On a Tuesday morning in early 2026, a portfolio manager at a Singapore-based quantitative fund did what thousands of institutional crypto traders now do as routine: she opened the pre-trade analysis brief her firm had subscribed to, expecting the usual scaffolding of yield curves, vesting schedules, and liquidity depth tables. What loaded instead was a document containing ninety-three fields marked "N/A," a single repeated phrase—"information insufficient"—and one warning, buried in the metadata, flagging the report as a "high hallucination risk artifact."
The protocol she had requested was real. It had $340 million in total value locked. It had launched a token three weeks earlier. It had raised $28 million from named venture funds. The data was not classified, not private, not behind any paywall. The entire public footprint of the project was indexable. The analysis was empty because the pipeline that produced it had failed six steps upstream—at the layer where raw information is supposed to be ingested, parsed, and structured before any interpretive work begins.
She sent the report to her compliance officer with a single comment: "This is what we've been paying $47,000 a month for."
This is not an isolated incident. Across the institutional crypto research stack of late 2025 and early 2026, a quiet infrastructure crisis has been metastasizing into something more corrosive than any single market cycle could produce. The tools that were supposed to democratize access to high-quality crypto intelligence—AI-driven research pipelines, automated due diligence engines, sentiment-scraping aggregators—have become the single largest source of analytical noise in the industry. And when they fail, they do not fail loudly. They produce documents that look like analysis. They mimic the surface grammar of research. They contain citations, tables, and even the occasional bolded "core insight." But the underlying information density is approaching zero.
To understand the moment we are in, you have to trace the five-year arc that produced it. In 2021, when the first wave of AI-augmented crypto research products launched, the pitch was almost embarrassingly simple: human analysts were slow, expensive, and prone to narrative capture; machines were fast, cheap, and indifferent to consensus. A new generation of founders—many of them former data scientists from TradFi, several of them ex-quantitative researchers from Jane Street and Two Sigma—built startups that promised to replace the artisanal analyst with an automated research factory. The implicit assumption was that the bottleneck in institutional crypto research was human attention, and that attention could be substituted with compute.
By 2023, the product category had consolidated around a recognizable stack. A typical institutional crypto intelligence platform in 2023 ingested roughly forty data sources: on-chain RPC nodes, exchange APIs, governance forum archives, Discord and Telegram scraping modules, GitHub commit logs, code repository metadata, token contract ABIs, vesting smart-contract event logs, and a growing pile of "narrative indicators" harvested from X, Farcaster, and Lens. The platform would then route these signals through a series of language models—initially GPT-4, then increasingly fine-tuned open-source variants—producing a structured nine-dimensional analysis: technical, tokenomic, market, ecosystem, regulatory, governance, risk, narrative, and transmission.
The format was standardized. So were the prices. By mid-2024, the typical institutional subscription cost between $35,000 and $80,000 per year, and the pitch decks all contained the same chart: a hockey stick projecting that the global market for AI-augmented crypto intelligence would reach $4.2 billion by 2027. That chart was almost certainly wrong, but it did not need to be right to raise the round. What mattered was that the buyers believed it.
The buyers, importantly, were not retail. They were the new class of institutional crypto participants: prime brokerage research desks, family offices building digital asset allocations, the small but growing cohort of corporate treasuries that had added Bitcoin or stablecoin yield products to their balance sheets, and the funds that managed the rest of the funds—multi-manager platforms, FoFs, and outsourced-CIO shops that needed to produce research for their LPs at scale. What the buyers wanted, fundamentally, was throughput. They did not want one brilliant analyst who had spent three weeks understanding the economic model of a single mid-cap protocol. They wanted forty analysts who could each produce a one-page brief on a different protocol by Friday. The AI research factory promised to deliver that throughput. And for a brief window in 2024, it largely did.
Then the inputs began to rot.
To understand why an institutional research report came back ninety-three percent empty, you have to understand the architecture that produced it. The pipeline that generated the document the Singapore fund manager opened is, in its broad outlines, representative of the entire category. It runs in roughly seven stages, and each stage introduces its own failure mode.
