# Where genuinely new findings come from, and which Canopticon questions are best placed to produce one
Research document · Canopticon · published 2026-10-19
Canonical: https://canopticon.com/p/where-genuinely-new-findings-come-from-and-which-canopticon-questions-are-best-placed-to-produce-one
License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
---
# Where genuinely new findings come from, and which Canopticon questions are best placed to produce one

## TL;DR

The research that most reliably produces new knowledge out of existing material does one of three things. It joins pieces of evidence that already exist but sit in literatures that never cite each other (Swanson's fish oil and Raynaud's link; the streptokinase effect, which pooled trials showed in 1973, years before the large trials). It builds a new instrument or measure rather than a new theory (new methods are the type of novelty most tied to disruptive influence). Or it combines a conventional base with one unusual join (papers built that way were twice as likely to become hits across 17.9 million papers) [1][3][10][11]. Each route has a known failure mode. Novelty is often recognised late and is risky. Literature-based discovery has almost no record of validated hits beyond its founder. The headline claims for AI-driven discovery rest on single cases, and one prominent study's data were disowned by its own university [2][11][17][18]. Applied to Canopticon, the two questions best placed to yield a genuinely new result are *self-extinguishing maintenance* and the *self-service transfer*. The first is a cross-domain loop whose parts are each documented in separate fields that have never been joined. The second is a named but unmeasured shift of clerical work onto individuals. Both can be checked against data that already exist. The McLuhan *precision ratio* programme is original, but its next gain comes from a test, not more synthesis.

## Key Findings

- **Highest-impact papers pair a conventional base with a small intrusion of unusual combinations.** Across 17.9 million papers, those with this profile were twice as likely to be highly cited, and teams were 37.7% more likely than solo authors to insert such combinations [1].
- **Novel work pays late and unevenly.** Papers making distant new combinations are more likely to be top-1% cited in the long run and are cited more by distant fields. Their citation performance also varies more, and recognition comes late. Short-window metrics therefore penalise them [2].
- **New methods are the type of novelty most associated with disruptive influence.** This comes from text analysis of authors' own essays on Citation Classics [3].
- **Pooled existing evidence can settle a question years before the field acts.** Of 33 streptokinase trials run from 1959 to 1988, the first eight (2,432 patients) already showed a significant mortality reduction by 1973 (odds ratio 0.74; 95% CI 0.59–0.92), before the 1986 GISSI trial of 11,712 patients [10].
- **Literature-based discovery works in principle but is barely validated.** Systems are mostly judged by whether they can re-find Swanson's own discoveries. The field's evaluators found that Swanson was the only researcher who managed to publish his discoveries in medical journals [11]. His 1986 fish-oil/Raynaud's hypothesis was followed by a small positive double-blind trial in 1989 [11][12].
- **Modelling which experts could plausibly see a link improves prediction of future discoveries by up to 400%.** The same models can be tuned to "avoid the crowd" and generate hypotheses no one is likely to reach soon [13].
- **Outsiders win more often.** Across 166 corporate science challenges and over 12,000 solvers, winning was positively related to the distance between the solver's field and the problem's field [14]. Small teams disrupt more than large ones across more than 65 million papers, patents and software products [4][5].
- **The measured decline in disruptive science is real but disputed in size.** An inventory of 105 studies reports a decline across papers, patents, products, legal cases, music and art [5]. Critics attribute part of it to records with zero references [6]. The original authors, who also wrote the inventory, rebut this [7].
- **AI-augmented scientists gain individually while science narrows.** In 41.3 million papers, AI-using scientists published 3.02 times more papers and received 4.84 times more citations. AI adoption shrank the volume of topics studied by 4.63% and cut scientists' engagement with each other by 22.00% [15].
- **AI-generated ideas look better on paper than in practice.** When 43 researchers executed LLM and human ideas, the LLM ideas' review scores fell much more than the human ideas', closing or reversing the gap seen at the idea stage [16].
- **The strongest AI-discovery case is one lab grading its own test.** Google's AI co-scientist proposed, as its top hypothesis, a mechanism that matched the lab's unpublished result, plus a conjugation idea the lab is now testing [17]. The best-known study claiming large AI discovery gains (44% more materials) was disowned by MIT, which said it had no confidence in its data [18][19].

