Last month, a nine-section due-diligence template landed on my desk with every field filled in the same way. Not bullish. Not bearish. The same four words, nine times over: “insufficient information available.”
On paper, it was a failure. A research pipeline had gone out to retrieve a token’s unlock schedule, its audit history, its sequencer topology, its legal domicile — and it came back with nothing. Nine dimensions, zero information points. So the system reported zero. It reported zero in the supply table. It reported zero in the risk matrix. It reported zero in the Howey test, which is genuinely difficult, because the Howey test will accept almost any input you hand it.
Twenty minutes later, someone upstream asked for the finished article anyway.
That request — not the empty template — is the story. Because the blank document was the most honest thing produced that week.
I have been editing crypto research long enough to remember when a sideways market meant fewer articles. Not anymore. In a consolidation phase the premium on micro-signals goes up, because readers are waiting for direction and will buy anything that looks like direction, while the cost of producing a plausible-sounding signal has collapsed to near zero. A human analyst on my desk produces two to four genuine deep notes a week. A well-plumbed pipeline produces two hundred. Both land in the same inbox. One of them carries a byline that says “Research Desk.”
The inputs feeding those pipelines are mostly real: on-chain indexers, governance forums, GitHub commit histories, exchange order books, unlock calendars, DEX liquidity snapshots. Those are the fields where a machine can be right. But look closely at the nine-dimension template and you notice something structural. The fields that actually determine whether a token deserves a treasury manager’s attention — contract upgrade authority, legal domicile, team identity, whether the auditors were paid in the token they audited — are precisely the fields that have no API. They live in PDFs, in Discord threads, in a corporate registry in a jurisdiction nobody wants to name. They are the fields where the model has nothing.
And they are the fields where the fabrications are most expensive.
The shape of the tool
Start with the mechanical reason, because the moral reason is less interesting.
A large language model is a completion engine. Show it a header row — Team | Investors | Community | Treasury — and ask what comes next, and it answers based on what followed that header in the text it was trained on. In that corpus, tables are full. Empty cells are rare. The string “N/A — insufficient information” appears almost nowhere in a decade of crypto research writing, because the people who produced that corpus were paid to fill the table. The string “Team: 15%, 4-year vest, 1-year cliff” appears thousands of times.
So the model’s statistical prior points hard toward a plausible fabrication. This is not a bug in the model. It is the model working exactly as specified, on a distribution curated by an industry with no tolerance for blanks.
Silence speaks louder than hype — but only a human can decide to be silent. A model asked to complete a form will complete the form. When you hand a completion engine a template, you have already issued the instruction. The honest blank and the confident lie are the same pipeline, one prompt revision apart.
The template is the accelerant
I spent six months of 2017 manually auditing time-crowdsale contracts in Warsaw, and the lesson that stuck was not about reentrancy. It was about checklists. The format of a question determines the quality of the answer more than the analyst does. A checklist that asks “Is the contract reentrant?” gets a yes or a no. A checklist that asks “Assess contract security” gets an essay — and essays are where reentrancy vulnerabilities go to hide.
The nine-dimension template does the same thing at the narrative layer. It presumes that nine dimensions of information exist for every asset under review. That presumption is not neutral. It forces the model — and the junior analyst, who behaves identically under deadline — to produce content for dimensions that are empty. The output looks like comparability. Four of those nine rows will be invented, and the reader has no way to know which four.
Absence of evidence, sold as evidence of absence
This is the failure mode that costs money, and it is subtler than a fake number.
When a risk field comes back as “N/A,” downstream readers do not read it as “no search was performed.” They read it as “no risk was found.” Those two statements are opposites, and the gap between them closes in about four hours under market pressure.
I watched this happen at scale in 2022. During the Terra collapse I ran a fact-checking team against a 10,000-member Telegram group for three weeks, verifying on-chain data before anything got amplified. The hardest corrections were never false claims. False claims are easy; you post the transaction hash and the thread dies. The hard ones were missing values that people had already filled in with zero. “We have no on-chain evidence of a reserve shortfall” became “the reserves are fine” in the time it took to forward a message. By the time the data arrived, the belief had already settled. We reduced member attrition by roughly 40% against the industry baseline that quarter, and I still believe most of that came from refusing to answer questions we couldn’t answer yet.