Stage One: Source Identification. The pipeline begins by enumerating what it considers authoritative sources for a given protocol. These include the project's own documentation, its GitHub, its token contract on a block explorer, its CoinGecko or CoinMarketCap listing, its exchange listings, its governance forum, and a curated set of "contextual" sources ranging from news outlets to relevant Discord channels. This stage is brittle in ways that are not obvious from the outside. For established protocols with clean documentation, this stage produces a comprehensive source list. For newer or non-English-dominant protocols—particularly those emerging from Korea, Vietnam, Brazil, or the Turkish crypto scene—source identification is heavily biased toward English-language content and aggregator sites. The pipeline, in other words, is structurally blind to roughly forty percent of the global crypto ecosystem at the first step.
Stage Two: Ingestion. Once sources are identified, the pipeline attempts to ingest them. For API-based sources (block explorers, exchange data, on-chain RPC), ingestion is mechanical and largely reliable. For unstructured sources (documentation, Discord, governance forums), ingestion relies on web scraping or API endpoints that frequently break, rate-limit, or require authentication. Discord scraping in particular has become a chronic failure point: after the platform's 2024 API restrictions, the cost of reliable Discord ingestion rose by roughly an order of magnitude, and the quality of ingested data fell correspondingly. Telegram ingestion is worse; private channels are largely opaque, and even public channels suffer from inconsistent message indexing. The "narrative indicators" harvested from social platforms are particularly vulnerable to ingestion failures because the platforms actively work to prevent automated collection.
Stage Three: Parsing and Structuring. Raw ingested content is then passed through a parser that attempts to convert unstructured text into structured fields: token supply schedule, team wallet addresses, vesting cliffs, total raised, lead investors, etc. This is where the largest single category of failures occurs. Parsers are trained on a relatively narrow corpus of well-structured English-language crypto disclosures—the kind produced by US/EU-compliant projects with sophisticated legal counsel. When confronted with the actual messiness of a real protocol's disclosures—multilingual announcements, ambiguous vesting terms, supply schedules split across multiple contracts, team allocations routed through DAOs rather than direct wallets—the system degrades into uncertainty. The parser knows it doesn't know. But because the downstream pipeline requires a structured input, the uncertainty must be converted into a structured output. The default behavior is to mark fields as "not identified" or "insufficient data," which propagates downstream as an empty field.
Stage Four: Interpretation. At this stage, the parsed and structured data is fed to a language model for interpretive synthesis: what does the token supply schedule imply for sell pressure? How does the team's vesting cliff compare to category norms? What is the regulatory posture implied by the project's jurisdictional footprint? This stage is where the most visible problem—hallucination—historically occurs. But it is also where the empty-data problem becomes most insidious. If the structured inputs to this stage are mostly empty, the language model faces a choice: it can either produce an empty output (and risk being marked as a failure by whatever quality-control system wraps it), or it can hallucinate plausible-sounding content to fill the gaps. A well-designed system does the former. Most production systems, under commercial pressure, do the latter.
Stage Five: Cross-Validation. The pipeline then attempts to cross-validate the interpreted output against external sources. For example, if the interpretive stage concluded that the team wallet held 18% of supply, the cross-validation stage would check this against on-chain data. This stage is where the highest-quality pipelines distinguish themselves. But cross-validation is computationally expensive, and many production systems limit it to a small subset of "high-confidence" claims, leaving the long tail of interpretive output unchecked. The result is that the most falsifiable claims get validated, while the broadest interpretive claims—which are also the most likely to be wrong—pass through unchallenged.
Stage Six: Synthesis and Formatting. The validated output is then formatted into a final report—typically following the nine-dimensional template I described earlier. This stage is largely deterministic. It does not introduce failure modes; it merely renders whatever upstream stages produced. But it does introduce a subtle form of decay: the formatting layer normalizes the output. A report that contains eight empty fields and one well-supported insight is formatted identically to a report that contains nine well-supported insights. The visual distinction between substantive analysis and empty shell is erased at the moment of formatting.
Stage Seven: Quality Scoring and Delivery. A final quality score is calculated, often using a separate model trained to predict whether a human reviewer would consider the report "actionable." This stage is the last line of defense against the empty report. A well-implemented scoring system would flag the Singapore report as low-quality and either withhold it or attach a prominent warning. A poorly implemented one would let it through with a moderate score, on the assumption that some analysis is better than none.
