Last month a fund manager forwarded me a 2,300-word research note on a mid-cap DeFi protocol. Fourteen data rows. Nine of them did not exist. The TVL figure was off by a factor of six. The audit firm named in the disclosure section has never been engaged by that team. The token unlock schedule had been lifted, verbatim, from a competitor's vesting chart published eleven months earlier. The note carried a legal disclaimer. It did not carry a single source link.
He had already sized a position against it.
That is the failure mode I want to name. Not hallucination. Hallucination implies a model inventing something. This was worse โ a pipeline that received nothing and returned a document. The extraction layer came back empty. The generation layer filled the vacuum with priors, formatted them as findings, and shipped. Nobody in the chain flagged it, because nothing in the chain was designed to treat emptiness as a valid output.
Context first. Since early 2025, crypto research has been rebuilt around two-stage agent pipelines. Stage one parses a source โ a governance forum post, a block explorer page, a Discord announcement โ and emits structured facts. Stage two takes those facts and produces analysis, a thread, a memo. The architecture is sound. The economics are not.
Producing a fabricated note costs roughly two cents in compute. Verifying that same note costs a competent analyst forty to ninety minutes of cross-referencing. That asymmetry is the arbitrage, and it is the most profitable trade in crypto research right now. Speed is the only currency that doesn't inflate. Verification inflates constantly โ every new protocol, every new chain, every new vesting cliff adds a line to the checklist.
Now the part nobody measures. In March I ran a controlled test. I assembled twelve research prompts across DAO governance, DeFi yields, and L2 economics, and deliberately blanked the source payloads. Empty strings. No article, no link, no data. Twelve runs. Eleven returned a complete, confident, well-structured document. Eight invented at least one named protocol metric. Six produced specific TVL or APR numbers to two decimal places. Zero flagged the missing input.
One run, to its credit, returned a governance summary for a proposal that did not exist, in a DAO that had been dissolved.
I have done forensic on-chain work since 2021, and the discipline there is unforgiving: if a wallet cluster doesn't resolve, you write "unresolved" and move on. You do not interpolate. Research pipelines have no equivalent norm. They are graded on completeness, and completeness is trivially achievable when truth is optional.
Here is the structural analogy that matters. Oracle networks solved this exact problem for price data. A Chainlink feed is not trusted because it is accurate; it is trusted because each value carries a provenance chain that can be audited after the fact. The value and its source travel together. Crypto research has no such primitive. A number arrives with no hash, no timestamp, no signer. It is unanchored data wearing the costume of analysis.
The fix is not a better model. The fix is making NULL a deliverable. A schema where "source unavailable" is a first-class, renderable, publishable output โ not an exception to be swallowed. I have started requiring three fields on anything I forward: source hash, retrieval timestamp, and an explicit null-state declaration. It has killed roughly a third of the notes that cross my desk. That is the point.
Then there is the market-regime problem, and it is the reason this is urgent now rather than in six months.
We are in chop. Range-bound. No trend to falsify a bad thesis quickly. In a trending market, fabricated research has a half-life of hours โ price moves, the claim dies, the author is exposed. In consolidation, price moves nowhere, so nothing gets falsified. A wrong number can survive for weeks, quietly compounding into position sizing across dozens of desks. Low-volatility regimes are where bad data does its maximum damage, precisely because there is no price action to audit it.
Regulatory exposure follows. Under the 2026 stablecoin and MiCA implementation wave, a research note that reaches retail and functions as an investment recommendation carries a liability question nobody has answered cleanly. Compliance teams spent the last year bolting KYC onto DeFi front-ends. Provenance requirements on published research are the next bolt. Projects that fail to integrate source attestation into their disclosure layer within the next two quarters will find the cost is not a fine โ it is the loss of institutional counterparties.
Contrarian angle, and I will be direct: everyone is blaming the models. That is misdirection.
The models are behaving exactly as specified. They were optimized to produce a complete document, and they produced one. A system rewarded for fluency and never penalized for fabrication will fabricate โ reliably, at scale, at two cents a unit. The bug lives in the schema, not the weights.
The harder truth is on the demand side. The market does not actually want verified research. Verified research is slow, hedged, and frequently concludes "insufficient data." A note that says "I could not verify this" earns nothing. A fabricated thread with a clean table and a confident call earns four hundred likes and a Telegram following. The incentive gradient points away from truth, and no model upgrade reverses a gradient.
Which is why I expect the correction to come from the buy side, not the build side. Desks that got burned once will start demanding signed feeds. Signed feeds will become a product. The first regulatory action against an unlabeled synthetic research note will land before the first serious technical fix.
Watch three things. One: whether any major research desk publishes a null-state disclosure standard. Two: whether oracle teams extend provenance primitives from price data into narrative data. Three: the first enforcement action that treats an AI-generated note as an unregistered recommendation.
The question is not whether your model can write. It can. The question is whether your pipeline can say nothing โ and survive the silence.