The bear market has a way of clarifying what actually matters, and what matters now is provenance. When you read a research report — human or generated — ask one question: what is the weakest link in the chain from this sentence back to a fact? If you cannot identify the chain, there is no chain, and the sentence is decoration.
I have spent my career refusing to write promotional whitepapers during the ICO frenzy, tracing transaction hashes nobody asked me to trace, decompiling contracts while the market celebrated. The reason is not asceticism. It is that the only durable product in this industry is a claim that survives contact with reality. Everything else is marketing, and marketing does not survive a drawdown.
The generated report that returns N/A is, paradoxically, the most trustworthy artifact in the current research market. It is the only one that has told you the truth about its own limits. The 6,000-word report full of confident adjectives is the dangerous one, and it is dangerous precisely because it looks like what you were taught to trust.
So the forward-looking question is not whether AI will replace crypto analysts. It is who signs the report. When a pipeline emits a conclusion, a human puts their name on it, and that name is the only thing standing between a hallucination and your portfolio. The industry is currently building pipelines that scale the generation of conclusions while scaling away the accountability for them. That is not a research industry. It is an accountability-laundering operation with a formatting layer.
I measure risk in gas units, not in hope. And the risk in 2026 is that we have automated the production of hope and called it due diligence. The code didn't lie — it just had nothing to say. The lie was in the reading.