Pith. sign in

REVIEW 3 major objections 7 minor 67 references

Even when generative AI improves research outputs, it can still degrade research by eroding the practices that form scholarly judgement.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 22:01 UTC pith:GQAR4NQ7

load-bearing objection Solid conceptual package on formation vs. output; the recognition/generation hinge is asserted harder than the short-horizon evidence supports, but the paper is still worth refereeing. the 3 major comments →

arxiv 2607.04049 v1 pith:GQAR4NQ7 submitted 2026-07-04 cs.AI

What is Left for Us? Second Scholarship Against the Degradation of Research by AI

Claudio Novelli , Luciano Floridi This is my paper
classification cs.AI
keywords generative AIscholarly judgementsecond scholarshipresearch formationtacit knowledgedeep readingacademic trustlarge language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that generative AI degrades research not mainly by producing bad papers, but by displacing the lived practices through which scholarly judgement is formed and academic trust is built. Those practices—reading closely, arguing through friction, committing personally, and being socialised into a community—are constitutive of knowledge, not mere means to an output that AI can simulate. When researchers hand central tasks of inquiry to large language models, the individual result may look better while the researcher fails to develop; at scale, a field then loses the very judgement needed to check the technology. Merely keeping humans in the loop as prompters or quality-checkers is therefore insufficient. What is required is “second scholarship”: a deliberate, post-critical re-appropriation of the craft after experiencing what AI can and cannot do, grounded in four non-automatable sources—tacit knowledge, personal commitment, socialisation, and deep reading—and backed by institutions that prize what cannot be delegated.

Core claim

Even when generative AI assistance improves individual research outputs, it may still degrade the formative conditions of research: researchers who delegate central acts of inquiry stop doing the work through which scholarly judgement is formed, so the person behind the work fails to develop. “Humans in the loop” as checkers cannot replenish that formation. The alternative is second scholarship—the re-appropriation of scholarly craft after a critical encounter with AI—resting on tacit knowledge, personal commitment, socialisation, and deep reading.

What carries the argument

Second scholarship: the post-critical re-appropriation of scholarly craft, chosen after a critical experience of what generative AI can and cannot do. It carries the argument by naming the stance that preserves formation (rather than mere oversight of outputs) and by anchoring it in four non-automatable sources and warrants: tacit knowledge, personal commitment, socialisation, and deep reading.

Load-bearing premise

Supervising or verifying an AI’s output does not exercise the generative capacity that actually doing the inquiry builds, so a policy of oversight spends the capital of formation without replenishing it.

What would settle it

A controlled longitudinal comparison in which trainees who critically supervise LLM outputs throughout their training later match, on independent generative tasks (producing original arguments, designs, or interpretations without AI), the judgement, knowledge transfer, and ownership of those who performed the formative work themselves.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Academia’s reward systems must shift from output volume toward activities hard to automate—live defence of claims, deep reading, and communal exchange—or formation will continue to atrophy while metrics rise.
  • Early training must deliberately protect first-scholarship practices; otherwise later researchers have nothing left to re-appropriate as second scholarship.
  • If reviewers and authors rely on LLMs for reading and writing, the community’s main checks on accuracy weaken exactly when AI-generated volume increases.
  • Individual delegation at scale erodes the aggregate of formed judgements on which the reliability of any AI-assisted research programme depends.
  • What resists automation—answerability for claims one has actually reasoned one’s way to—will tend to become what the profession most visibly prizes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If generative capacity atrophies while recognition remains, fields may lose the ability to spot genuine anomalies and become locked into ever more refined “normal science.”
  • The live-music analogy implies that oral defences, vivas, and high-quality conference exchange could become premium credentials once text is cheap to produce.
  • Hermeneutic drift is testable: compare independent reconstructions of a source text after iterative LLM dialogue versus after direct deep reading, scoring fidelity to the original and the reader’s ability to defend the reading without the model.
  • Institutions that only add disclosure rules for AI use will leave the researcher-in-the-loop paradigm intact and therefore leave the formation problem unaddressed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper argues that generative AI can degrade research even when it improves individual outputs, because delegating central acts of inquiry erodes the formative practices through which scholarly judgement and academic trust are built. Against the standard remedy of keeping humans in the loop as prompters or quality-checkers, the authors contend that supervision without generative exercise spends the capital of formation without replenishing it. They defend re-embodiment of research via four non-automatable sources and warrants—tacit knowledge, personal commitment, socialisation, and deep reading—and introduce ‘second scholarship’ (after Ricoeur’s second naïveté) as the post-critical re-appropriation of scholarly craft. They also name ‘hermeneutic drift’ as progressive displacement of a source text’s meaning under iterative LLM dialogue, and sketch institutional reforms that would make unautomatable activities legible to career incentives.