## Details

### 1. What "novel" means when it is measured, and why it is hard to see in time

Three measurement traditions now exist, and they agree more than they disagree.

The first measures novelty as **unusual combination**. Uzzi, Mukherjee, Stringer and Jones (Science, 2013) scored 17.9 million Web of Science papers by how typical each pair of cited journals was, relative to chance [1]. The highest-impact papers were grounded in highly conventional combinations of prior work and also contained an intrusion of unusual combinations. Papers with that profile were twice as likely to be highly cited. Novel combinations were rare, and teams were 37.7% more likely than solo authors to add them to familiar domains [1]. The practical reading is that a hit is a familiar argument carrying one surprising joint, not a wholly strange argument.

The second asks **what novelty costs**. Wang, Veugelers and Stephan define novel papers as those making new, and especially distant, combinations of referenced journals [2]. Such papers do better on most long-run measures. They are more likely to be among the top 1% most cited, more likely to inspire highly cited follow-on work, and more often cited by distant fields. But their citation performance varies more, and recognition is delayed [2]. A later summary by Franzoni and Stephan of the same evidence puts the timing plainly. For the first three years after publication, a highly novel paper is less likely than a non-novel one to be in the top 1%. After that it pulls ahead [2]. Any judge working on a short horizon (a peer reviewer, a grant panel, an editor, or a reader scoring an AI idea on first sight) will therefore tend to undervalue the novel option. The authors subtitled the paper *A Cautionary Tale for Users of Bibliometric Indicators* [2].

The third measures novelty by its **effect on what came before**. The CD index scores a work as disruptive when later work cites it without citing its references, so that it eclipses its predecessors. It scores a work as consolidating when later work cites both. Leahey, Lee and Funk (American Sociological Review, 2023) linked this index to essays in which authors of Citation Classics described what was new about their own papers [3]. New methods tended to be more disruptive than other types of novelty [3]. That fits an older observation in the sociology of science: instruments and techniques travel across fields more easily than results or theories.

These measures sit inside a larger argument about whether new ideas are getting harder to find. Bloom, Jones, Van Reenen and Webb find research effort rising sharply while research productivity falls. Their best-known example is Moore's Law: sustaining the doubling of chip density takes a research workforce "more than 18 times larger than the number required in the early 1970s" [8]. Jones's burden-of-knowledge model explains part of this. As knowledge accumulates, innovators specialise more narrowly and depend more on teams, and he finds evidence consistent with each of the model's six predictions in micro-data on inventors [9].

**The disruption-decline dispute.** Park, Leahey and Funk reported in Nature in 2023 that papers and patents have become less disruptive since the Second World War, across roughly 45 million papers and 3.9 million patents [5]. The debate since then is the clearest live disagreement in this evidence base, and it bears on how much weight any decline should carry.

- *For the decline.* An inventory of 105 empirical studies, assembled from 1,587 works citing four seed papers, finds the large majority report decline. This holds across citation measures, patent text, firm displacement rates, product-similarity networks, audio features of Brazilian forró music, and image features of Japanese woodblock prints [5]. The authors argue that non-citation measures cannot share citation artefacts. They also argue that structured exceptions, such as rising disruption in fetal surgery and in AI after 2014, would not arise from a mechanical artefact [5]. Their summary: "Across measures and domains, that evidence points consistently toward decline" [5].
- *Against its size.* Holst and colleagues (Nature, Matters Arising, August 2026) attribute part of the decline to database entries with zero references, which mechanically score the maximum CD of 1 and have become rarer as metadata improved [6]. In one data source, dropping those entries cut the measured decline by 97% [6]. Petersen and colleagues find trends "at the level of noise" after controls in a 1995–2015 sample, and attribute the decline to citation inflation [5].
- *The reply.* Park, Leahey and Funk respond that the critics' own regression shows large, significant declines (P < 0.01). They also say the critics' dataset contains three times as many zero-reference works as theirs, inflated by at least 2.8 million editorials, obituaries and comments, 1.5 million books and proceedings, and 254,000 reviews. In the critics' sample, non-research content fell from 40% to 8% between 1945 and 2010, which is enough on its own to produce the zero-reference trend [7].

