At 3:14 on a Tuesday morning in February, a market-monitoring script in a Warsaw office pulled nothing back from its upstream feed.
Not a crash. Not a timeout. An empty payload โ a well-formed JSON envelope with every expected key present and every value null. The indexer behind it had rotated credentials overnight. That part is unremarkable. Infrastructure does this constantly, and most of the time the failure is loud.
This time it was quiet. Ninety seconds after the empty response landed, an automated research agent consumed it and produced roughly nine hundred words of confident market commentary. It named a protocol. It described a liquidity migration. It attributed a short-term price swing to whale rotation. Every sentence was grammatical, internally consistent, and built on nothing.
Nobody flagged it for six hours. By then it had been syndicated twice and quoted in a Telegram channel with forty thousand members. What finally broke the illusion was not a sophisticated forensic tool. It was a human editor asking a pedestrian question: where did that wallet address come from?
Silence speaks louder than hype. But silence left unattended gets filled, and the machinery that fills it does not blush.
Crypto has always run on narrative, and narrative has always carried noise. That is not the new part. What has changed is the texture of the noise.
In 2017, the noise sounded like a whitepaper promising a decentralized dental records protocol, written in a hurry by three people who had never deployed a contract. You could smell it. I spent six months that year manually auditing time-crowdsale contracts for three mid-tier ICOs in Warsaw, and the tells were human. Reentrancy bugs sat next to typos. Founders overpromised in ways that were visibly nervous. When I put fifteen thousand dollars of my own savings into a healthcare-token project, it was because I had read the code and could see that the team had actually done the boring work.
By 2020, the noise had a different accent. DeFi Summer produced yield numbers that made no sense, and the tells were still human: anonymous teams, unaudited forks, Telegram admins who answered questions at three in the morning. I wrote a guide on Aave's risk parameters that year, interviewing twelve risk managers, and the central lesson was that user safety has a structure you can inspect. The people who got hurt were the ones who skipped the inspection.
In 2022, the noise became a firehose. During the Terra collapse, I ran a crisis desk for a community of ten thousand, fact-checking rumors in real time. Three weeks of on-chain verification. The rumors were human-made, and human-made rumors have a fingerprint โ urgency, screenshots without timestamps, a suspicious eagerness to be shared.
Then came 2024, and the ETF. I led a series profiling small Polish businesses using Bitcoin rails for cross-border payments, thirty interviews deep. The noise that year was different again: institutional-sounding language wrapped around thin claims.
But every one of those cycles shared a property. The fabrication layer was human, and humans leak. They get excited. They get sloppy. They leave a trail of motive.
That trail is gone.
The most dangerous output of an automated research system is not a wrong answer. It is a fluent answer to a question that had no data behind it.
Those are different failure modes, and the industry keeps treating them as one. A wrong answer can be corrected. A wrong answer has a shape you can argue with. A fluent answer to a null input is worse, because there is no argument to be had โ the sentence is not false, it is unmoored. You cannot falsify a claim that was never anchored to an observation in the first place.
Let me describe the mechanics precisely, because this is where most coverage goes soft.
When a data feed returns null, a well-behaved system should propagate the null and stop. That is called null propagation, and it is boring, correct, and easy to implement. Instead, most modern pipelines do one of three things.
The first is silent substitution. A schema validator sees a missing field, applies a default โ often the last known value โ and marks the record valid. This is where you get a dashboard showing a protocol with forty percent of its liquidity intact when the actual figure is unknown. The dashboard is not lying. It is remembering.
The second is interpolation. The system notices a gap and draws a line through it. In a sideways market, where price action is compressed and most metrics genuinely move in narrow bands, interpolation looks like analysis. This is not a coincidence. Consolidation is precisely the regime in which fabricated continuity is least detectable, because the real data would have looked boring anyway. A sideways tape is camouflage.
The third, and the one that actually worries me, is generation. The model receives a null, and because it has no mechanism to refuse, it produces the most probable continuation of the surrounding context. If the surrounding context is a summary of a liquid staking token's week, the model writes about liquid staking. It writes well. And it writes as if it had seen something.
I want to give two concrete examples from this year, because abstraction lets people nod along without changing anything.
Start with Layer 2 sequencing. Several rollups publish "sequencer health" dashboards with metrics that read like decentralization evidence โ batch frequency, uptime, validator distribution. In practice, a large share of those numbers originate from a single operator-controlled endpoint. I have watched one of these endpoints return a null payload during a credential rotation, and I have watched the dashboard above it keep rendering a smooth uptime chart. Nobody interpolated maliciously. The default value was simply defined as one hundred percent, because that is what the schema said a healthy day looked like. Decentralized sequencing has been a slide deck for two years, and the dashboards that claim otherwise can be rendered meaningless by a single unrotated API key.