Significance. If the formation thesis holds, the paper reframes the AI-and-research debate from output quality and reliability toward constitutive conditions of inquiry, with direct implications for training, peer review, and research assessment. The framing of second scholarship, the four warrants, and the hermeneutic-drift diagnosis are original contributions that synthesise Polanyi, Dewey, Longino, Collins, and Ricoeur with current empirical work on metacognitive laziness, cognitive offloading, and sycophancy. The institutional proposals in §6, though schematic, correctly identify that disclosure rules leave the researcher-in-the-loop paradigm intact. The manuscript is carefully hedged on provisional evidence (e.g., Kosmyna et al. 2025) and engages objections in §5 rather than ignoring them. Its significance is conceptual and policy-facing rather than empirical.

major comments (3)
  1. §5 (objection on recognition vs generation): The hinge against ‘humans in the loop’ is the claim that supervising or verifying LLM output does not exercise generative capacity and therefore cannot form the judgement first scholarship builds. The supporting citations (Fan et al. 2025; Lee et al. 2025; Kosmyna et al. 2025) concern short-horizon student essay tasks and knowledge-worker surveys, not longitudinal formation of researchers who iteratively critique, redesign, and defend AI-assisted work under peer pressure. The paper acknowledges the objection but treats the asymmetry as settled by those studies. Either (a) weaken the claim to a risk that needs empirical investigation at research timescales, (b) supply a stronger philosophical argument distinguishing generative formation from recognitional competence independent of those studies, or (c) engage the counter-possibility that sustai
  2. §4.4 (hermeneutic drift): Hermeneutic drift is introduced as a central failure mode of AI-mediated reading and is used to motivate second scholarship’s appropriation requirement. The authors correctly state they have not measured it and ground it in sycophancy and cognitive-bias literature. For a concept that does this much work, the manuscript should more explicitly mark it as a testable hypothesis rather than a diagnosed mechanism, and specify what would count as evidence for or against progressive displacement of source meaning across iterative LLM dialogue. Without that, the concept risks functioning as a name for the conclusion rather than an independently checkable intermediate claim.
  3. §6.2 (institutional proposals): The three lines of reform—reweighting in-person participation, rethinking metrics, and raising submission costs—are presented as directions entailed by the diagnosis. The authors flag gaming risks (Matthew Effect on conference programmes; successive proxy capture) but do not resolve how quality of live exchange or unautomatable craft is to be made legible without new proxies that recreate the pathology. Because the paper’s constructive answer depends on institutions that reward what cannot be automated, this design gap is load-bearing for the ‘what is left for us’ claim. At minimum, state more clearly which proposals are diagnostic illustrations versus actionable recommendations, and what success criteria would distinguish re-embodiment from prestige-weighted theatre.
minor comments (7)
  1. Introduction: ‘Ths can be seen as a form of re-embodiment’ — typo for ‘This’.
  2. §5 and bibliography: ‘Felgenbaum 1977’ should be ‘Feigenbaum’ (Edward A. Feigenbaum).
  3. Throughout: special-character ligatures (e.g., efficiency, insufficient, difficulty) appear inconsistently and may cause PDF/search issues; normalise to standard ASCII or proper Unicode.
  4. §4.1: The distinction between relational, somatic, and collective tacit knowledge (Collins 2013) is useful; a brief forward pointer to how ‘epistemic disturbance’ maps onto that typology would help readers who know only Polanyi.
  5. §3: The Tao astronomy anecdote is effective; a citation or source for the public conversation would aid verification.
  6. Bibliography: Several arXiv and preprint items (Kosmyna et al. 2025; Liu et al. 2026; Mann et al. 2025) are appropriately used as provisional; ensure status labels remain consistent in the final version.
  7. §6.1: Nguyen’s ‘trust as an unquestioning attitude’ and ‘agential gullibility’ are well deployed; a one-sentence contrast with reliance (already present later) earlier in the subsection would tighten the calibration argument.

