{"path":"research/tools-and-analogies-as-engines-of-discovery/analogy-in-science-sources-as-read-2026-11-05.md","content":"# Analogy in science: sources as a helper read them (2026-11-05)\n\nThis is the report of a read-only helper (a language model, Sonnet) made on 2026-11-05 for\n[tools-and-analogies-as-engines-of-discovery.md](../tools-and-analogies-as-engines-of-discovery.md),\nkept verbatim as it came back. Every number carries its source and a read-status tag the helper\nassigned. The helper was told what the corpus already holds and to skip it; its interpretation is\nmarked as its own at the end. What Claude made of it: the main note, § 4 and § 5.\n\n---\n\nRESEARCH REPORT: analogy in science, read against Night Science ep. 83\nDate of research: 5 Nov 2026. Nothing written to disk. Tags: [full text] [part] [abstract only] [secondary].\n\n0. THE EPISODE\n- No transcript exists on the episode page. It gives show notes only, so I did not summarise the audio. [part] https://nightscience.buzzsprout.com/1744020/episodes/18691165-83-how-science-is-secretly-driven-by-analogy-melanie-mitchell\n- Listed length is 33:19. Published 16 Feb 2026.\n- The notes say analogies often arise \"during mental rest\" and \"don't need to be accurate\". They also list: core concepts that started as metaphors, analogies that illuminate and mislead, whether LLMs grasp abstraction, Einstein, evolutionary \"landscapes\", and \"generative night science and critical day science\".\n- That list is a useful checklist for your transcription.\n\n1. MITCHELL'S OWN ACCOUNT\n\na) Annals NYAS 2021 / arXiv 2102.10717 [part: introduction read via search excerpts]\n- https://arxiv.org/abs/2102.10717 and https://nyaspubs.onlinelibrary.wiley.com/doi/10.1111/nyas.14619\n- What analogy IS: \"The process of abstraction is driven by analogy, in which one mentally maps the essence of one situation to a different situation.\"\n- Also: \"the primary way we humans make sense of novel situations is by making analogies to situations we have previously experienced.\"\n- She cites Hofstadter: \"a concept is a package of analogies.\"\n- What it is FOR in science: \"Many scientific insights are based on analogies, such as Darwin's realization that biological competition is analogous to economic competition and Von Neumann's analogies between the computer and the brain.\"\n- Two claims she makes: \"analogy-making is not a rare and exalted form of reasoning, but rather a constant, ubiquitous, and lifelong mode of thinking.\" And analogy is \"a central mechanism for unlocking meaning from perception.\"\n- Her misleading example is not about science. The pandemic was described as a fire, tsunami, tornado, volcano, \"another Katrina\" or a war.\n- Models: \"no current AI system is anywhere close to a capability of forming humanlike abstractions or analogies.\" [abstract only]\n\nb) ConceptARC (Moskvichev, Odouard & Mitchell 2023), https://arxiv.org/abs/2305.07141 [abstract only]\n- Humans \"substantially outperform\" the machine solvers. The solvers were the top two programs from the 2021 ARC competition, plus GPT-4.\n- I found no accuracy numbers in what I read.\n\nc) Substack \"AI: A Guide for Thinking Humans\", archive https://aiguide.substack.com/archive [part: titles only]\n- Relevant posts by title:\n  - 6 Oct 2025, \"Do AI Reasoning Models Abstract and Reason Like Humans?\"\n  - 14 Jan 2026, \"On Evaluating Cognitive Capabilities in Machines (and Other 'Alien' Intelligences)\"\n  - 10 Feb 2026, \"On Brian Cantwell Smith and the Promise of AI\"\n  - 21 Mar 2025, \"AI learns to reason (or does it?)\"\n  - \"Misleading Metaphors and Real Risks\" (year not shown)\n- I found no Substack post on analogy in scientific discovery.\n\nd) 6 Oct 2025 post, https://aiguide.substack.com/p/do-ai-reasoning-models-abstract-and [full text, via summarising fetch]\n- ARC is \"a test of abstraction and analogy, based on 'core knowledge' priors\".\n- Tested o3, Claude Sonnet 4 and Gemini 2.5 Pro on 480 ConceptARC tasks. Each rule was rated \"correct as intended\", \"correct but unintended\" or \"incorrect\".