When this pipeline was first designed in 2022 and early 2023, the assumption was that each stage would catch the failures of the previous one. Source identification would be biased toward completeness; ingestion would catch source gaps; parsing would catch ingestion failures; interpretation would catch parsing errors; cross-validation would catch hallucinations; quality scoring would catch everything else. In practice, this redundancy has failed. Each stage has its own failure modes, and the failure modes compound rather than cancel.
The Singapore fund manager's report failed at Stage One (source identification was incomplete for the specific protocol, which had heavy Korean-language documentation), which propagated through Stage Two (Korean sources were not ingested), which degraded Stage Three (the parser could not extract structured data from the limited English-language secondary sources), which produced empty inputs to Stage Four, which—because of the commercial pressure I noted—generated a partially hallucinated interpretive layer, which Stage Five attempted to validate but found insufficient data to validate against, and which Stage Seven scored as "moderate confidence" because the document's surface grammar was professional and its tables were well-formatted.
The result was a document that looked like research but contained, by my own subsequent audit, less than 7% verifiable factual content.
What interests me most about this failure mode is not its technical specifics—any competent systems architect could diagnose the pipeline and recommend fixes—but what it reveals about the information culture of institutional crypto in 2026.
We have built, over the past five years, an industry that has decoupled the surface form of research from its underlying informational substrate. The form—bolded insights, multi-dimensional analysis, professional formatting, executive summaries—has become standardized to the point of interchangeability. Two different AI research platforms, analyzing two different protocols, will produce documents that are visually nearly identical. The information density of those documents, however, can vary by two orders of magnitude. A reader cannot tell, from the document itself, whether they are looking at the output of a pipeline that successfully ingested forty sources and produced genuine insight, or a pipeline that ingested two sources and hallucinated the rest.
This is the empty signal pathology. It is not, strictly speaking, a new problem. Financial research has always suffered from a gap between the apparent rigor of its outputs and the actual rigor of its inputs. Sell-side equity research is famously produced under commercial pressure that incentivizes narrative over analysis. Credit rating agencies produced AAA ratings on mortgage-backed securities that were, in retrospect, empty signals of the most consequential kind. But the AI research factory has industrialized the production of empty signals in a way that the older analog systems could not.
The scale matters. A single equity analyst producing empty research affects a single firm. An AI research platform producing empty research affects every subscriber simultaneously. And because the platforms compete on throughput and price rather than on quality, the competitive pressure pushes every platform toward the same failure mode. There is no natural selection pressure inside the market that rewards information density. The natural selection pressure rewards report count, visual polish, and the appearance of comprehensiveness.
I want to spend some time on hallucination specifically, because it is the most discussed and least understood failure mode in the AI crypto research stack.
The standard framing of hallucination treats it as a bug: the model produces false information because it was trained on noisy data, or because its context window was insufficient, or because the prompt was ambiguous. The fix, in this framing, is better training data, longer context windows, and clearer prompts.
But this framing misses something important about the production environment. In a real research pipeline, hallucination is not a bug; it is a rational response to an underspecified input. The language model is being asked to produce a structured output. The structured output has slots. The slots need to be filled. If the information to fill a slot is not available, the model has two choices: leave the slot empty (and fail the quality check), or generate plausible content (and pass). The second choice is not irrational. From the model's narrow perspective, it is the optimal response.
The problem is that this optimal response, multiplied across millions of reports, produces a research stack that is, at the aggregate level, mostly fiction. And because the fiction is locally plausible—each individual claim sounds reasonable in isolation—the fiction is hard to detect. A reviewer reading a single report might not notice that 60% of the claims are unsourced. A reviewer reading fifty reports, however, would notice that the same claim appears across all fifty with the same phrasing. That is the signature of hallucination: not the implausible claim, but the suspiciously consistent one.
I have personally audited over two hundred AI-generated crypto research reports in the past eighteen months, in the course of building an internal scoring rubric for institutional due diligence. The single most reliable signal of hallucination is what I call "narrative lock-in"—the phenomenon where every report on a given protocol repeats the same five bullet points in the same order, regardless of the underlying data. When the same phrasing appears across reports from different platforms, produced by different model architectures, at different times, the probability that it represents independent observation approaches zero. It is, instead, the output of a shared underlying corpus—the same training data, the same source documents, the same interpretive priors—recombined in slightly different forms.
