The Null-Input Problem: How AI Slop Is Quietly Rewriting Crypto's Risk Map

The document landed in my inbox at 2 a.m. Paris time, forwarded through four group chats, its letterhead stripped. It was a machine's refusal notice. An AI analyst had been asked to produce a nine-dimension teardown of a token and had instead returned a single verdict: input empty, valid information points — zero, analysis aborted. No price target. No tokenomics chart. No team breakdown. Just a table of missing fields and one line I could not shake: a fabricated deep report is more dangerous than no report at all. By morning it had been screenshotted across three trading desks. Not because it was clever. Because every analyst on those desks recognized the thing it was refusing to do. Here is the mechanism nobody priced in. Since late 2024, the marginal crypto "research note" is no longer written by a human staring at a block explorer. It is generated — assembled from a headline, a ticker, and whatever the model half-remembers about a project. The economics are brutal and simple. A human analyst costs a salary and a week. A content farm costs a few dollars of inference and forty seconds, and it can ship two hundred token "deep dives" a day. Google's 2026 algorithm refresh was supposed to kill this. It did not kill it; it repriced it. The update rewards "information gain" — content that tells the reader something they could not have found by reading the top ten results. Here is the trap: a model that hallucinates a custodial clause in a filing technically produces information gain. It is new. It is just not true. So the slop did not disappear. It learned the new vocabulary. Now every generated report opens with a "proprietary data point" and closes with a "non-consensus view," and neither one exists. I spent 2017 through 2019 auditing pre-mainnet contracts, and the lesson that stuck was not about reentrancy. It was about the difference between a claim and a receipt. A whitepaper is a claim. A verified contract address on Etherscan is a receipt. Fabricated research is dangerous precisely because it is fluent in the language of receipts while holding none. Three tests separate a real note from a generated one, and none of them require trust. First: does every number resolve to a source you can independently fetch? A legitimate report cites a contract address, a block height, a governance proposal ID. A generated one cites "on-chain data" and "recent filings" and moves on. Pull the thread. If the citation cannot be pasted into a block explorer and return the same number, it is decoration. Second: does the timeline actually close? Models are compulsive time-travelers. They will confidently place a mainnet launch in a quarter where the project's GitHub had a single commit. I have watched a "research" note describe a protocol's TVL migration six weeks before the migration's governance vote was even proposed. The math was beautiful. The calendar was fiction. Third, and this is the one people skip: does the conclusion survive a null input? This is the discipline the leaked document was practicing. Strip the token's name, strip the ticker, strip the narrative, and ask whether your thesis still holds on raw on-chain behavior alone. If the entire bull case evaporates the moment you remove the branding, you were never analyzing a protocol. You were analyzing a story. And stories are exactly what the market overpays for. Here is where it gets uncomfortable. Slop is not a victimless crime committed against readers. It is a liquidity event. When a fabricated note circulates widely enough, it does not just mislead — it moves. Wallets trade on it. Price follows. And then the price becomes the evidence, and the next model scrapes that price and calls it confirmation. The fabrication launders itself into a fact through the only mechanism crypto trusts: volume. The chart lies. The volume speaks — except when the volume is quoting a hallucination. I learned to watch for a specific fingerprint during the 2020 DeFi Summer, when I was livestreaming yield-farm breakdowns to an audience that mostly could not read the contracts. The tell was never the headline number. It was the velocity of the explanation. Real analysis is slow and lumpy — it stalls, it contradicts itself, it says "I do not know yet." Generated analysis is frictionless. It answers everything, instantly, in the same confident register. Fluency that never stumbles is not expertise. It is autocomplete wearing a suit. The contrarian read — the one the desks forwarding that screenshot missed — is that the leaked document is not a warning about AI. It is a demonstration of the only edge left. Everyone is panicking about fabricated research flooding the market. But flooding cuts both ways. When ninety percent of published analysis is generated, the scarcity value of a verifiable, sourced, deliberately incomplete note goes up, not down. The null output is not a failure state. It is the rarest signal in the market: a source that knows what it does not know. The danger is not that machines write research. It is that we stopped requiring receipts. Panic sells. I just watch — and what I am watching now is a slow bifurcation. On one side, an infinite supply of confident, citation-free narrative. On the other, a shrinking set of documents that can survive being checked. The market will keep pricing the first. The first will keep failing. Alpha does not wait for permission — and it does not wait for a model to finish hallucinating a thesis either. The next twelve months will be decided by a boring question, asked fast: can I verify this before the candle closes? Build the habit of stripping the ticker. Read the receipts. And when a report hands you a perfect answer with no source, treat the fluency itself as the red flag. The most dangerous document in crypto is not the one that says nothing. It is the one that says everything and means none of it.