**Which account is more credible.** The direction of the decline is better supported than its size. Its direction is backed by measures that use no citations at all. Its size depends on cleaning choices the two sides have not settled. Note that the 105-study inventory was written by the original authors, so it is an interested synthesis, although its entries can be checked. One finding in it matters more for Canopticon than the trend line. Li and colleagues report that the share of papers with a positive disruption score fell from nearly 50% in 1965 to under 20% by 2020, while the absolute number of highly disruptive papers stayed roughly stable [5]. Disruptive work has not disappeared. It is being diluted by a growing volume of consolidating work.

### 2. Does synthesis of existing literature actually produce validated discoveries?

Yes, under narrow conditions, and the evidence is stronger for pooling than for linking.

**Pooling: the cumulative meta-analysis record.** Lau, Antman, Jimenez-Silva, Mosteller and Chalmers (NEJM, 1992) re-ran the trials of 15 treatments for heart attack in chronological order. They pooled each new trial with all those before it to see when the answer became clear [10]. For intravenous streptokinase, 33 trials were run between 1959 and 1988. A consistent, statistically significant reduction in total mortality (odds ratio 0.74; 95% CI 0.59–0.92) had already appeared by 1973, after only eight trials with 2,432 patients [10]. The two very large trials that followed, GISSI in 1986 (11,712 patients) and ISIS-2, added precision but not direction [10]. This is the cleanest demonstration that a "discovery" can sit unassembled in published data for over a decade. Nothing new was measured. Someone arranged existing results in an order that let them speak.

**Linking: Swanson's undiscovered public knowledge.** The record's *Net-Casting* document already covers Swanson's method: literature A links to B, B links to C, but A and C never cite each other. The new question here is how often such links hold up. Swanson's 1986 paper proposed that dietary fish oil might help Raynaud's syndrome, by joining a literature on blood viscosity and platelet behaviour in Raynaud's to a separate literature on the effects of fish oil [11]. In 1989, DiGiacomo, Kremer and Shah published a double-blind controlled trial in the American Journal of Medicine [12]. In the fish-oil group, the median time to onset of a cold-induced Raynaud's attack rose from 31.3 minutes at baseline to 46.5 minutes at six weeks (p = 0.04). Five of 11 patients with primary Raynaud's on fish oil could no longer be induced to develop an attack, compared with one of nine on placebo [12]. That counts as a validation, but a small one.

The base rate behind that example is the weak point. Yetisgen-Yildiz and Pratt's review of how literature-based discovery systems are evaluated names the core difficulty: if such systems succeed, "by definition they are capturing new knowledge that has yet to be proven useful" [11]. In practice, most systems were judged by whether they could re-find Swanson's own handful of discoveries (migraine–magnesium, Raynaud's–fish oil, Alzheimer's–estrogen, Alzheimer's–indomethacin, somatomedin C–arginine, schizophrenia–calcium-independent phospholipase A2) using literature from before the discovery date [11]. The review lists three weaknesses of that method:

- It says nothing about the hundreds or thousands of other links the system proposes.
- Developers who know the target can tune their systems to hit it.
- It cannot rank two systems that both succeed [11].

Time-sliced evaluation, which runs a system on literature before a cut-off and checks which links appear afterwards, does better. But it gives only a lower bound on precision, because links that appear later still count as misses [11]. The review's verdict on real-world validation is blunt: "Among all LBD researchers, Swanson is the only researcher who could publish his discoveries in the medical journals" [11]. It notes one laboratory follow-through. Wren and colleagues ranked chlorpromazine third among 19,718 candidate targets for cardiac hypertrophy, then showed in mice that it reduced isoproterenol-induced hypertrophy [11].