Second, real-world assets. Tokenized treasury products report supply and yield through a mix of on-chain contracts and off-chain attestations. The on-chain half is verifiable. The off-chain half is a letter. When I have asked teams what happens to their reported AUM if the attestation feed goes down, the honest ones say the number carries forward. That is a three-year storytelling exercise with a null-handling bug at the center, and nobody wants to say the quiet part out loud: the institutions being courted do not need a public chain to know how many bonds they hold.
Both examples share a structure. The data was absent. The interface was confident. The human reader had no way to tell the difference without going to the source, and the source was the thing that failed.
This is what my team and I have been building against since early 2026, in a joint project with a Warsaw-based AI startup. We built a cross-referencing tool that pairs AI-generated sentiment readings against raw on-chain whale movements. The premise is simple: if a generated sentence claims accumulation, there should be wallet activity that sustains it. We published the first open-source dataset on algorithmic manipulation risks, and roughly two thousand independent journalists now use it as a sanity layer.
The most useful thing we learned was not about sentiment. It was about refusal.
A model that cannot output the word "null" is not a research tool. It is a content generator with a research-shaped interface. We started mandating what we call a refusal contract: every automated report must carry an explicit enumeration of which input fields it actually received, and any field that was empty must appear in the output as empty. Not zero. Not "no change." Empty. If the upstream wallet feed did not resolve, the report must say so in the same font size as the conclusion.
Adoption has been slower than it should be, and the reason is instructive. Refusal contracts make reports look broken. An honest report that says "liquidity data unavailable, price action unattributed" reads worse than a fabricated one that names a whale. The incentive gradient points toward fluency. It always has.
There is a detection heuristic worth handing out freely, because it costs nothing to use. Look at the confidence distribution across an automated report's claims. Human-written analysis is lumpy โ some assertions are hedged, some are firm, and the hedges cluster around whatever the author was least sure of. Generated commentary built on a null input has a flat distribution. Every claim carries the same weight. Nothing is hedged, because hedging requires knowing where the uncertainty lives.
Code does not lie, only humans do. But code that has been taught to never say "I don't know" will produce human-sounding lies at machine speed, and that is a genuinely new problem.
Here is where I part ways with the current consensus.
The dominant response to synthetic misinformation is provenance tooling โ watermarking, content credentials, classifiers that estimate whether a paragraph was machine-written. That work is worth doing, and it addresses the wrong layer.
Provenance tells you who wrote something. It does not tell you whether the thing was written about anything. A perfectly watermarked report, cryptographically signed by a verified publisher, can still be a fluent response to an empty payload. You will have established authorship with total certainty and learned nothing about whether the underlying observation existed. That is a substrate problem disguised as an authenticity problem.
Truth is often buried under the noise, but the noise is now generated by systems that were designed to be trustworthy, which means the old smell test is gone.
There is a second blind spot, and it is one I hold myself to. The demand side is complicit. When a feed goes down and a reader is told "no data," they go somewhere else. I have watched our own newsletter open rates react to honest emptiness. The audience punishes refusal and rewards confidence, and automated systems are optimized for the audience. Any accountability framework that ignores reader incentives will fail, because it will be competing against a version of itself that never says no.
And there is a final inversion worth sitting with. Sometimes a null is not a failure. It is the answer.
A protocol with no active LPs. A chain with no unique users. A treasury attestation that stopped arriving three weeks ago because the issuer stopped sending it. In each case, the absence is the most informative data point available, and our entire tooling culture is built to convert absence into continuity. We have automated the exact reflex that loses people money.
I do not think this gets solved by better models. It gets solved by making emptiness legible โ by treating a null payload as a headline rather than an inconvenience.
Watch for three things over the next two quarters, because quiet markets are where this becomes visible. First, whether any major analytics provider publishes an explicit data-availability rate alongside its coverage claims; a provider that cannot state what percentage of its inputs resolved has never measured it. Second, whether rollup teams start publishing raw sequencer endpoint logs rather than derived health scores, which would make interpolation impossible to hide. Third, whether the AI research tools now entering newsrooms ship with refusal contracts by default or as a paid tier.
Every one of those is a test of the same instinct: the willingness to publish a page that says nothing, when saying something would perform better.
One question for the builders. If your model produced a perfect, fluent, nine-hundred-word report this morning, and the feed underneath it returned nothing โ would you be able to tell? Would anyone?