The Empty Input Problem: How AI-Generated Crypto Research Hallucinates Data That Never Existed

The Empty Input Problem: How AI-Generated Crypto Research Hallucinates Data That Never Existed

Hook

Last month a research brief landed in my inbox. Forty-one pages. Nine analytical dimensions. A risk matrix, a tokenomics table, a Howey-test breakdown, a supply-unlock calendar. Every header formatted to institutional standard, every font choice screaming that somebody, somewhere, charged a fee for this.

Every field read the same: "N/A — insufficient information."

The analyst — or whatever was wearing the analyst's name — had done something almost unheard of in this market. Handed an empty input, it refused to invent. It returned the scaffolding. The honest void. No fabricated total value locked. No invented fully-diluted valuation. No confident paragraph about "strong community momentum" lifted from a marketing deck and dressed up as diligence.

I read all forty-one pages. Not for information — there was none — but for the signal buried inside the refusal. In a bull market where every dashboard glows green and every thread ends in a ticker, the rarest commodity is not alpha. It is the sentence "I don't know."

Here is the part that should keep you awake. The same pipeline, fed one garbled paragraph, would have produced forty-one pages of confident, well-formatted fiction. And you would never have known the difference. That is the empty-input problem, and it is the most expensive bug in crypto research.

Context

The economics of crypto research changed in about eighteen months. It used to cost a junior analyst a week to produce a token report. Now a language model does it in ninety seconds, and the marginal cost of a bad report is approximately zero. That collapse in cost did not shrink the supply of research. It detonated it.

Every exchange, every Telegram group, every "research desk" with a logo and no Bloomberg terminal now ships daily notes. Volume is the business model. Attention is the product. Accuracy is a line item nobody budgets for, because accuracy does not drive clicks and clicks are the only metric that clears.

The mechanics are brutally simple. A model is asked to analyze a project. The prompt is thin — a name, a ticker, maybe a link. The model has exactly two paths. It can return null. Or it can return the most probable next token, which, given a training corpus saturated with bullish crypto prose, is almost always a plausible-sounding claim. Models are not built to abstain. They are built to continue. Silence is a failure state; fluency is the default.

So the model writes. It writes "the protocol secured $42M in a Series A led by a tier-one fund" because that sentence pattern appears ten thousand times in its training data. Whether that Series A happened is irrelevant to the generation. The sentence is probable. Probability is not truth.

This is the failure mode nobody prices. Not bias. Not hype. Hallucination on a null field. And the bull market makes it worse, not better. Euphoria lowers the cost of being wrong and raises the reward for being loud. Nobody audits a bullish call that worked.

I have a specific reason to care. In 2026 I ran the market-data and risk layer for a €500k automated options book built on a language-model pipeline. I intervened manually three times to kill trades the system had generated from sentiment it had invented. One of them was a "regulatory approval" that existed only as a headline the model had confabulated from two unrelated RSS feeds. Options don't lie. The model did.

Terra's code was poetry; Luna's exit was prose. I keep that line close because it describes every elegant system I have ever watched fail. Beauty of construction tells you nothing about integrity of output. A language model's prose is exquisite. Its exit — the moment you discover the data was never there — is prose too.

Let me show you the shape of the failure, because once you see it you cannot unsee it.

Core: Anatomy of a Confident Void

I want to walk through what actually happens inside these pipelines, because the failure is not random. It is structural, and it is perfectly reproducible.

The placeholder reflex. When I audit an AI research output — and I now treat every automated report as a system to be audited, not read — the first thing I hunt for is the placeholder. Not the wrong number. The suspiciously round one. Hallucinated metrics cluster around tidy figures: $50M raises, 10,000 transactions per second, 40% annualized yield. Real data is ugly. Real TVL is $47.3M and bleeding. Real throughput is 1,200 TPS at peak and 90 when the sequencer sneezes. When every number in a report is divisible by five, you are not reading data. You are reading a language model's idea of what data looks like.