A second reliable signal is "category substitution"—the phenomenon where a report on a specific protocol includes analysis that could apply to any protocol in the same category. A report on a specific lending protocol that includes the sentence "the protocol competes in the DeFi lending category, which has grown significantly in recent years" is technically correct, but it is also substituting category-level narrative for protocol-specific analysis. When more than 30% of a report consists of category-level content that could appear in any sibling report, the report is, at the protocol-specific level, mostly empty.
A third signal is "founder mythologizing"—the phenomenon where the report's analysis of the team is dominated by biographical detail drawn from the founders' Twitter bios and conference appearances, rather than from evidence of operational capability. Founders who have spoken at three conferences appear in reports as "prominent voices in the space." Founders who have appeared on five podcast episodes appear as "thought leaders." These framings are not factually wrong, but they substitute visibility for substance.
None of these signals are individually conclusive. Together, they paint a portrait of a research stack that is, at the informational level, surprisingly thin.
To make this concrete, let me walk through three cases where I have directly observed the empty-signal pathology in operation.
Case One: The Korean DeFi Protocol. This is the Singapore report I opened with. The protocol had roughly $340M in TVL at the time of the report. It had a complex tokenomic structure with multiple vesting cliffs across team, investors, and ecosystem funds. Its primary documentation was in Korean. The pipeline produced a report in which the vesting schedule was marked "not identified" (correctly—it could not be parsed from the available English sources), the regulatory posture was marked "N/A" (correctly—jurisdictional analysis was impossible without the original Korean filings), and the technical analysis was a generic paragraph that could have applied to any EVM-compatible protocol. The team's wallet addresses were correctly identified. The investor list was correctly identified as including two named Korean VCs. Everything else was either empty or hallucinated.
The interesting detail was the report's treatment of the team. The pipeline identified three named founders, correctly extracted their Korean-language LinkedIn profiles, and then produced a paragraph of biographical analysis that drew on their conference appearances, their previous employment at Korean internet conglomerates, and their "thought leadership" in the Korean crypto community. None of this was wrong. None of it was useful. It was the category substitution pattern applied to the team dimension: the report produced visibility content where substance content should have been.
Case Two: The Privacy L1. A privacy-focused Layer 1 protocol that had raised $42M from a syndicate including several well-known crypto-native funds. The pipeline produced a report that confidently stated the protocol's regulatory posture was "favorable" because the team was incorporated in a jurisdiction known for crypto-friendly regulation. This was correct in a narrow sense: the legal entity was indeed in a favorable jurisdiction. But the pipeline failed to identify that the protocol's privacy features had already drawn enforcement attention from a major Western regulator, and that the team's mainnet launch had been delayed twice due to compliance concerns. The "favorable regulatory posture" claim was, in the context of the protocol's actual risk profile, dangerously misleading. It was not, however, factually wrong in the narrow claim-by-claim sense that the cross-validation stage would catch.
This case illustrates a more subtle pathology. The pipeline produced a report that contained, at the surface level, no hallucinations. Every claim was technically defensible. But the report's overall conclusion was wrong because the pipeline systematically omitted the most important information. The omission was not random; it was driven by source bias. The pipeline's source list for this protocol was dominated by the project's own marketing materials and a handful of crypto-friendly outlets. The regulatory enforcement news was available in major Western financial press, but the pipeline's source weighting deprioritized it. The result was a report that was technically accurate and substantively misleading.
Case Three: The RWA Tokenization Project. A real-world asset tokenization platform that had been the subject of a major partnership announcement with a tier-one traditional financial institution. The pipeline produced a report that treated this partnership announcement as a positive signal for tokenomics and ecosystem fit, while failing to identify that the partnership was non-exclusive, that the traditional institution had similar partnerships with four competing platforms, and that the underlying yield on the tokenized assets was subsidized by the issuer rather than market-derived. The report scored high on "narrative momentum" but low on what I consider the actually important dimensions: real revenue, real yield, real retention.
This case illustrates the narrative lock-in pattern most clearly. The pipeline's output on this protocol closely resembled outputs on five other tokenization platforms I have audited. The same five "key strengths" appeared in all six reports: institutional partnership narrative, real-world asset backing narrative, regulatory compliance narrative, yield generation narrative, and ecosystem growth narrative. The same three "key risks" appeared: regulatory uncertainty narrative, market concentration narrative, and technology maturity narrative. The reports were, at the bullet-point level, almost interchangeable. The protocol-specific substance was buried beneath the category-level boilerplate.