\n- Models are \"substantially more prone to use unintended 'shortcuts'\" than humans.\n- She concludes \"it is essential to go beyond accuracy\".\n- The post gives no aggregate percentages.\n\ne) 14 Jan 2026 post, https://aiguide.substack.com/p/on-evaluating-cognitive-capabilities [full text, via summarising fetch]\n- When the model's output grid was correct, it stated the intended rule about 70% of the time. Humans did so about 90%.\n- Lewis and Mitchell found GPT models much less robust than humans on altered letter-string problems. (You hold this material already.)\n- \"Rather than harder benchmarks, we need more rigorous evaluation methods.\"\n- o3 scored 87.5% on a 100-task semi-private ARC set at roughly $5,000 per task.\n\nf) Not found: any 2025–26 Mitchell paper or SFI talk on analogy in scientific discovery, incubation or mind-wandering. Episode 83 may be the only source.\n\n2. ANALOGY IN WORKING SCIENCE, MEASURED\n\na) Dunbar 1997, \"How scientists think: on-line creativity and conceptual change in science\" [part, long excerpt of full chapter]\n- https://www.ida.liu.se/divisions/hcs/seminars/cogsciseminars/Papers/Dunbar,%20K%20(1997)%20How_scientists_think_On-line_creativity_and_conceptual_change_in_science.pdf\n- Design: four molecular biology labs, one year, 4 meetings coded per lab (16 in all), two independent coders.\n- Frequency: 99 analogies in 16 meetings, mean 6.1 per meeting, range 2–14. Per lab: 25, 30, 31, 13.\n- Range, by the chapter's own labels: Within Organism 40, Other Organism 57, Non-Biological 2.\n  - Quote: \"only 2 of the 99 analogies used by the scientists were of this type.\"\n  - Dunbar had two more non-biological analogies elsewhere in his database. All four were used to explain.\n- Goals: explain 45, hypothesis 23, design experiment 21, fix experiment 10.\n- Distant analogies were \"rare and generally used for explanations rather than the generation of new hypotheses\". Hypotheses drew on other organisms, for example a gene in malaria and a similar gene in clams.\n- The TICS 2001 version (Dunbar & Blanchette) reports 3 to 15 analogies per one-hour meeting. [part] https://pages.ucsd.edu/~scoulson/203/dunbarTICS.pdf\n- A 1999 paper (Dunbar, J. Applied Dev. Psych) gives the same 99 and 16, and says: \"despite the anecdotal accounts of distant analogies having a major role in science, our findings indicate that distant…\" (text cut off). [part] https://nuovoeutile.it/wp-content/uploads/2015/01/How_scientists_think.pdf\n\nb) Where the popular accounts conflict with the primary data\n- The terms \"local / regional / long-distance\" do not appear in the 1997 chapter. It uses within organism / other organism / non-biological. Check the 1995 chapter if you need the three-way wording.\n- Secondary sources say successful labs used more long-distance analogies. I could not confirm that against the 1997 text, which says distant analogies were 2 of 99 and used for explaining. Treat the secondary claim as unverified. [secondary]\n- A later review says \"more than half of the analogies\" in labs and politics \"showed no apparent surface similarity\" (Kretz & Krawczyk, via search snippet). [secondary]\n- The Frontiers commentary says the molecular biologists \"still exhibited mostly within-domain analogizing\". [part] https://www.frontiersin.org/articles/10.3389/fpsyg.2014.01487/full\n- Blanchette & Dunbar 2001, Memory & Cognition 29(5):730–735, covers politics, not science. [abstract only]\n- Not found: Dunbar's finding on how distant analogies are remembered or forgotten. The only related item is secondary: scientists under-reported their analogy use in interviews.\n\nc) Nersessian [part]\n- Her 2002 chapter: https://mechanism.ucsd.edu/bill/teaching/F12/cs200/Readings/Nersessian.cognitivebasisofmodel-basedreasoninginscience.2002.pdf\n- Claim: \"analogies serve as sources of constraints for constructing models\".\n- Maxwell's source domain was continuum mechanics, with machine mechanics added later. \"No direct analogy existed\", so Maxwell built a hybrid.