Circularity Check

1 steps flagged

No significant circularity: conceptual argument with independent external warrants; only minor non-load-bearing self-positioning via Floridi’s onlife thesis.

specific steps
  1. self citation load bearing [Introduction / §4 (re-embodiment paragraph)]
    "in a world that is by now thoroughly onlife, in which the boundary between online and offline has dissolved to the point where it is no longer sensible to ask on which side of it one stand (Floridi 2014, 2015), there is no offline to which one could retreat. Re-embodiment, as we use it, is the continued exercise of the formative activities that AI delegation displaces, carried on within the onlife condition rather than against it."

    Co-author Floridi’s prior ‘onlife’ concept is cited to frame re-embodiment as remaining inside a hybrid environment. This is ordinary conceptual self-positioning, not a load-bearing premise: the formation claim and four warrants stand independently of the onlife thesis, which only rules out one rejected alternative (unplugging). It does not make the conclusion true by construction.

full rationale

This is a philosophical essay, not a derivation with equations, fitted parameters, uniqueness theorems, or empirical predictions. Its central claim (that AI-improved outputs can still degrade research by eroding formative practices of judgement) is supported by four external warrants—tacit knowledge (Polanyi, Collins), personal commitment (Polanyi’s fiduciary programme), socialisation (Longino, Hardwig, Lave & Wenger), and deep reading (Wolf)—plus empirical citations (Fan et al. 2025, Lee et al. 2025, Kosmyna et al. 2025) that are independent of the authors. ‘Second scholarship’ is an explicit adaptation of Ricoeur’s second naïveté, not a renaming that forces the conclusion by construction. The sole self-referential element is co-author Floridi’s prior ‘onlife’ framing (Floridi 2014, 2015), used only to dismiss wholesale unplugging as infeasible; it does not define or force the formation argument, the recognition/generation asymmetry, or the four warrants. No step reduces a claimed result to its own inputs by definition, fit, or load-bearing self-citation chain. Score 1 reflects the minor, non-load-bearing self-positioning only.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 3 invented entities

The paper is conceptual rather than quantitative; it has no free parameters fitted to data. Its load-bearing commitments are domain assumptions drawn from philosophy of science and cognitive science, plus two invented entities that organize the positive proposal. No mathematical axioms are used.

axioms (5)
  • domain assumption Tacit knowledge (Polanyi) is in principle non-codifiable and is transmitted only through structured social immersion, not text alone.
    Invoked throughout §4.1 as the reason LLMs can produce only “pseudo-interactional expertise.”
  • domain assumption Personal commitment (Polanyi’s fiduciary programme) is required for a claim to be answerable and for reputation to function as an epistemic warrant.
    §4.2; without it, the social value of the formed scholar collapses.
  • domain assumption Scientific knowledge and trust relations are irreducibly social (Longino, Hardwig, Collins) and require moral agents who can be held accountable.
    §4.3; used to exclude LLMs from genuine participation in research communities.
  • domain assumption Deep reading circuits develop with use and atrophy with disuse; AI-mediated reading removes the integrative cognitive work that builds them.
    §4.4, drawing on Wolf and meta-analyses of screen vs. print comprehension.
  • domain assumption There is a standing asymmetry between the capacity to recognize a good argument and the capacity to generate one; only the latter forms judgment.
    §5 reply to the supervision objection; load-bearing for the claim that human-in-the-loop is insufficient.
invented entities (3)
  • second scholarship no independent evidence
    purpose: Names the post-critical, deliberate re-appropriation of scholarly craft after critical experience of generative AI, by analogy with Ricoeur’s second naïveté.
    Central positive proposal of the paper; organizes the institutional recommendations in §6.
  • hermeneutic drift no independent evidence
    purpose: Names the progressive displacement of a source text’s meaning across iterative LLM dialogue under sycophancy and user framing.
    Introduced in §4.4 as a specific failure mode of AI-mediated reading; authors note it has not been measured.
  • first scholarship no independent evidence
    purpose: Labels the pre-critical, inherited practice of research before AI, to set up the three-stage movement to second scholarship.
    Contrast term introduced in the Introduction; not independently evidenced.

pith-pipeline@v1.1.0-grok45 · 25972 in / 3126 out tokens · 33824 ms · 2026-07-11T22:01:48.344401+00:00 · methodology