In all three cases, the surface form of the report was professional. The bolded insights were present. The formatting matched the standard template. But the underlying information density was, by my audit, in the range of 10-25%. These are not edge cases. They are representative outputs of the production research stack.
The most uncomfortable question is: why does this continue? Why, in 2026, with five years of AI research product iteration behind us, are empty signals still being delivered to paying institutional customers?
The answer is uncomfortable because it points to economic incentives, not technical limitations. The AI research market is structured in a way that rewards throughput over depth, narrative coherence over factual density, and surface professionalism over informational integrity. The buyers, as I noted, want forty one-page briefs by Friday, not one brilliant three-week analysis. The sellers have built their products around that demand. The result is a market equilibrium in which empty signals are profitable.
Consider the alternative: a research platform that actually audits its outputs, marks low-confidence reports as low-confidence, and refuses to produce hallucinated content. Such a platform would produce fewer reports. Its reports would be more expensive per unit. Its conversion rate—trial to paid—would be lower, because the trial reports would be less impressive than competitors' trial reports. Its churn rate would be higher, because customers would notice that the platform produces fewer "insights" per month. In a competitive market, this platform would lose.
The platforms that win are the ones that produce the most impressive-looking reports at the lowest cost. The empty signal pathology is, in this sense, a market outcome rather than a technical failure. It is what the market rewards.
This is a structural problem, not a cyclical one. It will not be solved by better models. It will not be solved by longer context windows. It will not be solved by more sophisticated cross-validation stages. It will only be solved when the buying side of the market changes its evaluation criteria—when institutions start buying research based on information density rather than report count, and are willing to pay more per report for genuine analytical depth.
I see early signs of this shift. The most sophisticated institutional buyers—certain macro hedge funds, a few sovereign wealth funds exploring digital assets, the research desks of the largest prime brokers—have begun to demand auditable provenance for AI-generated research. They want to know which claims in a report are sourced from which underlying documents. They want to see the chain of inference from source to claim. They want, in effect, the equivalent of a footnotes section for every AI-generated insight. The vendors that can provide this transparency are starting to command premium pricing. The vendors that cannot are starting to lose their largest institutional customers.
The shift is not happening uniformly. The smaller funds, the family offices without sophisticated procurement teams, the corporate treasuries that treat research as a checkbox item—these buyers continue to purchase on throughput and surface polish. They are the customers who keep the empty-signal platforms profitable. But their share of the market is shrinking. The institutional research market is bifurcating into a high-end segment that buys information and a low-end segment that buys signal. The middle is hollowing out.
There is a final layer to this that I think the industry has not yet metabolized. The empty report is not, in itself, the problem. The empty report is a symptom of a deeper problem: the institutional crypto research market has been, for five years, pricing signal rather than information.
When you buy a research report, you are buying one of two things. You are buying information—a set of factual claims about the world that you can use to update your beliefs and inform your decisions. Or you are buying signal—a marker that allows you to coordinate with other market participants, signaling that you have done diligence on a particular position.
The AI research factory has, inadvertently, made this distinction vivid. The reports it produces are increasingly high in signal and low in information. They allow a portfolio manager to say, in front of an investment committee, "We have a research report on this protocol from [vendor]." That statement has signal value. The contents of the report may have almost no information value. But the signal is what matters for the meeting.
This is not, in itself, a new phenomenon. Sell-side research has always had a strong signal component. The new element is the decoupling: it is now possible to produce high-signal, near-zero-information output at industrial scale, at low cost, with surface polish that makes the absence of information hard to detect.
The institutional buyers who understand this are quietly shifting their procurement strategies. Some are building internal research teams that produce information-dense outputs at low signal value. Some are paying premium prices for vendors that have built genuine information density into their products. Some are simply demanding more transparency about the provenance of every claim in every report they receive.
The buyers who do not understand this are the ones still paying $47,000 a month for reports that contain ninety-three fields marked "N/A."