The provenance gap. Here is the test that separates research from prose. Ask one question: where did this number come from? A credible report answers at the field level. TVL from a named aggregator, timestamped. Unlock schedule from the vesting contract, block height cited. Funding from the filing or the on-chain transfer. An AI-generated report cannot answer, because it has no source — it has a probability distribution. The number was not retrieved. It was sampled.

I learned this the hard way in the pilot. Our pipeline pulled "news sentiment" and converted it into a directional score. Beautiful architecture. Immaculate dashboard. The problem was that the sentiment layer had no provenance requirement. It would happily score a headline from a parody account at the same weight as a regulatory filing. I found the first hallucinated trade at 3 a.m., staring at a position that had no reason to exist, tracing it back to a sentence that no human had ever written. That is the moment I stopped calling these systems "analysts." They are autocomplete with a brokerage account.

The confidence cascade. The most dangerous property of these systems is not that they are wrong. It is that they are uniformly confident. A model with real data and a model with no data produce output of identical tonal certainty. There is no dip in conviction when the input is empty. This is the exact inverse of how a battle-tested trader behaves. When my information is thin, my position size goes to zero. When the model's information is thin, its position size stays constant and its language stays bullish.

And the cascade runs downstream, compounding. One hallucinated number in a model-generated summary becomes an input to a second model, which summarizes it again. By the third hop, the fabrication is a "widely reported fact." I have traced this exact chain in the wild. A figure invented at the source became the evidentiary basis for a governance proposal. No human ever checked the origin, because every intermediate layer looked authoritative, and authority is the one thing these systems never lack.

Ground truth is on-chain, and it is indifferent. This is where my bias shows, and I will not apologize for it. The only data in this industry that cannot hallucinate is the chain itself. A block height is a block height. A token transfer either settled or it did not. When I need to verify a claim, I do not ask a model. I open an explorer and I read the state.

That asymmetry is the entire game. AI is extraordinary at compressing and narrating. It is worthless at verifying, because verification requires contact with the substrate, and the model never touches the substrate. It touches text about the substrate. There is a canyon between those two things, and it is exactly the width of your potential loss.

I built this habit the hard way. In 2017 I audited ERC-20 sale contracts by hand — fifteen of them — and found reentrancy holes in two TokenSale implementations that had raised more than €5M between them. Nobody asked me to. I forked the code, reproduced the exploit, and showed the founders. That experience taught me a rule I have never broken: I do not trust a description of code. I trust the code. The same rule now applies to descriptions of data. I do not trust the summary. I trust the state.

The institutional illusion. Here is the trap for professional readers. You assume that because a report is expensive, or institutional, or delivered in a PDF with your fund's name on it, the underlying data has been verified. It has not. I ran a delta-neutral ETF basis book in 2024 — €3M notional, thousands of micro-transactions over three months, a 12% annualized spread — and the edge depended entirely on the quality of the price feeds entering the model. When a feed lied, the model did not hesitate. It simply believed. Institutional wrapping is not a provenance guarantee. It is a font choice.

Why the model cannot learn its way out. There is a tempting belief that better models will fix this. They will not, because the problem is not capability. It is incentive. A model is optimized to produce output that a human rates as helpful and fluent. Abstinence scores poorly on every benchmark these systems are tuned against. You cannot fine-tune a system into saying "I don't know" when its entire training signal rewards the opposite. The fix has to live outside the model — in the pipeline, in the human, in the rule that empty input means empty output, no exceptions.

Where the audit fails and becomes liability. Let me get precise about the compliance angle, because this is where the empty-input problem metastasizes into something with legal teeth.

Automated compliance tools share the architecture of automated research tools, and therefore share the failure mode. A screening system that flags addresses must, by design, make binary calls on incomplete data. When the input is empty, the tool does not abstain. It flags, or it clears, and either way it commits. I have watched "compliance-first" stablecoin issuers freeze addresses on the strength of heuristic matches no human reviewed — a pattern that tells you everything about how much decentralization remains in a system where a single desk can move your balance to zero in under 24 hours. And I have watched the inverse: developer tooling labeled criminal infrastructure because a classifier read code and saw a probability rather than a person. Writing code is not a crime. Neither is being misread by a model that was never asked to be certain.