0 comments
read the original abstract

We argue that generative AI can degrade research by eroding the very practices through which scholarly judgement is formed and academic trust is built. As constitutive conditions for the production and validation of knowledge, these practices cannot be reduced to the final outputs of research, which is what AI so effectively simulate. Accordingly, when researchers delegate central tasks of inquiry to systems like Large Language Models, they may stop enacting these practices and, with them, lose access to the formation they provide. An individual research output generated by AI may even appear improved but the researcher behind it fails to develop. Against this risk, merely keeping humans in the loop as prompters or quality checkers of AI outputs is insufficient to preserve research as a site of intellectual formation. What is needed instead is a renewed commitment to research as a lived practice in which judgement is formed gradually, often through frictions, and participation in a scholarly community. We defend it because it rests on four sources and warrants of research that cannot be automated: tacit knowledge, personal commitment, socialisation, and deep reading. This practice enacts what we call second scholarship, by which we understand the reappropriation of scholarly craft, chosen out of a critical experience of what generative AI can and cannot do. What cannot and should not be delegated becomes what research communities must value and answer for. This is what is left for us.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 31 canonical work pages

  1. [1]

    humans in the loop

    about their application to empirical research. For instance, the combination of LLMs with large, publicly available datasets – notably biomedical ones – has been associated with a proliferation of low-quality or formulaic research output (Naddaf 2025). Similarly, the use of LLMs to generate synthetic data as a substitute for primary quantitative survey me...

    2025

  2. [2]

    prepared mind

    Inside the Loop: Is That Enough? A growing body of scholarship argues that AI-generated scientific output is not reliable without some form of human input, revision, or verification. Agrawal, McHale, and Oettl (2026, NBER WP 34953, pre-publication at the time of writing), for instance, offer a useful framework by dividing scientific production into four main ...

    2026

  3. [3]

    progressive

    to suggest that humans excel at the kind of abductive inference that AI currently struggles to replicate, not least because the creative “guess” involved in formulating an explanatory hypothesis also requires a degree of personal commitment from the scientist (Agrawal et al. 2026, 18–19). Although LLMs are highly capable of 6 This is taken from Charles Pe...

    2026

  4. [4]

    The mathematician Terence Tao, in a recent public conversation on AI and scientific discovery, used the history of astronomy to illuminate a related risk

    at the expense of paradigm-breaking inquiry. The mathematician Terence Tao, in a recent public conversation on AI and scientific discovery, used the history of astronomy to illuminate a related risk. He noted that Ptolemy's geocentric model had been refined over a millennium through increasingly elaborate modifications, becoming remarkably precise (and yet w...

    1970

  5. [5]

    optimisation

    Beyond the loop: Re-emboding Research That LLMs make writing easier is by now a commonplace, and it has made us worry that writing, as a mode of thinking, is under pressure. A quieter crisis concerns reading. An undergraduate who wants to grasp Wittgenstein's picture theory of language no longer needs to read the Tractatus; they can ask an LLM for a summa...

    2010

  6. [6]

    desirable difficulties

    reports EEG evidence of reduced brain connectivity among essay writers using ChatGPT and a diminished sense of ownership over their texts. This preprint should be treated as provisional: the sample is small (n = 54 in the main sessions; n = 18 in the critical fourth session), institutionally homogeneous (Boston-area university students, ages 18–39), and t...

    1991

  7. [7]

    Objections to our argument Before moving to the constructive, so-what?, part of the paper, let us address two potential objections against our analysis. A first objection, both against our analysis so far and what comes next, may arrive from that strand in the philosophy of cognitive science which argues that LLMs are best understood as the latest instance...

    2024

  8. [8]

    Without structural support, they remain normative aspirations easily overridden by existing incentive pressures

    What Is Left for Us: How to Develop Second Scholarship? The four arguments for re-embodiment presented above do not become operative on their own. Without structural support, they remain normative aspirations easily overridden by existing incentive pressures. Some concrete measures to contrast the uncontrolled use of AI in research have already been adopt...

    2022

  9. [9]

    rethink its metrics

    and the Leiden Manifesto were formulated for precisely this kind of pathology, though before AI made it acute. An update may be necessary. We want to be careful here, however, because the conclusion that academia should “rethink its metrics” is easier to state than to act on. The pull of metrification is not just that the current proxies are bad; it is tha...