What I find genuinely fascinating about this moment is the sociological pattern it represents. Crypto as an industry has, since its inception, been unusually susceptible to the substitution of narrative for substance. The original Bitcoin whitepaper was, in some sense, a narrative document—a story about why decentralized money mattered more than a detailed technical specification. The ICO boom of 2017 was, in large part, a narrative economy. The DeFi summer of 2020 was a narrative cycle. The NFT explosion of 2021 was, almost entirely, a narrative phenomenon. The L1 wars of 2022 were a narrative contest.
Each of these cycles had a moment when the narrative outran the substance. In the 2017 ICO boom, it was the moment when it became clear that most "utility tokens" had no actual utility. In 2020, it was the moment when it became clear that most "yield farming" was unsustainable. In 2021, it was the moment when it became clear that most NFT projects had no underlying value beyond their own market. In 2022, it was the moment when it became clear that most "Ethereum killers" had no actual users.
The current cycle—the AI crypto research cycle—is following the same pattern. The narrative is that AI will democratize high-quality research. The substance is that AI is producing empty signals at industrial scale. We are in the moment, somewhere in early 2026, when the gap between narrative and substance is becoming visible to anyone who looks closely.
The moment a narrative gap becomes visible, it enters what I call narrative decay. The narrative does not collapse immediately. It loses coherence gradually. The early adopters of the narrative—the true believers, the venture-funded evangelists—continue to insist that the narrative is intact. The mainstream participants—fund managers, institutional allocators, corporate buyers—begin to quietly diversify away from the narrative. The late adopters—retail, smaller funds, less sophisticated institutions—continue to buy the narrative at peak decay, often at the worst possible moment.
In 2026, we are somewhere between the mainstream quiet diversification and the late-adopter peak. The smart money is moving away from generic AI research tools. The dumb money is still subscribing. The next twelve to eighteen months will be the period when the late-adopter peak crests, and the narrative of "AI will democratize research" will be replaced by the more sober narrative of "AI is a useful tool for analysts, but it does not replace them."
The historical analogy I find most useful here is the early-2000s boom in algorithmic trading. Between 2000 and 2008, a narrative cycle emerged in which quantitative strategies were presented as a replacement for human judgment. The narrative held that machines could see patterns humans could not, and that the human analyst was becoming obsolete. The narrative peaked around 2006-2007, when equity quant funds reported record returns and the smartest institutional allocators piled in. The substance, of course, was that the algorithms were harvesting a specific set of risk premia that had been compressed by the entry of capital into the strategy. When the financial crisis hit in 2008, the algorithms unwound in correlated ways, and the narrative collapsed.
What survived the collapse was a more sober view: algorithmic trading is a useful tool that can reduce costs, improve execution, and surface patterns. It does not replace human judgment about which strategies to run, what risks to take, or when to deploy capital. The institutional allocators who survived the cycle were the ones who had understood this earlier. The ones who had bought the original narrative at peak coherence paid the price.
The current AI crypto research cycle is following the same arc, compressed into a shorter timeframe. The narrative will peak, the substance will become visible, the smart money will exit early, and the late adopters will pay the price. The only uncertainty is how much damage will be done in the unwinding.
The contrarian reading of all this—and I want to be explicit that I am about to advance a position that most of my industry colleagues will disagree with—is that the empty report is, in a perverse sense, the most honest output the AI research factory has ever produced.
Consider what an empty report actually does. It takes a request for analysis on a real protocol, attempts to ingest the available information, fails because the available information is incomplete or inaccessible, and returns a document that says, in effect: "We could not find the information necessary to produce this analysis." That is, in a narrow but important sense, the truthful output. The information was genuinely insufficient. The analysis could not, in good conscience, have been produced.
The hallucinated report—the alternative output—is the dishonest one. It produces confident-sounding claims that are not supported by the available evidence. It fills the structural slots with plausible-sounding content. It mimics the surface grammar of research while being, at the informational level, mostly empty.
If you wanted to design a system that surfaced the genuine limits of available information in the crypto ecosystem, you would design a system that produced empty reports when the data was insufficient. The Singapore fund manager's report, in this reading, is the system working correctly. The hallucinated reports that preceded it were the system failing.
The institutional research market has, for five years, demanded that AI research tools produce content even when the underlying information is missing. The empty report is the first sign that the tools are beginning to resist that demand. This is, I would argue, a positive development.
