A nine-part research framework — the kind now stapled to exchange dashboards, Telegram bots, and "AI analyst" newsletters across crypto — was fed a source document last week and returned a full verdict. Nine dimensions. Clean headers. A confidence column. And underneath all of it, one repeating token: N/A. The pipeline had a single rule hardcoded at the top: do not manufacture conclusions from empty data. It obeyed. That rule is now the rarest asset in this market.
Because everywhere else, the opposite is happening. Feed a void into a model tuned for engagement and you get a thesis, a price target, and a twelve-post thread. On a sideways tape, that output gets consumed as signal. It is not signal. It is formatting — a shape that resembles analysis because the model has absorbed ten thousand analyses and learned the silhouette.
The numbers are not exotic. A mid-size aggregator running one of these pipelines can push forty to sixty "analyst notes" a day. Multiply that across the newsletters, bots, and Discord channels that resyndicate them, and the same unpacked claim touches a hundred thousand retail wallets before anyone checks whether the underlying fact exists. Velocity without verification is just amplification.
Why now: the signal drought
Bitcoin has been chopping in a range for months. Range-bound markets do something specific to information markets: they starve them. No breakout, no liquidation cascade, no surprise ETF headline. Demand for "signals" climbs while the supply of genuinely verifiable events collapses. That gap is the whole business model.
The tooling industry read that correctly. Through 2025 and into this year, every exchange, aggregator, and analytics shop shipped some version of an AI analyst. Most are thin wrappers over a language model with a crypto-flavored system prompt. They are cheap to build, cheap to run, and they produce output on demand — which, in a sideways market, is the only product anyone wants to buy.
In a range, readers don't want conviction — they want a map. Which means the honest product is not a thesis. It's a set of signals: who is actually shipping, where liquidity is quietly repositioning, which contracts changed. That is the work, and it is unglamorous.
I have watched this pattern before, from the other side. In late 2017, during the CryptoKitties congestion, I didn't wait for a press release — I sat on the Ethereum mainnet and watched gas climb past 500 Gwei block by block, then pulled Dapper Labs' backend devs into a Discord call and confirmed the pause-contract decision before the announcement hit. Two hours, one breakdown, real block numbers. Nobody was selling me a narrative. The chain was the source.
That is the opposite of what the current research layer does. Today's pipelines are trained to produce the shape of an answer, and in a drought the shape is all anyone can afford to look at. The nine-dimension template — technicals, tokenomics, market, ecosystem position, regulatory, team and governance, risk, narrative, transmission — is a perfectly reasonable scaffold. It is not evidence.
The three-way laundering of empty input
Here is the mechanism, and it is worth naming precisely, because "the AI hallucinated" is a lazy diagnosis that lets the operators off the hook.
First, template inheritance. The model knows what a crypto analysis looks like. It has seen the headers. Given a thin prompt, it fills slots. Technical position: unstated. Token model: unstated. Confidence: still printed. A scaffold with every beam missing.
Second, retrieval drift. Retrieval-augmented systems pull whatever sits nearest the query. In crypto, "nearest" is a swamp: recycled press releases, paid KOL threads, three-week-old exchange blog posts, and a Reddit comment from an account with a six-day history. The system cites it all with equal weight because it cannot tell a primary source from a repost. Chainlink's own oracle design is a neat metaphor here — a "decentralized" feed supplied by a permissioned node set, where the decentralization is branding and the latency is real.
There is a concrete version of this I keep running into. A project announces a partnership. The partnership is a logo swap and a co-marketing post. Four hours later, an AI note frames it as "ecosystem expansion," the note gets quoted, the quote gets quoted, and by tomorrow the logo swap is infrastructure. Nobody lied at any single step. The system just averaged its way from nothing to something.
Third, confidence theater. This is the part that actually costs people money. A confidence score attached to a field that reads N/A is not a confidence score. It is decoration. It converts an absence into a number, and numbers get screenshotted, and screenshots become positions.