    2019

  10. [10]

    Working Paper No

    ‘AI in Science’. Working Paper No. 34953. Working Paper Series. National Bureau of Economic Research, March. https://doi.org/10.3386/w34953. Autor, David

    work page doi:10.3386/w34953

  11. [11]

    Working Paper No

    ‘Polanyi’s Paradox and the Shape of Employment Growth’. Working Paper No. 20485. Working Paper Series. National Bureau of Economic Research, September. https://doi.org/10.3386/w20485. Bainbridge, L

    work page doi:10.3386/w20485

  12. [12]

    In Analysis, Design and Evaluation of Man–Machine Systems, edited by G

    ‘IRONIES OF AUTOMATION’. In Analysis, Design and Evaluation of Man–Machine Systems, edited by G. Johannsen and J. E. Rijnsdorp. Pergamon. https://doi.org/10.1016/B978-0-08-029348-6.50026-9. Bisbee, James, Joshua D. Clinton, Cassy Dorff, Brenton Kenkel, and Jennifer M. Larson

    work page doi:10.1016/b978-0-08-029348-6.50026-9

  13. [13]

    Political Analysis 32 (4): 401–16

    ‘Synthetic Replacements for Human Survey Data? The Perils of Large Language Models’. Political Analysis 32 (4): 401–16. https://doi.org/10.1017/pan.2024.5. Bjork, Elizabeth Ligon, and Robert A. Bjork

    doi:10.1017/pan.2024.5 2024

  14. [14]

    Journal of Applied Research in Memory and Cognition (Netherlands) 9 (4): 475–79

    ‘Desirable Difficulties in Theory and Practice’. Journal of Applied Research in Memory and Cognition (Netherlands) 9 (4): 475–79. https://doi.org/10.1016/j.jarmac.2020.09.003. Broek, Paul van den

    doi:10.1016/j.jarmac.2020.09.003 2020

  15. [15]

    Science (New York, N.Y.) 328 (5977): 453–56

    ‘Using Texts in Science Education: Cognitive Processes and Knowledge Representation’. Science (New York, N.Y.) 328 (5977): 453–56. https://doi.org/10.1126/science.1182594. Buçinca, Zana, Maja Barbara Malaya, and Krzysztof Z. Gajos

    work page doi:10.1126/science.1182594

  16. [16]

    Proceedings of the ACM on Human-Computer Interaction 5 (CSCW1): 1–21

    ‘To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making’. Proceedings of the ACM on Human-Computer Interaction 5 (CSCW1): 1–21. https://doi.org/10.1145/3449287. Carter, J. Adam

    doi:10.1145/3449287

  17. [17]

    Synthese 197 (7): 2937–61

    ‘Intellectual Autonomy, Epistemic Dependence and Cognitive Enhancement’. Synthese 197 (7): 2937–61. https://doi.org/10.1007/s11229-017-1549-y. Cassinadri, Guido

    work page doi:10.1007/s11229-017-1549-y

  18. [18]

    Chapuis, Kevin, Patrick Taillandier, and Alexis Drogoul

    https://doi.org/10.1007/s13347-024-00701-7. Chapuis, Kevin, Patrick Taillandier, and Alexis Drogoul

    work page doi:10.1007/s13347-024-00701-7

  19. [19]

    Science Studies 4 (2): 165–85

    ‘The TEA Set: Tacit Knowledge and Scientific Networks’. Science Studies 4 (2): 165–85. https://doi.org/10.1177/030631277400400203. Collins, Harry

    work page doi:10.1177/030631277400400203

  20. [20]

    AI & SOCIETY 40 (3): 1249–63

    ‘Why Artificial Intelligence Needs Sociology of Knowledge: Parts I and II’. AI & SOCIETY 40 (3): 1249–63. https://doi.org/10.1007/s00146-024-01954-8. Collins, Harry, and Simon Thorne

    doi:10.1007/s00146-024-01954-8

  21. [21]

    Csibra, Gergely, and György Gergely

    https://doi.org/10.1007/s11229-026-05531-y. Csibra, Gergely, and György Gergely

    work page doi:10.1007/s11229-026-05531-y

  22. [22]

    Trends in Cognitive Sciences 13 (4): 148–53

    ‘Natural Pedagogy’. Trends in Cognitive Sciences 13 (4): 148–53. https://doi.org/10.1016/j.tics.2009.01.005. Delgado, Pablo, Cristina Vargas, Rakefet Ackerman, and Ladislao Salmerón

    doi:10.1016/j.tics.2009.01.005 2009

  23. [23]

    Educational Research Review 25 (November): 23–38

    ‘Don’t Throw Away Your Printed Books: A Meta-Analysis on the Effects of Reading Media on Reading Comprehension’. Educational Research Review 25 (November): 23–38. https://doi.org/10.1016/j.edurev.2018.09.003. DeStefano, Diana, and Jo-Anne LeFevre

    work page doi:10.1016/j.edurev.2018.09.003 2018

  24. [24]