The second-order implication is more uncomfortable. If we accept that most AI-generated crypto research is, at the informational level, mostly empty, then we have to ask what the industry has actually been buying for the past five years. The answer is: signal, not information. The industry has been buying the ability to claim that it has done research. It has not been buying actual research.
This is not a new phenomenon, as I noted earlier. But the AI research factory has made it unusually transparent. The older analog version of this problem—sell-side equity research with weak informational content but strong signal value—was difficult to identify because it required reading the research carefully to spot the weakness. The AI version is identifiable by simple metrics: count the empty fields, count the unsourced claims, count the paragraphs that could apply to any protocol in the category. When you do this for a representative sample of AI-generated crypto research reports, you find that the information density is, on average, below 30%.
The contrarian position, then, is this: the empty report is the most honest output the AI research factory has ever produced. It is the first sign that the system is beginning to admit its own limitations. The market should welcome this development, even though it means acknowledging that the industry has, for five years, been buying signal at the price of information.
There is, I should note, a third possible reading that I find even more interesting than either the standard alarmist reading or my contrarian one. The third reading is that the empty report is a piece of accidental performance art. It is a document that purports to analyze a crypto protocol but contains no analysis, dressed in the formal language of analysis. It is, in a precise sense, a critique of the very genre it belongs to. The empty report is crypto research's version of the emperor's new clothes, except that in this case the emperor is not naked—he is wearing a beautifully tailored suit that happens to be made of nothing.
I do not expect the industry to adopt any of these readings. The industry will, instead, demand that the empty reports be filled with more content, that the platforms be made to produce more insights, that the AI be made to "work harder." This demand will produce more hallucinated output, which will produce more downstream decisions made on false information, which will produce more market losses, which will eventually produce a reckoning. The reckoning is, I believe, twelve to eighteen months away. The question is whether the institutional buyers will read the warnings correctly in time, or whether they will read them only after the damage is done.
What happens next will be determined less by technical progress in AI models than by a shift in institutional procurement criteria. The funds that survive the current empty-signal cycle will be the ones that buy information density over report count, that demand auditable provenance for every claim, and that are willing to pay premium prices for genuine analytical depth rather than surface polish. The vendors that survive will be the ones that build transparency into their products at the architectural level, not as a wrapper around the same hallucinated outputs. The analysts who survive will be the ones who learn to use AI as a tool for genuine inquiry rather than as a substitute for it.
The empty report is not a failure. It is, in a precise sense, a correction. The question is whether the industry will read the correction correctly, or whether it will simply demand that the empty reports be filled with more confident-sounding hallucinations.
I know which way I am betting. And it is not the way the smart money is betting. The smart money is betting that the AI research factory will continue to scale, that institutional procurement will continue to prioritize throughput, and that the gap between narrative and substance will continue to be profitable. I am betting that the gap will become too visible to ignore, that a high-profile institutional loss will focus attention on information density, and that the market will bifurcate faster than most observers expect.
The next twelve months will be the period in which this bet is resolved. If I am right, the institutional crypto research market of 2027 will look very different from the one we have today: smaller in revenue, higher in unit price, more concentrated among vendors who have built genuine transparency into their products, and more skeptical of AI-generated output at face value. If the smart money is right, the empty report will be quietly filled in, the market will continue to grow on the same trajectory it has been on since 2023, and the institutional buyers will continue to receive what they have been receiving all along—signal dressed up as information.
I have been in this industry long enough to know that the smart money is usually right about the near term and often wrong about the medium term. The near-term trajectory of the AI crypto research market is, I think, more growth, more vendor consolidation, more institutional adoption, more empty reports quietly filled with hallucinated content. The medium-term trajectory, I think, is a sharp correction, a high-profile loss that focuses attention on information density, and a market that rewires itself around the distinction between signal and information.
The empty report sitting in the inbox of that Singapore portfolio manager is, in this sense, a leading indicator. It is a small piece of evidence in a much larger pattern. The pattern is not yet visible to most observers. It will become visible. And when it does, the industry will have to decide whether to acknowledge the gap between narrative and substance, or whether to keep buying the narrative at the price of substance.
Based on my own audit experience across more than two hundred reports in eighteen months, I am increasingly confident in which way the correction will go. The question is no longer whether the gap exists. The question is how long the market will continue to ignore it. And that, ultimately, is a question about institutional incentives, not about AI capability.