The four fields where the polite lie is expensive
Tokenomics. Unlock schedules are the single most fabricatable number in the asset class and one of the most price-sensitive. A hallucinated cliff date is not a rounding error; it is a trade. If a model invents a twelve-month team vest for an allocation that actually unlocks in month three, every reader acting on it is positioned on the wrong side of a supply shock they were told did not exist.
Audit status. “Independently audited” is the cheapest sentence in crypto to generate and one of the most expensive to believe. A real audit has four components: a named firm, a commit hash, a date, and a scope document. If any of the four is missing from a report, the field is empty no matter how confident the sentence sounds.
Sequencer topology. This is where the gap between roadmap and deployment is widest, and where AI summarization does the most damage, because the roadmap is written in language a model loves. Every major rollup’s published strategy contains the phrase “decentralized sequencing.” Every major rollup’s production deployment runs a single sequencer, operated by a single entity, behind a single upgrade key, on infrastructure a well-funded adversary could plausibly pressure. The shared-sequencer projects are real and they are early. An AI reading governance forums and engineering blogs will faithfully summarize the direction of travel and hand you a note that reads like a decentralized network. The chain tells a different story. I have been making a version of this argument in print for two years, and it remains the least popular technically correct sentence in this space: code does not lie, only humans do. Models trained on human text inherit the human weakness — and the weakness is optimism about planned versus shipped.
Real-world assets. Here the fabrication is not a number but a relationship. A pilot on a permissioned ledger with a friendly counterparty becomes “institutional partnership” in the summary. The institution never touched a public chain, never held a public token, and has no stated intention of routing settlement through one. I have watched that sentence get rewritten three times up the chain of a research brief until it read like adoption. It is not adoption. It is a press release with a database behind it.
What the divergence actually looks like
Since early 2026 I have been running a joint research project with a Warsaw-based AI startup to build a verification framework for AI-generated crypto market reports. The method is unglamorous. We take a model’s sentiment output for an asset, pull the corresponding whale movement and exchange net-flow data, and measure the gap. We published the first open-source dataset on Algorithmic Manipulation Risks, and roughly 2,000 independent journalists have used it to triage suspicious reports.
The finding that matters is not that models are wrong. It is that the wrongness has a signature. When a generated sentiment score moves sharply and on-chain flow does not, you are almost always looking at summarized narrative rather than summarized data. When flow moves and sentiment does not, you are usually looking at something real that nobody has written about yet. The gap is the signal. That sounds obvious written down. It was not obvious at 3am in February, and it took four months of false positives to learn which direction of divergence to trust.
The contrarian read
Here is the part I have not seen argued, and I think it is the correct one: the model is not the problem. The incentive is the problem, and the template is the second problem.
Ask what happens to a human analyst who files a nine-dimension report with four fields reading “insufficient information.” In most newsrooms they get a note from an editor. In most research shops the report never ships. The analyst who fills all nine fields with confident prose gets the byline, the client, and the bonus. That selection pressure existed long before language models were good at anything. AI did not introduce fabrication to crypto research; it reduced the marginal cost of fabrication to approximately zero and removed the fatigue that used to cap how much of it one person could produce in a day.
Which leads to the deeper problem with the template itself. Nine dimensions presumes a world in which every asset has nine dimensions of substance. Most do not. A governance token with a three-person team, no revenue, and an admin key held by the deployer does not have a tokenomics section worth writing — it has a tokenomics section worth noting as absent. The template manufactures the appearance of comparability, and then the pipeline fills the manufactured space. We built a nine-box form to make assets sortable, and we got a machine that sorts fabrications.
The honest move is to allow fewer boxes, and to treat an empty box as a finding rather than a gap.
Takeaway
The next premium in crypto research will not go to the desk with the most coverage. Coverage is now free and effectively infinite. It will go to the desk that publishes its nulls — that says, in writing, “we queried six sources, spent eleven days, and found nothing, and we are not going to guess.” Truth is often buried under the noise, but it is also frequently buried under an empty field somebody was too embarrassed to leave empty.
When anyone can produce two hundred briefs a day, the scarce asset is the willingness to publish zero. The desks that learn that will be trusted in the next drawdown. The ones that don’t will keep getting the assignment, keep filling the table, and keep being right — right up until the unlock actually happens.