    Computers in Human Behavior, Including the Special Issue: Avoiding Simplicity, Confronting Complexity: Advances in Designing Powerful Electronic Learning Environments, vol

    ‘Cognitive Load in Hypertext Reading: A Review’. Computers in Human Behavior, Including the Special Issue: Avoiding Simplicity, Confronting Complexity: Advances in Designing Powerful Electronic Learning Environments, vol. 23 (3): 1616–41. https://doi.org/10.1016/j.chb.2005.08.012. Dewey, John

    work page doi:10.1016/j.chb.2005.08.012 2005

  25. [25]

    How We Think

    How We Think. How We Think. D C Heath. https://doi.org/10.1037/10903-000. Douglas, Heather

    work page doi:10.1037/10903-000

  26. [26]

    British Journal of Educational Technology 56 (2): 489–530

    ‘Beware of Metacognitive Laziness: Effects of Generative Artificial Intelligence on Learning Motivation, Processes, and Performance’. British Journal of Educational Technology 56 (2): 489–530. https://doi.org/10.1111/bjet.13544. Felgenbaum, Edward A

    doi:10.1111/bjet.13544

  27. [27]

    Springer

    The Onlife Manifesto - Being Human in a Hyperconnected Era. Springer. https://doi.org/10.1007/978-3-319-04093-6. Fonagy, Peter, and Elizabeth Allison

    doi:10.1007/978-3-319-04093-6

  28. [28]

    Psychotherapy (US) 51 (3): 372–80

    ‘The Role of Mentalizing and Epistemic Trust in the Therapeutic Relationship’. Psychotherapy (US) 51 (3): 372–80. https://doi.org/10.1037/a0036505. Gerlich, Michael

    work page doi:10.1037/a0036505

  29. [29]

    Societies 15 (1)

    ‘AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking’. Societies 15 (1). https://doi.org/10.3390/soc15010006. Ginzburg, Carlo

    doi:10.3390/soc15010006

  30. [30]

    Episteme 8 (1): 112–25

    ‘The Division of Epistemic Labor’. Episteme 8 (1): 112–25. https://doi.org/10.3366/epi.2011.0010. Goldberg, Sanford C

    work page doi:10.3366/epi.2011.0010 2011

  31. [31]

    The Journal of Philosophy 88 (12): 693–708

    ‘The Role of Trust in Knowledge’. The Journal of Philosophy 88 (12): 693–708. https://doi.org/10.2307/2027007. Heersmink, Richard, Barend de Rooij, María Jimena Clavel Vázquez, and Matteo Colombo

    work page doi:10.2307/2027007

  32. [32]

    Hou, Chenyu, Gaoxia Zhu, Vidya Sudarshan, Fun Siong Lim, and Yew Soon Ong

    https://doi.org/10.1007/s10676-024-09777-3. Hou, Chenyu, Gaoxia Zhu, Vidya Sudarshan, Fun Siong Lim, and Yew Soon Ong

    work page doi:10.1007/s10676-024-09777-3

  33. [33]

    Computers & Education 234 (September): 105329

    ‘Measuring Undergraduate Students’ Reliance on Generative AI during Problem-Solving: Scale Development and Validation’. Computers & Education 234 (September): 105329. https://doi.org/10.1016/j.compedu.2025.105329. Howells, Jeremy

    doi:10.1016/j.compedu.2025.105329 2025

  34. [34]

    Technology Analysis & Strategic Management 8 (2): 91–106

    ‘Tacit Knowledge’. Technology Analysis & Strategic Management 8 (2): 91–106. https://doi.org/10.1080/09537329608524237. Kambhampati, Subbarao

    work page doi:10.1080/09537329608524237

  35. [35]

    Communications of the ACM 64 (2): 31–32

    ‘Polanyi’s Revenge and AI’s New Romance with Tacit Knowledge’. Communications of the ACM 64 (2): 31–32. https://doi.org/10.1145/3446369. Keller, Timothy A., Robert A. Mason, Aliza E. Legg, and Marcel Adam Just

    work page doi:10.1145/3446369

  36. [36]

    Khoo, Shaun Yon-Seng

    https://doi.org/10.1038/s41539-024-00232-y. Khoo, Shaun Yon-Seng

    work page doi:10.1038/s41539-024-00232-y

  37. [37]

    LIBER Quarterly: The Journal of the Association of European Research Libraries 29 (1): 1–18