**What has changed since.** Sourati and Evans (Nature Human Behaviour, 2023) offer the strongest recent evidence that the linking idea scales when it accounts for people [13]. Models that simulate which inferences are cognitively accessible to experts, meaning which scientists could plausibly connect two ideas given what they have published, improved prediction of future discoveries by up to 400% over models built on research content alone. The gain was largest where the relevant literature was sparse [13]. The same models can be tuned "to avoid the crowd", generating hypotheses that are scientifically promising but unlikely to be pursued until the distant future [13]. In effect, this formalises Swanson's insight. The discoverable gaps are social, not only logical: they lie where no single expert's reading spans both sides.

**Inference (mine).** Pooling has a strong record because its output is checkable: one estimate, one confidence interval, one date. Linking has a weak record because its output is a ranked list of candidates with no built-in test. A synthesis meant to produce a validated finding should therefore end in something that can be checked against data that already exist or will soon exist, not in a list of plausible connections.

### 3. Which people, teams and moves produce the breakthroughs

The evidence here is correlational and mostly citation-based. Its consistency across very different designs is what gives it weight.

**Distance from the problem.** Jeppesen and Lakhani analysed 166 science challenges that R&D-intensive companies broadcast on InnoCentive, with over 12,000 external scientists self-selecting to solve them under blind review [14]. In their words, "The provision of a winning solution was positively related to increasing distance between the solver's field of technical expertise and the focal field of the problem" [14]. Women, whom the authors place in the outer circle of the scientific establishment, performed significantly better than men [14]. Their explanation is that marginal solvers bring different perspectives and heuristics. Self-selection means this is not a randomised result: outsiders who chose to enter may have seen a route insiders missed.

**Small teams and generalists.** Wu, Wang and Evans (Nature, 2019) found across more than 65 million papers, patents and software products that small teams disrupt and large teams develop. Small teams have also become less common [4][5]. Studies catalogued in the disruption inventory add further mechanisms:

- Generalist teams were 31% more likely to produce highly disruptive papers, while the share of generalist scientists fell from 55% to 40% (Risha et al.) [5].
- A one-standard-deviation increase in topic switching made a paper 39% more likely to be disruptive (Yang 2025) [5].
- Conceptual churn in a field predicted disruption (Kedrick et al.) [5].
- Knowledge independence, meaning reliance on references that are not themselves heavily interlinked, strongly predicted disruption (Yu et al.) [5].
- Remote collaboration fused fewer breakthrough ideas across 20 million papers and 4 million patents (Lin et al.) [5].
- In marketing, conceptual articles were more disruptive but were declining in number (Lee & Kim) [5].

**New instruments over new results.** As section 1 noted, new methods are the novelty type most tied to disruptive influence [3]. The record's *Telescope Problem* document reached the same conclusion for economics from case evidence: the field-reorganising contributions came from new instruments such as online price indexes and private-data trackers, not new theory. The bibliometric evidence and the case evidence point the same way.

**Inference (mine).** For a small, independent research operation, the conditions associated with disruption describe a structural advantage that is easy to waste. Canopticon is a small team, generalist, and switches topics often. Its documents span fields that do not cite each other. Those are the conditions the evidence associates with disruption. They are wasted if each document stays inside one field's literature.

### 4. What AI changes, and what the evidence for AI-driven discovery actually shows

The record's article on language models and idea diversity, and the "Net-Casting" document, already cover homogenisation within and across sessions. The newer evidence concerns outcomes.

**At the scale of science.** Hao, Xu, Li and Evans used a language model to identify AI-augmented research in 41.3 million natural-science papers. Validated against expert labels, the classifier had an F1 score of 0.875 [15]. Scientists doing AI-augmented work published 3.02 times more papers, received 4.84 times more citations, and became project leaders 1.37 years earlier. At the same time, AI adoption shrank the collective volume of topics studied by 4.63% and decreased scientists' engagement with one another by 22.00% [15]. AI-augmented work moves toward the areas richest in data, and the authors conclude that "AI tools appear to automate established fields rather than explore new ones" [15]. For anyone using AI to look for new findings, this is the central warning. The tool's default pull is toward where data, and therefore other people, already are.

