When a bridge collapses, everyone knows. The feedback is immediate, public, and fatal. Engineers build within disciplines where reality pushes back. A rocket explodes on the pad. A building fails inspection. The concrete cracks.
Deep research has none of this.
Five companies built nearly identical products in eight months. OpenAI, Google, Perplexity, xAI, Mistral. All five perform the same operation: parse a question, browse the web, synthesize what they find, present a polished report. The price floor dropped from $200/month to free. The reports look authoritative. The errors are invisible. And the user, lacking the domain expertise to verify what they’re reading, accepts the output and moves on.
Deep research tools have no feedback mechanism. When the report is wrong, nothing falls down. Nobody checks. The errors accumulate silently, and the market selects for the features that make this problem worse.
Five companies, one product
The convergence happened fast. Google launched Gemini Deep Research in December 2024. OpenAI followed in February 2025 at $200/month. Perplexity launched its own Deep Research days later. Grok offered real-time X/Twitter data for free. Mistral launched Le Chat with on-premise deployment and EU regulatory alignment.
Strip away the branding and they all do the same thing. Browse, retrieve, synthesize, format. The differences are retrieval variations. Gemini pulls 100 to 450+ sources. Perplexity prioritizes freshness. Grok has social media data. These are not different products; they are different configurations of the same pipeline.
What none of them ship is more telling. As of early 2026, no product offers vocabulary reframing, the ability to help users find terms they don’t know to search for. No product offers verification against the user’s existing knowledge. No product offers navigation, the capacity to notice when a search is heading in the wrong direction and pivot.
The DEFT benchmark evaluated roughly 1,000 reports across 14 failure modes. The best performer, Gemini, scored 51 out of 100. The price collapse tells the same story: the core operation has no pricing power because it has no defensible value. On GAIA, a separate benchmark for AI assistants, HuggingFace scored 55% within a day of OpenAI’s launch, against OpenAI’s 67%. A 30B parameter model ran on consumer GPUs by November 2025. The commoditization was complete.
Everyone built the same thing. The question is why.
The tradeoffs they chose
Every gap in these products traces to a rational design decision, not incompetence. The tradeoffs become clearer when you separate infrastructure choices from strategic ones.
Infrastructure: several products use Bing’s API rather than Google Search. Cheaper to integrate, smaller index. They browse with headless browsers that hit CAPTCHAs and paywalls, seeing a degraded version of the internet. Most don’t search Hacker News or GitHub at all, making entire expert communities invisible.
Then there’s the single context window. A report that pulls hundreds of sources crams them into one context. At that scale, synthesis degrades. The model can’t distinguish signal from noise across that many documents any better than a human skimming the same number of tabs. Worse, if those sources come from the wrong part of the information landscape, every one confirms the wrong answer.
Strategic choices are even more revealing. Batch delivery instead of conversation. Simpler to build, but it locks the human out of the research process. Plans must change mid-execution, and they can’t. Format polish instead of uncertainty markers. Polished reports win demos. Uncertainty markers lose them. Retrieval metrics instead of navigation quality. You can measure how many sources the tool found. You cannot easily measure whether it found the right ones.
Each decision is rational for a company optimizing for adoption. Each has a cost the user absorbs without knowing it.
A few people checked.
The floor went soft
Derek Lowe, a medicinal chemist with decades of drug-discovery experience, tested deep research on thalidomide toxicity. He found three failure modes. Completeness bias: the tool presented a comprehensive-looking answer that omitted critical pharmacological details. Semantic confusion: it conflated related but distinct biochemical mechanisms. Scope-of-question rigidity: it omitted that thalidomide research spawned an entirely new field — targeted protein degradation — because that wasn’t explicitly asked about.
Lowe’s summary is precise: “you have to know the material already to realize when your foot has gone through what was earlier solid flooring.”
That sentence contains the verification paradox. The domain expertise needed to verify a deep research report is the same expertise that makes the report unnecessary. The tool is most dangerous exactly when it seems most useful: on topics where the user lacks the knowledge to spot errors.