Fourth, transmission drift. This is the one that leaves the terminal. A note generated from nothing doesn't stay on the desk — it gets echoed into a KOL thread, then into a subgroup chat, then into a position, then into a support level that gets cited as "the market's view." By the third hop, the empty input has become consensus, and consensus is the hardest thing in this industry to argue with, because arguing with it sounds like arguing with the crowd.
The guardrail pipeline I described did the boring, correct thing: it stopped. It returned a framework with every substantive line marked as insufficient. No thesis. No target. Just an honest map of what it did not know. Sixteen years in this industry, and that is still the rarest output I see.
What actual verification costs
I have a bias, and I will state it: I do not trust an analysis I cannot re-derive from a contract, a block, or a signed message.
During DeFi Summer in 2020, I treated protocol docs like assembly instructions and deployed real capital — small, but real — through Uniswap and Compound, just to feel impermanent loss instead of describing it. That is how I caught the discrepancy in Curve's early emission schedule, and the unpatched admin-key exposure, hours before the token launch. I was not smarter than the desk analysts. I was just unwilling to publish a claim I hadn't touched.
Nothing about that method has changed with better models. When a token emission schedule didn't reconcile with the audit, I didn't ask a chatbot. I opened the contract. When Terra de-pegged in May 2022, I didn't read the panic coverage — I worked with independent security researchers to trace the flash-loan sequence against Anchor's ledger, watched the causal chain in order from liquid staking derivatives outward, and pivoted the framing from "code failure" to "regulatory vacuum" only after the treasury movements made that the truer story.
When the spot Bitcoin ETFs cleared in early 2024, the lag wasn't in the approvals — it was in custody. I went around the PR desks to an operations manager and asked the technical questions directly: how are the multi-sig keys managed, who holds the cold-storage shards, what happens on a signer failure. That conversation produced something no model could have generated, because it wasn't public. It was a person telling me how the machine actually runs.
Each of those took hours I could have spent formatting. That trade — hours for structure — is exactly the trade the market is currently making in the wrong direction.
There is a version of this argument that ends in "just use better tools." I don't buy it. Tools are downstream of incentives. A better model pointed at a swamp returns a better-formatted swamp. The fix is upstream: pay for primary data, cite the block, timestamp the claim, and publish the null results alongside the hits.
The contrarian angle: the models are innocent
Everyone wants to blame the models. Wrong target.
The empty-input problem is not an AI failure. It is a data-market failure. Crypto's public information layer is structurally thin, and it is thin by design. Press releases are marketing. Influencer threads are inventory. There is no economic incentive to publish the boring verification, because verification is expensive and non-viral, while a confident framework is cheap and screenshots beautifully.
The one mechanism I have seen actually pay for public goods without a committee deciding whose friend gets funded is Optimism's RetroPGF — retrospective, outcome-based, no grant-committee gatekeeper. It is not perfect. It is the only one that isn't nepotism wearing a governance hat.
There is an uncomfortable incentive stacked on top of all this. The desks producing these notes are not paid by accuracy. They are paid by engagement, seat retention, and the appearance of coverage. A blank page doesn't trend. An "analysis" — even one built on N/A — does. The operators know this. Some of them are quietly rebuilding around retrieval from primary sources: block explorers, contract ABIs, governance forums, signed attestations. That work is slow and invisible, which is exactly why it will be a moat when the market stops being bored.
So the real question is not "why is the AI making things up." The real question is: what is the market paying for? Right now it pays for the silhouette. Which means the silhouette is what it gets. That is not a technology problem. It is a buyer's problem, and buyers can fix it.
What to watch
Watch for the first research product that ships "insufficient input" as a visible feature instead of an error state. Watch for desks that publish their null results — the projects they looked at and couldn't verify. That is the tell. Because the drought will not last. Ranges resolve. Liquidity returns, cascades return, and the tape starts handing out verdicts to everyone who published a confident call. That is when the receipts get checked — and the desks holding nothing but N/A will find out, publicly, which of their nine dimensions were ever real.
When that moment comes, the survivors will not be the ones with the cleanest templates. They will be the ones who knew the difference between a model that had something to say and a model that had something to prove.