**At the scale of ideas.** Si, Hashimoto and Yang had 43 expert researchers execute randomly assigned research ideas, some written by an LLM and some by human experts, and then had blind reviewers score the results [16]. Human ideas kept almost the same scores after execution (changes of −0.010 for novelty, +0.078 for excitement and −0.052 for effectiveness on a 10-point scale). LLM ideas fell significantly more on every metric, by 1.049 points on novelty for example. The gap that favoured LLM ideas at the proposal stage closed, and on several metrics human ideas ranked higher [16]. An idea's apparent novelty at first reading is a poor predictor of novelty once it is built.

**The best case for AI discovery.** Penadés and colleagues had spent over a decade asking how identical capsid-forming phage-inducible chromosomal islands (cf-PICIs) turn up in different bacterial species. They had answered it in a paper that was still confidential and under review [17]. They gave Google's AI co-scientist, a multi-agent system built on Gemini 2.0, a one-page brief of published background and asked why. Its top-ranked hypothesis was capsid–tail interactions: tail-less cf-PICI capsids pick up tails from different phages, and those tails determine which host the DNA reaches. The authors describe this as "remarkably recapitulating the novel hypothesis and key experimental findings of our unpublished paper" [17]. Five other systems given the same brief (OpenAI o1, o3-mini-high and deep research; Gemini 2.0 Pro and Flash Thinking) did not reach the mechanism [17]. The co-scientist also proposed that satellites might hitchhike on conjugative elements via oriT sequences. The authors say this had never been considered in their field, and their lab is now testing it [17].

The limits are equally clear. This is one question, evaluated by the lab whose result it was, against its own paper. The co-scientist ran for days, while the comparison systems ran for seconds or minutes and were sampled once [17]. The authors themselves caution that "consistency among AI-generated hypotheses should not be equated with validity or superiority" [17]. The conjugation idea is a candidate, not a finding.

**The cautionary case.** The most widely cited quantitative evidence for AI accelerating discovery was a preprint by an MIT economics PhD student, Aidan Toner-Rodgers. It reported that AI-assisted materials scientists at a large US firm produced 44% more new materials, filed 39% more patents and generated 17% more product innovations [19]. On 16 May 2025, MIT announced that after a confidential review it had "no confidence in the provenance, reliability or validity of the data". It asked arXiv to withdraw the paper and notified the Quarterly Journal of Economics, where it had been submitted [18]. Daron Acemoglu and David Autor, who had raised the concerns, said the findings "should not be relied on in academic or public discussions of these topics" [18]. Before that, the paper had been covered in Nature and widely discussed [19]. Any synthesis that cites a dramatic AI-productivity number should check whether the source still stands.

**Which account is more credible.** At the level of the whole field, the evidence (Hao et al.) is large-sample and consistent with the homogenisation findings already in the record. The evidence for AI producing validated new discoveries consists of a handful of self-evaluated cases and one withdrawn study. The defensible position is that AI widens the range of candidates quickly, is biased toward data-rich and crowded areas, and does not by itself validate anything. The Sourati–Evans result suggests the bias can be reversed deliberately: by modelling where experts are not looking, AI can be steered toward the gaps instead of the crowd [13].

### 5. Rules for choosing a question that can produce a new finding

This section is my inference. It turns sections 1–4 into criteria and states which evidence each rests on.

| Criterion | What it asks of a question | Evidence it rests on |
|---|---|---|
| **A. Disjoint pieces** | Are the parts of the answer already published, in literatures that do not cite each other? | Swanson; Sourati & Evans; cumulative meta-analysis [10][11][13] |
| **B. Conventional base, one odd joint** | Is it mostly built on solid, mainstream evidence, with one unusual connection? | Uzzi et al. [1] |
| **C. A new instrument or measure** | Does answering it require measuring something nobody has counted? | Leahey et al.; the record's Telescope Problem [3] |
| **D. Checkable end-point** | Does it end in a prediction or estimate that existing or soon-to-exist data can confirm or refute? | LBD evaluation failures; ideation–execution gap; MIT withdrawal [11][16][18] |
| **E. Away from the crowd** | Is it in an area where data-rich, AI-assisted work is not already concentrating? | Hao et al.; Sourati & Evans [13][15] |
| **F. Distance advantage** | Does the person asking bring a field the question's home field lacks? | Jeppesen & Lakhani; generalist and small-team findings [4][5][14] |

Two warnings come with the table. First, criterion B cuts against the instinct to chase the strangest idea. The hit profile is a familiar argument with one surprising joint [1]. Second, novel work is recognised late [2]. Early indifference from readers or reviewers is weak evidence against a question that scores well on A–F.