Lowe was not alone. A JMIR study from UCL and Moorfields found hallucinated citations, open-access bias, and plausible-sounding fabrications in medical literature reviews. A post on the American Journal of Nursing blog found deep research tools searched only one or two scholarly databases. The DEFT benchmark measured strategic content fabrication at roughly 19% of all annotated failure instances, the single most common failure mode.
The errors don’t distribute randomly. They cluster at navigational boundaries. Within a broad topic, retrieval is competent. The tool finds papers, extracts relevant passages, assembles them coherently. But where the research requires judgment about what to look for next, where the vocabulary shifts, where one subdomain borders another, the tool fails. Pharmacology becomes toxicology. Legal precedent becomes fabricated citation. Accounting frameworks get confused.
That fabrication rate is format-amplified. Three forces compound. RLHF training rewards confident, complete-sounding outputs. Users expect polished reports. And the report format itself makes gaps visible, turning every “I don’t know” into a hole that looks like failure. Format doesn’t create the fabrication incentive. It amplifies it by making honesty look like incompetence.
This produces what I’ll call the Conclusive Illusion. A report that looks detailed invites checking. A report that looks done discourages it. “This seems comprehensive” prompts the reader to verify a claim or two. “This is the complete answer” prompts the reader to close the tab. The distinction matters because deep research reports are designed to look done. Every formatting choice, every confident assertion, every clean section header signals completion. The epistemic loop closes on questions that remain open.
The hard problem isn’t finding things
Every deep research product optimizes retrieval. Better search, more sources, faster browsing. But the hard problem in research isn’t finding things. It’s knowing what to look for next.
This is the distinction between retrieval and navigation. Retrieval is the operation of fetching relevant documents given a query. Navigation is the process of formulating the right query in the first place, recognizing when the results suggest you’re looking in the wrong place, and adjusting.
A study by Furnas and colleagues in 1987 found that when people name the same database concept or object, they agree on the term less than 20% of the time. The vocabulary problem is not an edge case; it’s the default condition. If you’re searching for “slow file manager” and the actual cause is an xdg-desktop-portal D-Bus timeout, no amount of retrieval quality bridges the gap. You need a human, or a system, that can jump between vocabularies.
Batch architecture makes this impossible in current products. Deep research tools take a question, generate a search plan, execute it, deliver results. The human is locked out during execution. But research plans need to change mid-execution. You read three papers and realize the question was wrong. You find a term you didn’t know existed and need to pivot. You hit a dead end and need to try a different approach.
Simon Willison’s trajectory illustrates the progression. In April 2025, he wrote that AI search had “finally crossed the line into being genuinely useful.” By September 2025, he had coined the term “Research Goblin” for GPT-5’s agentic search, which iterates internally — searching, reasoning about results, searching again — before delivering a final answer. A step beyond batch, but the human still only reviews afterward. Post-hoc review catches errors; it can’t change direction. OpenAI added plan editing in February 2026. A half-measure. You can see the plan, but you still can’t redirect mid-execution.
Three principles compress the structural problem:
Research is ignorance management. The researcher’s job is not to know things. It’s to manage what they don’t know, to identify which unknowns matter, which can be deferred, which are hiding. Current tools can’t model their own ignorance.
The hard problem is navigation, not retrieval. Retrieval has been commoditized ($200 to free in months). Navigation, the capacity to recognize you’re in the wrong part of the information landscape and reorient, has not been addressed by any product at scale.
Format inversely correlates with epistemic safety. The more polished and complete a report looks, the less likely the reader is to question it. Every dollar spent on format quality makes the output epistemically worse.
They predict which problems will get worse. The answer depends on what kind of question you’re asking.
The river problem
Some questions don’t need intelligence. “What is TCP?” Type it into any search engine, any deep research tool, any LLM. You’ll get a correct answer. The information landscape for this question is convex: every path through it converges on the same destination. Fall into the river, and the current carries you.