\n- The vortex–idle-wheel model shows \"a generic dynamical relational structure\" (generic abstraction). It was not meant as a literal picture of the aether.\n\nd) Gentner et al. 1997, J. Learning Sciences 6(1):3–40 [part]\n- https://www.tandfonline.com/doi/abs/10.1207/s15327809jls0601_2\n- Analogy changes knowledge in four ways: highlighting, projection, rerepresentation, restructuring.\n- Applied to Kepler's chain: light → vis motrix (a precursor of gravity), then magnetism → vis motrix.\n\n3. HOW ANALOGY MISLEADS, WITH CASES\n\n- **Brain metaphors.**\n  - Cobb 2021: Descartes' hydraulic idea failed because \"sectioning nerves showed there was no such fluid\". Telegraph followed in the 1830s. Telephone exchange was called flexible but had no feedback loops. Computer came last. https://www.frontiersin.org/journals/ecology-and-evolution/articles/10.3389/fevo.2021.760269/full [part]\n  - Gomez-Marin 2022: \"Neural plasticity soon hindered the [telegraph] analogy.\" https://www.frontiersin.org/journals/computer-science/articles/10.3389/fcomp.2022.890531/full [part]\n  - Corrected by new experiments (nerve sectioning, plasticity), not by a better analogy.\n- **Kelvin's vortex atom (1867).**\n  - Mapped Helmholtz's permanent vortex rings onto atoms. Silliman: it was \"eventually forgotten\" after Thomson, Rutherford and Bohr. https://webhomes.maths.ed.ac.uk/~v1ranick/papers/silliman.pdf [part]\n  - Falconer 2019: stability could not be shown. The energy of a vortex ring is a saddle point, and Kelvin proved stability only for a single columnar vortex. https://royalsocietypublishing.org/rsta/article/377/2158/20180451/58118/Vortices-and-atoms-in-the-Maxwellian-era [part]\n  - Wilczek: it failed through \"sterility\" and rival theories, not internal contradiction. https://www.pbs.org/wgbh/nova/article/beautiful-losers-kelvins-vortex-atoms/ [part]\n- **Rutherford's planetary atom.**\n  - Classical electrodynamics says orbiting electrons radiate and collapse. Bohr imposed stability \"by fiat\" using Planck's constant.\n  - Arabatzis & Ioannidou argue the negative analogy, electrons repel each other unlike Saturn's ring particles, motivated Bohr. http://publ.royalacademy.dk/backend/web/uploads/2020-02-14/AFL%206/SDM_1_00_00_2015_5858/SDM_1_18_00_2015_5974.pdf [part]\n  - Here the misleading analogy was productive.\n- **Immune system as army.**\n  - War metaphors \"shaping the very way scientists think… build theories\" (cites Martin 1994). The strong/weak defence framing neglects immune activity unrelated to defence. https://pmc.ncbi.nlm.nih.gov/articles/PMC9929241/ [part]\n- **Genetic program / organism as machine.**\n  - Nicholson 2014 calls it \"logically incoherent\". It gives biologists \"the illusion that they can predict and control development\". https://doi.org/10.1016/j.shpsc.2014.08.003 [part]\n  - I did not read Keller 2000 itself.\n- **Selfish gene.**\n  - Noble 2011: a metaphor \"that struggles to chime with modern biology\". No experiment distinguishes it from \"cooperative gene\", and Dawkins conceded he doubted any experiment could. https://pmc.ncbi.nlm.nih.gov/articles/PMC3060581/ [part]\n  - Noble & Noble 2025 argue it misled medical genomics. This is a partisan source. https://link.springer.com/article/10.1007/s11692-025-09651-0 [abstract only]\n- **Junk DNA.**\n  - Eddy 2012: Ohno's term was \"polarizing\". The idea was supported by Orgel/Crick and Doolittle/Sapienza (1980) with the selfish-DNA mechanism. Only about 5–20% of the human genome is under detectable selection. ENCODE's \"80%\" claim was about biochemical activity. https://www.cell.com/current-biology/fulltext/S0960-9822(12)01154-2 [part]\n  - Moran notes Ohno did not coin the term. This is a metaphor dispute, not a clean \"misled\" case.\n\nEmpirical bias evidence: Thibodeau & Boroditsky 2011, PLoS ONE, doi 10.1371/journal.pone.0016782 [full text, via summarising fetch] https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0016782\n- Crime as beast vs virus; the outcome is the share proposing enforcement.\n- Exp. 1: 74% vs 56%, χ²=13.94, p<.001 (485 recruited, 455 coded).