### 6. Applying the criteria to Canopticon's record

Given your decision to leave the scope to me, I scored the record's questions that already have research behind them against criteria A–F. The scores are my judgement, made from the record's own documents.

**1. Self-extinguishing maintenance (strongest candidate).** The record document *When Healed Is the Treatment Working* found the loop documented only in pieces. The parts are: the acute-model belief from Leventhal's Common-Sense Model, measured as the no-symptoms-no-asthma belief in 53% of 198 hospitalised adults; intentional non-adherence (stopping when one feels better), the main reason for stopping antidepressants at 46.6% in the WHO surveys; negative reinforcement and motivating operations from behaviour analysis; and the preparedness paradox from public health. The document found no source joining them into one cross-domain model, and none treating a persistent structural cause as the defining feature. Your article on self-extinguishing maintenance has since named the loop.

- *A (disjoint pieces).* High. Four literatures, each solid, none citing the others on this point.
- *B (conventional base).* High. Every link is established. Only the joining is new.
- *D (checkable).* High, and this is what lifts it above an essay. The model makes a prediction existing data could test: the hazard of stopping should rise after remission. It should rise more in conditions whose cause is structural or persistent than in conditions where remission reflects real resolution. Electronically monitored adherence datasets (the record cites Vrijens et al.'s 4,783 antihypertensive patients) and discontinuation surveys are the obvious places to look.
- *E (away from the crowd).* High. Adherence research treats stopping as a decision, not as a loop created by success.
- *Risk.* The loop may exist under a name the record's search missed, for example in the treatment-fatigue or drug-holiday literatures, or in organisational studies of regulation repealed after it has worked. A deliberate prior-art search across those vocabularies should come before any claim of originality.

**2. The self-service transfer (strong candidate, the instrument route).** The record document *Why Self-Directed Tasks Stall While Imposed Work Gets Done* states plainly that no study it located quantifies how many hours of administrative work digitisation shifted from institutions to individuals. The concept itself is not new. Public administration measures administrative burden as learning, compliance and psychological costs in citizens' dealings with the state (Moynihan, Herd and Harvey) [20]. Craig Lambert's 2015 book named *shadow work*, the unpaid tasks people now do on behalf of businesses and organisations, from booking travel to self-checkout [21]. What appears to be missing is the number.

- *A.* Medium-high. Three literatures meet here and do not cite each other on this point: administrative burden, shadow work, and prospective memory with executive function, which the record holds on the ADHD side.
- *C (new measure).* High. This is the instrument route the evidence favours [3]. The likely ingredients are national time-use surveys and organisations' own accounts of the staff hours that self-service replaced. Whether Statistics Canada's time-use categories separate this work cleanly is unverified and needs checking first.
- *D.* High. An hours estimate is checkable, and it implies a further testable claim: the burden falls hardest on people with the executive-function profile the record's ADHD documents describe.
- *Risk.* Administrative-burden researchers may have partial estimates for specific programmes. The original contribution would be the aggregate and the cognitive mechanism, not the concept.

**3. The precision ratio and the -crasis papers (original, but needs a test, not more synthesis).** The record's foundation research found no source joining McLuhan's sense ratios to precision weighting, interoception, allostasis or addiction. My own brief search this pass also found none, though that is weak evidence. On A and B the programme scores well. The same document rates the screen→interoception link as thin: one experience-sampling study, one smartphone-attention study and one lab trade-off effect. By criterion D, the next new finding will not come from another synthesis pass. It will come from a direct test of that weakest link, specifying which dimension of interoception (accuracy, sensibility, awareness or precision) the theory predicts will change.