Stack Overflow, MDN, Wikipedia. These are paved rivers. The information is canonical, well-indexed, internally consistent. Deep research tools handle river problems well. Their demos use river problems. The impressive-looking 30-page reports with 200 citations are, almost always, river problems.
Non-river problems are different. “My file manager is slow” seems simple. But the answer might be an xdg-desktop-portal D-Bus timeout that has nothing to do with file managers. The information landscape has multiple basins. The vocabulary barrier between “slow file manager” and “D-Bus timeout” is a cliff, not a slope. A tool that searches with the user’s vocabulary stays in the wrong basin. Hundreds of sources from the wrong basin produce hundreds of confirmations of the wrong answer.
The Furnas finding makes the vocabulary problem concrete: in controlled conditions, the chance of two people independently choosing the same term for the same concept was below 20%. Deep research tools take one shot at the vocabulary, then execute. If that shot lands in the wrong basin, the entire report confirms the wrong answer with high confidence and professional formatting.
Taking one shot at the vocabulary and calling the result comprehensive is not methodology.
The distinction matters because the tools are being evaluated on river problems and deployed on non-river ones. A demo that correctly summarizes the TCP handshake protocol proves nothing about the tool’s ability to navigate a pharmacological question where the user doesn’t know the right terminology. The impressive accuracy on well-indexed questions masks the structural failure on the questions where accuracy matters most.
The immediate consequence is errors. But over 100 sessions, over a year of mostly-right answers, the consequences are personal.
The skills you lose
Accuracy kills the checking instinct. When a tool is right 90% of the time, you stop verifying. This isn’t laziness. It’s rational calibration to observed reliability. The problem is that the 10% failure rate stays constant while your detection rate drops.
Ethan Mollick has written about the underlying mechanism: AI doesn’t damage your brain, but unthinking use damages your thinking habits. The temptation to offload cognitive work to the tool means users skip the mental effort that builds expertise.
Four capabilities degrade, roughly in sequence.
Query formulation atrophies first. The tool rewrites your question, expands it, generates sub-queries. You stop learning how to formulate good searches because the tool does it for you. This seems like a feature. It’s a dependency.
Source credibility instinct goes next. The tool treats a peer-reviewed paper and a blog post as equivalent inputs. Over time, many users do too.
Verification habit follows. When the first 50 reports are right, checking the 51st feels like wasted effort. You stop.
Meta-calibration is the most dangerous loss. This is the instinct to know when to be careful. Not just whether a specific claim is true, but whether this is the kind of question where claims are likely to be wrong. It’s the difference between reading a Wikipedia article on photosynthesis (low stakes, high accuracy) and reading a deep research report on drug interactions (high stakes, unknown accuracy). Experienced researchers have this instinct. It atrophies when a tool presents both with identical confidence and formatting.
The error rate stays constant. The detection rate drops. The gap between the two is where damage accumulates.
Users already know
They’re running five tools in parallel. A Hacker News user identified as iandanforth described the workflow: run the same question through five AI tools, throw out 65 to 75% of the output, cross-reference the remainder. Multiple independent users converge on the same pattern. Gemini or Perplexity for breadth. ChatGPT or Claude for reasoning. Cross-check between them. Apply domain expertise to reconcile.
Nobody runs five browsers for a Google search.
The deeper pattern is that users are independently inventing conversational research. They are the navigator. The tools are retrievers. The human provides the vocabulary jumps, the judgment calls, the “wait, that doesn’t sound right” interventions. The product that wins will make this improvised workflow native rather than forcing users to jury-rig it across five tabs.
The open-source evidence reinforces the point. A Nature paper from February 2026 found that open-source tools outperformed commercial products on some literature review metrics. If the commercial products had solved the hard problem, independent builders would stop building. They haven’t stopped.
What users are compensating for is clear. What companies are investing in instead makes the problem worse.
Where the money goes
Investment flows toward format quality. Better-looking reports, more professional citations, cleaner section headers, longer documents. The better it looks, the less anyone questions it.