\n- Exp. 2: 71% vs 54%, χ²=6.50, p<.05 (347 recruited, 253 analysed).\n- Exp. 3 (synonym prime, no metaphor): 64% vs 65%, no effect.\n- Exp. 4: 40% vs 22% chose criminal justice, χ²=5.72, p<.05.\n- Exp. 5 (information seeking): 69% vs 64%, p=.59.\n- Only 3% (Exp. 1) and 7% (Exp. 2) cited the metaphor as influential.\n- The paper reports no effect sizes. The participants were students, not scientists.\n- I found no empirical test of analogy bias in science writing proper.\n\n4. INCUBATION AND INSIGHT\n\na) Sio & Ormerod 2009, Psychological Bulletin 135(1):94–120 [full text, extracted from the PDF]\n- 117 effect sizes, 3,606 participants, median 25 per study. 85 effects were positive.\n- Weighted mean d=0.29, 95% CI 0.21–0.39. Unweighted mean 0.41, or 0.36 after outliers were recoded.\n- Heterogeneity: Q=173.99, p<.001. The funnel-plot regression found no sign of publication bias (β=−.08, p=.41).\n- Table 4 weighted means:\n  - linguistic insight 0.22 (65 studies)\n  - second problem type 0.26 (35 studies; the label read \"Verbal\" in my text extraction, which I take to be \"visual\")\n  - creative / divergent 0.29 (14 studies, CI 0.02–0.55)\n  - high-load interpolated task 0.24 (76)\n  - low-load task 0.52 (22)\n  - unfilled rest 0.46 (16)\n- Longer preparation before incubation raised the effect (β=.03, p<.05).\n- For linguistic problems, a low-load task beat rest alone (β=.54, p<.05).\n- A secondary source gives d=.65 for divergent tasks. I could not verify that figure in the text, where creative problems come out at 0.29. Do not use .65 unchecked. [secondary]\n- The table shows \"rest\" is not the best filler. A mildly occupying low-load task did as well or better.\n\nb) Kounios & Beeman 2009, Current Directions in Psychological Science [abstract only] https://journals.sagepub.com/doi/10.1111/j.1467-8721.2009.01638.x\n- Insight is \"the culmination of a series of brain states and processes operating at different time scales.\"\n\nc) Jung-Beeman et al. 2004, PLoS Biology [part] https://journals.plos.org/plosbiology/article?id=10.1371%2Fjournal.pbio.0020097\n- fMRI: more activity in the right anterior superior temporal gyrus for insight solutions.\n- EEG: a gamma burst starting about 0.3 s before the solution response, at the right temporal site T8 (t[18]=3.48, p=.003). It was preceded by right posterior alpha, roughly 1.4 to 0.4 s before.\n- The authors tie this area to \"making connections across distantly related information\".\n\nd) Kounios et al. 2006, resting-state EEG predicts later problem-solving strategy [part] https://ncbi.nlm.nih.gov/pmc/articles/PMC2293274/\n- Group differences in resting-state EEG matched insight-style versus search-style solvers. I have no effect sizes.\n\ne) Default-mode network\n- A 2024 stereo-EEG study in 13 patients found DMN recruitment in both mind-wandering and alternate-uses tasks. Direct cortical stimulation of DMN regions \"preferentially decreased the originality\" of responses, with no effect on fluency or mind-wandering. [part] https://pmc.ncbi.nlm.nih.gov/articles/PMC11449134/\n- A 2022 ALE meta-analysis (27 divergent-thinking and 30 insight fMRI studies) found no overlapping areas between divergent thinking and insight. Divergent thinking leaned on the inferior parietal lobe, a DMN hub. [abstract only] https://pmc.ncbi.nlm.nih.gov/articles/PMC9582370/\n- I did not read Beaty's DMN papers.\n- Neither source links DMN activity to analogy in particular.\n\n5. MACHINES AND SCIENTIFIC ANALOGY, 2024–26\n\n- Shen, Zou et al. (Stanford), \"Unlocking LLM Creativity in Science through Analogical Reasoning\", arXiv 2605.11258, May 2026 [part, abstract and results text] https://arxiv.org/html/2605.11258v1\n  - Prompting for cross-domain analogies gave novel solutions about 50% of the time, against as little as 1.6% for a no-domain baseline.\n  - Binary novelty: Claude 1.6% / 21.6% / 50.4%, GPT 9.0% / 28.2% / 58.8%, Gemini 8.9% / 37.8% / 68.9% (no-domain / cross-domain / analogical).\n  - Humans rated the analogical solutions more novel than cross-domain ones 78% of the time (κ=0.445) and \"reasonable\" 67% of the time.