**4. Algorithmic rent-setting detection (*RealPage North*; instrument route, data-limited).** The Competition Bureau closed its inquiry into RealPage Canada and Yardi Canada on 10 November 2025, finding adoption not sufficiently widespread, and the class action was discontinued [record: *RealPage North*]. With both accountability tracks closed, any new evidence would have to come from a detection instrument built on rental-listing data. That scores high on C and E. Feasibility depends on data access, which the record's own feasibility study treats as the binding constraint.

**5. Announced-versus-started housing audits (high verifiability, lower novelty).** The record's Toronto audit and its GTA programme review sit on primary records and produce checkable numbers. The GTA review names a real gap: no published counterfactual for how many rental starts the GST rebate, the Apartment Construction Loan Program (ACLP) or municipal incentives caused. That is valuable accountability work. On criteria A and B, though, it is closer to rigorous journalism than to a new finding, unless it builds a counterfactual method others can reuse.

**Recommendation (inference).** Lead with candidate 1, because it scores high on A, B, D and E together and can be tested against existing adherence data. Run candidate 2 alongside it as the instrument project, because new measures are the kind of novelty most associated with lasting influence [3]. For candidate 3, write a test protocol rather than another research pass.

## Beyond this question

The evidence implies a change to how Canopticon's research passes are run, not only to which questions they take on. Three practices follow directly from sections 2 and 4.

- **Run a disjoint-literature check.** Before calling a synthesis original, record which literatures each link draws on and whether they cite each other. This is a manual version of what Sourati and Evans automate [13].
- **End every exploratory pass in a falsifiable statement with a date.** This is the time-sliced logic that gave literature-based discovery its only honest evaluations [11]. Logged predictions turn a speculative record into a track record.
- **Check that dramatic AI or metascience numbers still stand.** The Toner-Rodgers figures circulated for six months before MIT disowned them [18][19].

One further point bears on the article on idea diversity. The Hao et al. result [15] measures, at the scale of science, the pull toward the data-rich centre that the record documents within single AI sessions. The two findings are the same pattern at two scales.

## Open questions

- **No base rate exists for literature-based hypotheses being confirmed.** The field's own evaluators could point to one founder's handful of discoveries and one mouse study [11]. I found no systematic count of how many published A–B–C hypotheses were later tested, or how many held.
- **The size of the disruption decline is unsettled.** The direction is supported by non-citation measures, but the Holst–Park exchange over zero-reference works and data cleaning has not been resolved by an independent team [5][6][7]. The main synthesis of the evidence was written by one side.
- **Evidence that AI produces validated discoveries is anecdotal.** One self-evaluated biology case [17], one withdrawn economics study [18], and a controlled study showing AI ideas lose ground when executed [16]. No pre-registered, multi-question test of AI hypothesis generation against later-validated results was found.
- **Whether these findings transfer outside academic science is untested.** Every measure here is built on papers, patents and citations. Whether "conventional base plus one odd joint" or "new instrument beats new theory" holds for investigative or public-interest research like Canopticon's is an inference, not a finding.
- **Coverage was thin in places.** I relied on abstracts or search extracts, not full texts, for Lau et al. [10], DiGiacomo et al. [12], Leahey et al. [3], Si et al. [16], the two Nature Matters Arising pieces [6][7] and the administrative-burden article [20]. For Shadow Work [21], only a review was available. Prior-art checks for candidates 1–3 were brief and should be repeated before any originality claim is published.