    ‘Article Processing Charge Hyperinflation and Price Insensitivity: An Open Access Sequel to the Serials Crisis’. LIBER Quarterly: The Journal of the Association of European Research Libraries 29 (1): 1–18. https://doi.org/10.18352/lq.10280. Koskinen, Inkeri

    work page doi:10.18352/lq.10280

  38. [38]

    Social Epistemology 38 (4): 458–75

    ‘We Have No Satisfactory Social Epistemology of AI-Based Science’. Social Epistemology 38 (4): 458–75. https://doi.org/10.1080/02691728.2023.2286253. Kosmyna, Nataliya, Eugene Hauptmann, Ye Tong Yuan, et al

    work page doi:10.1080/02691728.2023.2286253 2023

  39. [39]

    arXiv:2506.08872

    ‘Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task’. arXiv:2506.08872. Preprint, arXiv, December

    Pith/arXiv arXiv

  40. [40]

    Krenn, Mario, Robert Pollice, Si Yue Guo, et al

    https://doi.org/10.48550/arXiv.2506.08872. Krenn, Mario, Robert Pollice, Si Yue Guo, et al

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.08872

  41. [41]

    Nature Reviews

    ‘On Scientific Understanding with Artificial Intelligence’. Nature Reviews. Physics 4 (12): 761–69. https://doi.org/10.1038/s42254-022-00518-3. Kuhn, T. S

    doi:10.1038/s42254-022-00518-3

  42. [42]

    Science 390 (6779): 1240–43

    ‘Scientific Production in the Era of Large Language Models’. Science 390 (6779): 1240–43. https://doi.org/10.1126/science.adw3000. Lakatos, I

    doi:10.1126/science.adw3000

  43. [43]

    In Criticism and the Growth of Knowledge: Proceedings of the International Colloquium in the Philosophy of Science, London, 1965, edited by Alan Musgrave and Imre Lakatos, vol

    ‘Falsification and the Methodology of Scientific Research Programmes’. In Criticism and the Growth of Knowledge: Proceedings of the International Colloquium in the Philosophy of Science, London, 1965, edited by Alan Musgrave and Imre Lakatos, vol

    1965

  44. [44]

    https://doi.org/10.1017/CBO9781139171434.009

    Cambridge University Press. https://doi.org/10.1017/CBO9781139171434.009. Lave, Jean, and Etienne Wenger

    work page doi:10.1017/cbo9781139171434.009

  45. [45]

    Lee, Hao-Ping (Hank), Advait Sarkar, Lev Tankelevitch, et al

    https://doi.org/10.3390/e28040377. Lee, Hao-Ping (Hank), Advait Sarkar, Lev Tankelevitch, et al

    work page doi:10.3390/e28040377

  46. [46]

    Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (New York, NY, USA), CHI ’25, April 25, 1–22

    ‘The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers’. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (New York, NY, USA), CHI ’25, April 25, 1–22. https://doi.org/10.1145/3706598.3713778. Lee, John D., and Katrina A. See

    doi:10.1145/3706598.3713778 2025

  47. [47]

    Human Factors 46 (1): 50–80

    ‘Trust in Automation: Designing for Appropriate Reliance’. Human Factors 46 (1): 50–80. https://doi.org/10.1518/hfes.46.1.50_30392. Liang, Weixin, Yaohui Zhang, Zhengxuan Wu, et al

    doi:10.1518/hfes.46.1.50_30392

  48. [48]

    Nature Human Behaviour 9 (12): 2599–609

    ‘Quantifying Large Language Model Usage in Scientific Papers’. Nature Human Behaviour 9 (12): 2599–609. https://doi.org/10.1038/s41562-025-02273-8. Liu, Jiachen, Jiaxin Pei, Jintao Huang, et al

    doi:10.1038/s41562-025-02273-8

  49. [49]

    arXiv:2604.24658

    ‘The Last Human-Written Paper: Agent-Native Research Artifacts’. arXiv:2604.24658. Preprint, arXiv, May

    Pith/arXiv arXiv

  50. [50]

    Longino, Helen E

    https://doi.org/10.48550/arXiv.2604.24658. Longino, Helen E

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.24658

  51. [51]

    McIntyre, Lee

    https://arxiv.org/abs/2509.14189v3. McIntyre, Lee

    Pith/arXiv arXiv

  52. [52]

    Nature 641 (8065): 1080–81

    ‘AI Linked to Explosion of Low-Quality Biomedical Research Papers’. Nature 641 (8065): 1080–81. https://doi.org/10.1038/d41586-025-01592-0. Naddaf, Miryam

    work page doi:10.1038/d41586-025-01592-0

  53. [53]