The competitive dynamics select for the most dangerous feature. In a demo, the polished report wins. The report with visible uncertainty and rough edges loses. Product teams optimize for demos because demos drive adoption. Adoption drives revenue. The feature that causes the most epistemic damage is the feature that wins the most customers.
The optimal deep research output would look worse. Rough formatting. Visible gaps. Uncertainty markers on every claim. Explicit statements about what the tool looked for and didn’t find. Confidence intervals instead of assertions. This product would lose every demo and win every serious research task.
No one will build it. The incentive gradient points the wrong direction.
In engineering, reality provides the feedback. The bridge falls. The rocket explodes. The test fails. These feedback mechanisms are brutal and expensive. They are also irreplaceable. They force convergence toward correctness because the cost of being wrong is visible and immediate.
Deep research has no equivalent. The report goes out. The user accepts it. Nobody checks. If the report is wrong about a drug interaction, the user doesn’t know. If it fabricates a citation, the user doesn’t look it up. If it misses the critical paper because it searched the wrong database, the absence is invisible. No bridge falls. No rocket explodes. The error enters the user’s understanding and stays there.
The information ecosystem is bifurcating along the same fault line. AI-optimized shallow content proliferates on the open web, designed to be retrieved and summarized. Expert knowledge retreats behind paywalls and institutional access, into specialized communities the tools can’t reach. The tools will get better at navigating worse content. The gap between what’s easy to find and what’s worth finding will widen.
The honest position
The steelman deserves full engagement. Evaluating deep research tools as substitutes for human research may be a category error. As complements, they do something genuinely useful: broad scanning of large information spaces at a lower accuracy bar. A researcher who needs to survey 500 papers to identify the 20 worth reading closely has a legitimate use case. The tool handles breadth. The human handles depth.
The PRISMA and Cochrane frameworks provide a relevant counter-example. Systematic reviews, the gold standard of medical evidence synthesis, use batch methodology deliberately. The protocol is defined in advance. The search is exhaustive. Human steering is explicitly minimized because steering introduces bias. Batch research works for exhaustive coverage of known question spaces.
The reconciliation: batch for known questions, conversational for exploratory ones. When you know exactly what you’re looking for and need comprehensive coverage, the batch model is appropriate. When you’re investigating something you don’t fully understand, when the question might be wrong, when the vocabulary might shift, batch architecture is structurally inadequate.
Why does research work at all? Three conditions. Smoothness: nearby queries give nearby results, so iterative refinement works. You can get closer by adjusting incrementally. Approximation: rough models capture enough structure. Herbert Simon argued in 1962 that nearly decomposable systems can be understood through simplified models. Convergence: independent paths producing the same answer is the strongest signal of truth. When three different approaches to a question produce the same answer, that answer is likely correct.
Current deep research products violate all three conditions. They don’t iterate (batch delivery). They don’t preserve rough approximations (format pressure demands polish). They don’t leverage convergence (single-tool, single-pass architecture).
Three predictions, each with a criterion that would prove it wrong:
Scale will produce negative returns. More sources and bigger models will not fix navigation failures. They may make them worse by increasing confidence in wrong-basin answers. This prediction is wrong if brute-force scaling pushes error rates below 2% on non-river problems.
The next breakthrough will be an interaction paradigm, not a model improvement. The product that lets users steer mid-research and jump vocabularies will outperform the product with the best retrieval. This prediction is wrong if a larger model achieves conversational-quality research in a batch architecture.
The category will stratify into expert tools and non-expert tools. Expert tools will emphasize navigation and human steering. Non-expert tools will emphasize polish and conclusive-sounding output. These will be different products for different markets. This prediction is wrong if an adaptive interface dissolves the tension between the two.
The honest recommendation is proportional trust. Use deep research for river problems and broad surveys. Verify everything that matters. Run multiple tools when stakes are high. Treat polished formatting as a warning sign, not a quality signal. And calibrate your trust to the stakes: the cost of a wrong answer about TCP is a retry, the cost of a wrong answer about a drug interaction is not.
The bridge hasn’t fallen because there is no bridge. That has not made you more careful.