## Sources

1. [Atypical Combinations and Scientific Impact](https://www.kellogg.northwestern.edu/faculty/uzzi/htm/papers/science-2013-uzzi-468-72.pdf) — Uzzi, Mukherjee, Stringer & Jones, Science 342:468 (2013)
2. [Bias against Novelty in Science: A Cautionary Tale for Users of Bibliometric Indicators](https://www.nber.org/system/files/working_papers/w22180/w22180.pdf) — Wang, Veugelers & Stephan, NBER Working Paper 22180 (published Research Policy, 2017)
3. [What Types of Novelty Are Most Disruptive?](https://journals.sagepub.com/doi/10.1177/00031224231168074) — Leahey, Lee & Funk, American Sociological Review 88(3) (2023)
4. [Large teams develop and small teams disrupt science and technology](https://www.nature.com/articles/s41586-019-0941-9) — Wu, Wang & Evans, Nature 566 (2019)
5. [Is Innovation Becoming Less Disruptive? An Inventory of the Literature](https://arxiv.org/html/2602.05140) — Wu, Wu, Park, Leahey & Funk, arXiv (2026)
6. [Dataset artefacts can partially drive the measured decline in disruption](https://www.nature.com/articles/s41586-026-10787-y) — Holst, Algaba, Tori, Wenmackers & Ginis, Nature 656, Matters Arising (August 2026)
7. [Reply to: Dataset artefacts can partially drive the measured decline in disruption](https://www.nature.com/articles/s41586-026-10788-x) — Park, Leahey & Funk, Nature 656 (12 August 2026)
8. [Are Ideas Getting Harder to Find?](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3035132) — Bloom, Jones, Van Reenen & Webb, NBER Working Paper 23782 (2017)
9. [The Burden of Knowledge and the 'Death of the Renaissance Man': Is Innovation Getting Harder?](https://www.nber.org/papers/w11360) — Benjamin F. Jones, NBER Working Paper 11360 (2005; Review of Economic Studies, 2009)
10. [Cumulative Meta-Analysis of Therapeutic Trials for Myocardial Infarction](https://www.nejm.org/doi/full/10.1056/NEJM199207233270406) — Lau, Antman, Jimenez-Silva, Kupelnick, Mosteller & Chalmers, New England Journal of Medicine 327:248–254 (1992)
11. [Evaluation of Literature-based Discovery Systems](https://faculty.washington.edu/melihay/publications/LBDChapter2009.pdf) — Yetisgen-Yildiz & Pratt, University of Washington (book chapter, 2009)
12. [Fish-oil dietary supplementation in patients with Raynaud's phenomenon: a double-blind, controlled, prospective study](https://www.amjmed.com/article/0002-9343(89)90261-1/abstract) — DiGiacomo, Kremer & Shah, American Journal of Medicine 86(2):158–164 (1989)
13. [Accelerating science with human-aware artificial intelligence](https://arxiv.org/abs/2306.01495) — Sourati & Evans, arXiv (2023; Nature Human Behaviour 7:1682–1696)
14. [Marginality and Problem-Solving Effectiveness in Broadcast Search](https://ideas.repec.org/a/inm/ororsc/v21y2010i5p1016-1033.html) — Jeppesen & Lakhani, Organization Science 21(5) (2010)
15. [Artificial Intelligence Tools Expand Scientists' Impact but Contract Science's Focus](https://arxiv.org/abs/2412.07727) — Hao, Xu, Li & Evans, arXiv v4 (2025; accepted at Nature)
16. [The Ideation–Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas](https://arxiv.org/pdf/2506.20803) — Si, Hashimoto & Yang, arXiv (2025)
17. [AI mirrors experimental science to uncover a novel mechanism of gene transfer crucial to bacterial evolution](https://www.biorxiv.org/content/10.1101/2025.02.19.639094v1.full) — Penadés et al., bioRxiv (2025; later published in Cell)
18. [Assuring an accurate research record](https://economics.mit.edu/news/assuring-accurate-research-record) — MIT Department of Economics (16 May 2025)
19. [MIT says a high-profile AI productivity study used data that cannot be trusted](https://the-decoder.com/mit-says-a-high-profile-ai-productivity-study-used-data-that-cannot-be-trusted/) — The Decoder (17 May 2025); used only for the paper's reported figures, which MIT's statement does not restate
20. [Administrative Burden: Learning, Psychological, and Compliance Costs in Citizen-State Interactions](https://digitalgovernmenthub.org/library/administrative-burden-learning-psychological-and-compliance-costs-in-citizen-state-interactions/) — Moynihan, Herd & Harvey, Journal of Public Administration Research and Theory (2015), via Digital Government Hub
21. [Shadow Work: The Unpaid, Unseen Jobs That Fill Your Day (review)](https://www.libraryjournal.com/review/shadow-work-the-unpaid-unseen-jobs-that-fill-your-day) — Library Journal review of Craig Lambert's book (Counterpoint, 2015); no stronger source read