    Nguyen, C

    https://doi.org/10.1038/d41586-026-01973-z. Nguyen, C. Thi

    work page doi:10.1038/d41586-026-01973-z

  54. [54]

    American Psychologist (US) 23 (1): 27–43

    ‘Logic and Psychology’. American Psychologist (US) 23 (1): 27–43. https://doi.org/10.1037/h0037692. Polanyi, Michael

    work page doi:10.1037/h0037692

  55. [55]

    AI and Ethics 5 (2): 1499–521

    ‘The Ethics of Using Artificial Intelligence in Scientific Research: New Guidance Needed for a New Tool’. AI and Ethics 5 (2): 1499–521. https://doi.org/10.1007/s43681-024-00493-8. Ricoeur, Paul

    doi:10.1007/s43681-024-00493-8

  56. [56]

    arXiv:2310.13548

    ‘Towards Understanding Sycophancy in Language Models’. arXiv:2310.13548. Preprint, arXiv, May

    Pith/arXiv arXiv

  57. [57]

    Sher, Gila

    https://doi.org/10.48550/arXiv.2310.13548. Sher, Gila

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.13548

  58. [58]

    Erkenntnis 72 (2): 151–76

    ‘Epistemic Friction: Reflections on Knowledge, Truth, and Logic’. Erkenntnis 72 (2): 151–76. https://doi.org/10.1007/s10670-009-9202-x. Stadler, Matthias, Maria Bannert, and Michael Sailer

    work page doi:10.1007/s10670-009-9202-x

  59. [59]

    Computers in Human Behavior 160 (November): 108386

    ‘Cognitive Ease at a Cost: LLMs Reduce Mental Effort but Compromise Depth in Student Scientific Inquiry’. Computers in Human Behavior 160 (November): 108386. https://doi.org/10.1016/j.chb.2024.108386. Talwar, Victoria, Jennifer Lavoie, and Angela M. Crossman

    doi:10.1016/j.chb.2024.108386 2024

  60. [60]

    Applied Developmental Science 26 (3): 553–66

    ‘Socialization of Lying Scale: Development and Validation of a Parent Measure of Socialization of Truth and Lie-Telling Behavior’. Applied Developmental Science 26 (3): 553–66. https://doi.org/10.1080/10888691.2021.1927732. Tennant, Jonathan P., François Waldner, Damien C. Jacques, Paola Masuzzo, Lauren B. Collister, and Chris. H. J. Hartgerink

    work page doi:10.1080/10888691.2021.1927732 2021

  61. [61]

    Thornton, Tim

    https://doi.org/10.12688/f1000research.8460.3. Thornton, Tim

    work page doi:10.12688/f1000research.8460.3

  62. [62]

    Tomasello, Michael

    https://doi.org/10.1186/1747-5341-1-2. Tomasello, Michael

    work page doi:10.1186/1747-5341-1-2

  63. [63]

    Proceedings of the National Academy of Sciences 115 (34): 8491–98

    ‘How Children Come to Understand False Beliefs: A Shared Intentionality Account’. Proceedings of the National Academy of Sciences 115 (34): 8491–98. https://doi.org/10.1073/pnas.1804761115. Turner, Stephen

    work page doi:10.1073/pnas.1804761115

  64. [64]

    Proceedings of the National Academy of Sciences 122 (47): e2518075122

    ‘The Potential Existential Threat of Large Language Models to Online Survey Research’. Proceedings of the National Academy of Sciences 122 (47): e2518075122. https://doi.org/10.1073/pnas.2518075122. Wolf, M., and Mirit Barzillai

    doi:10.1073/pnas.2518075122

  65. [65]

    arXiv:2411.11581

    ‘OASIS: Open Agent Social Interaction Simulations with One Million Agents’. arXiv:2411.11581. Preprint, arXiv, March

    Pith/arXiv arXiv

  66. [66]

    Zhu, Lingxuan, Yancheng Lai, Jiarui Xie, et al

    https://doi.org/10.48550/arXiv.2411.11581. Zhu, Lingxuan, Yancheng Lai, Jiarui Xie, et al

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.11581

  67. [67]

    Clinical and Translational Discovery 5 (4): e70067

    ‘Evaluating the Potential Risks of Employing Large Language Models in Peer Review’. Clinical and Translational Discovery 5 (4): e70067. https://doi.org/10.1002/ctd2.70067

    work page doi:10.1002/ctd2.70067