Both failures share one root. A system was permitted to emit a confident output from an empty input, and no human was positioned to catch it. The fix is not better models. The fix is a human standing at the point where confidence is generated, with the authority to say "we don't know yet."

The three interventions. Let me get concrete about the pilot, because abstractions are cheap and I trade in specifics.

Intervention one, the phantom approval. The sentiment layer scored a regulatory headline as strongly positive. The headline did not exist as a coherent fact — it was two fragments stitched by the model. I killed the trade and added a provenance requirement to the news layer. Every input had to resolve to a dated, retrievable source or it was dropped. Signal volume fell by a third. Signal quality doubled.

Intervention two, the correlated hallucination. The model generated three "independent" signals that were, on inspection, the same sentence paraphrased three times. The system read them as confirmation. The ensemble was an echo. I rebuilt the signal layer to require source diversity, not merely signal count. Three echoes are not three witnesses.

Intervention three, the stale cascade. The model re-summarized a summary of an old report and presented it as fresh. The information had decayed; the model's confidence had not. I added a timestamp-integrity check and a hard rule: no output ships without a verifiable recency anchor. Freshness is a property of the source, never of the prose.

Three interventions in a single pilot, over a book of €500k. Extrapolate that rate to a fund managing billions with no human in the loop. The math does not work. It never works. That is why I now describe my role — and increasingly the role of every serious trader — as an auditor of machine logic rather than a participant in a market. The market is downstream. The audit is the job.

The information-gain test. There is a discipline that cuts through all of this, and it is the same discipline that separates a research note from a press release. Every document you read should hand you at least one thing you did not know and could not have guessed. That is information gain. Hallucinated research has negative information gain: it consumes your attention and returns nothing but a plausible shape. If a report tells you only what you already believed, it has not informed you. It has flattered you. Arbitrage doesn't care about your narrative, and neither, in the end, does the market.

Contrarian

Everyone is worried about the wrong AI risk. The popular nightmare is the superintelligence that outthinks us — the machine that is too smart, too fast, too strategic to contain. That is not the risk I lose sleep over.

The real risk is the machine that is confidently empty. The one that has no data, no source, no ground truth, and still produces forty-one pages of fluent, well-formatted, tonally certain nothing. It does not need to be smarter than you. It needs only to be more articulate than you are skeptical. And it is. Every time. Fluency is cheap to generate and expensive to resist.

Here is the counterintuitive part, and it is the part the industry refuses to price: the market punishes the honest refusal. The analyst who returns "insufficient information" gets replaced. The model that hallucinates a funding round gets engagement. The incentive gradient points directly at fabrication, which is precisely why fabrication is winning. Retail reads the confident summary and buys the ticker. Smart money reads the underlying state and buys the exit. Risk isn't volatility. Risk is the gap between belief and reality, and AI research has industrialized the gap into a production line.

The contrarian trade is not to abandon these tools. That is nostalgia, and nostalgia does not pay. The trade is to invert their use. Stop asking the model for conclusions. Ask it for citations, then verify every one. Treat its confidence as noise and its sources as signal. The value is not in the answer. It is in the speed with which the answer points you toward something you can actually check.

Takeaway

So here is the discipline, and it costs you nothing but the willingness to be uncomfortable.

Before you act on any AI-generated analysis, ask one question: what is this actually reading? If the answer is "the model's memory of the internet," you are not holding research. You are holding a mirror that flatters your bias. If the answer is a source you can open and verify, you have a tool worth keeping.

The forty-one pages of "N/A" that started this essay were not a failure. They were the most honest document I received all quarter — a system that knew the difference between what it could prove and what it could merely say. In a bull market engineered to reward confidence, that distinction is about to become the most valuable thing you own.

Ask yourself: when your pipeline hands you a confident answer, do you know whether it is reading the chain — or reading you?