\n  - Novelty and quality were mostly LLM-judged. Four biomedical tasks showed gains, for example ρ=0.729 against published brain-region methods.\n  - \"Held up\" was measured on benchmark tasks, not on expert judgement of the analogies' soundness.\n- \"On the Diversity of Analogy Making in LLMs\", arXiv 2608.03233, Aug 2026 [part] https://arxiv.org/html/2608.03233v1\n  - Ten LLMs show \"domain homogeneity\". Asked for 10 distinct analogies, they produced about 2 semantic clusters on average.\n  - Best intra-output diversity was 0.49 (Gemini 3.1 Pro). Raising diversity cost quality.\n  - This is the closest result to your \"mostly famous ones\" question. It is about narrow domains, not fame as such.\n- LacMaterial, arXiv 2510.22312, Oct 2025: LLM cross-domain analogies for battery materials, validated computationally only. [abstract only] https://arxiv.org/html/2510.22312\n- ResearchBench, arXiv 2503.21248 [part] https://arxiv.org/html/2503.21248\n  - Picking the top 4% of candidate \"inspirations\", GPT-4o included the ground-truth inspiration 45.7% of the time.\n- Teaching analogies (Feb 2025, arXiv 2502.16895): GPT-4o analogies for science concepts show \"object confusion\" errors. [part] https://arxiv.org/abs/2502.16895\n- A systematic review of 83 analogy-generation papers finds LLM generation \"rarely operates without a human-in-the-loop curation step\". [abstract only] https://aclanthology.org/2026.findings-emnlp.463.pdf\n- I found no 2025–26 study that checks LLM-proposed cross-field analogies against later scientific outcomes.\n\n6. A TESTABLE PAIR WITH GROUND TRUTH: MAXWELL'S VORTEX MODEL vs KELVIN'S VORTEX ATOM\n- Same source domain, 1861–62 versus 1867. The sources are Falconer 2019 (RSTA) and Silliman, linked in section 3, plus Nersessian 2002 in section 2c.\n- Productive: Maxwell mapped a sea of ether vortices onto Faraday's lines of force.\n  - Vortices stood for the magnetic field and idle-wheel motion for electric current.\n  - Maxwell discarded the mechanism but kept the mathematical analogy between the field equations and vortex dynamics.\n  - Maxwell himself warned the contact mechanism was not to be read literally.\n- Misleading: Kelvin (1867) mapped Helmholtz's indestructible vortex rings in a perfect fluid onto atoms. Permanence, identity and vibration modes were meant to explain atomic spectra.\n  - The critical disanalogy: stability. Kelvin could not prove vortex-atom stability, and the energy argument failed.\n  - Falconer notes Moffatt's suggestion that a magnetic analogy might have fared better, since stable knotted magnetic structures do exist.\n  - Wilczek's gloss: it was sterile and outcompeted.\n- A candidate rule to test: does the mapping carry the target's invariants as a mathematical structure, or does it import a physical mechanism (permanence) that the source cannot guarantee? Maxwell kept the first and discarded the second. Kelvin relied on the second.\n- Caveat: Maxwell's model was also later dropped as a mechanism, and Kelvin's work enriched hydrodynamics. \"Productive\" and \"misleading\" are fuzzy here.\n- Alternative pair: Bohr's use of Rutherford's planetary atom, a \"wrong\" analogy that was productive via its negative analogy.\n\nMY INTERPRETATION (clearly mine)\n- Dunbar's counts (2 of 99 non-biological, with those used for explanation) undercut a story that science runs on far-flung analogy. Mitchell's claim in the Annals paper is about analogy as pervasive and structural, not distant. Check on the audio which she stresses.\n- The \"relaxed mind\" claim has modest support: d about 0.29 overall. A low-load task did better than empty rest. Nothing I found ties incubation to analogy specifically.\n- The LLM evidence leans toward \"novel but narrow\", with the quality check mostly done by LLM judges.\n- The Maxwell/Kelvin pair works as a test case only if you pre-register what counts as \"invariant carried\".\n\nNot verified, in one place: Keller 2000; Dunbar's memory/forgetting finding; Dunbar 1995's three-way local/regional/distant counts; any 2025–26 Mitchell piece on science specifically.\n"}