> For the complete documentation index, see [llms.txt](https://www.parapathology.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://www.parapathology.com/theories-and-frameworks/theoretical-frameworks-for-understanding-pathology-practice.md).

# Theoretical Frameworks for Understanding Pathology Practice: A Landscape Map

## TL;DR

* The theoretical landscape applied to pathology splits cleanly into three layers: **cognitive/perceptual theories** of the diagnostic act itself (gestalt, dual-process, expertise, illness scripts, visual search, recognition-primed decision making), **system/social theories** of pathology as an organized practice (complex adaptive systems, actor-network theory, Abbott's sociology of professions, NASSS, normalization process theory, communities of practice, resilience engineering, disruptive innovation), and **adoption theories** specifically explaining why AI/digital pathology stalls (NASSS, automation bias, trust calibration, sociotechnical co-evolution, medico-legal frameworks).
* The most empirically grounded frames *for pathology specifically* are **dual-process theory + illness scripts** (cognition), **eye-tracking/visual-search studies of expert pathologists** (Crowley, Mello-Thoms, Brunyé, Elmore), and **NASSS/normalization process theory** (adoption); the **complex-adaptive-systems "attractor" framing in the Garratt LinkedIn post** is metaphorical/heuristic rather than empirically validated, but maps to a real literature on healthcare CAS and to documented automation-bias and workflow-inertia phenomena.
* If you want to *explain why AI fails to take root in pathology*, the most defensible synthesis combines (i) the **H\&E-microscope sociotechnical attractor** (CAS + actor-network + Abbott jurisdiction), (ii) **NASSS** complexity domains (technology, value, adopter, organization, wider system, time), and (iii) **dual-process/expertise** accounts showing that pathologists' System-1 gestalt is genuinely fast and cheap to substitute against — i.e., AI must clear a high cognitive-ergonomic bar, not just a statistical one.

***

## Key Findings

1. **Pathology has a deep cognitive-science literature** (Crowley, Norman, Nodine, Mello-Thoms, Elmore, Brunyé) that explicitly applies gestalt perception, dual-process theory, expertise/deliberate practice, visual-search theory, and cognitive-bias theory to microscope diagnosis. These are established, peer-reviewed applications, not metaphors.
2. **System-level theories are more heterogeneously applied.** Complex adaptive systems and chaos theory are widely used in medicine generally (Goldberger, Plsek, Greenhalgh) and have been imported into pathology only loosely — most prominently by commentators like John Garratt who frames H\&E as a "stable attractor" in pathology's diagnostic CAS. Actor-network theory and Abbott's sociology of professions are rarely cited in pathology journals directly but underlie much of the STS literature on laboratory medicine.
3. **Implementation/adoption theories now dominate the AI-in-pathology discussion.** The NASSS framework (Greenhalgh 2017) has been explicitly recommended for digital-pathology deployment (Betmouni, *Digital Health*, 2021). Normalization process theory (May) has been used in actual pathology rollouts (Mikkelsen et al. 2022 in Denmark; Kamath et al. 2024 in dermatopathology at Mayo Clinic). Disruptive-innovation theory (Christensen) is invoked routinely by vendors and commentators.
4. **AI adoption literature now empirically documents automation bias and anchoring in computational pathology** (Rosbach et al. 2024/2026 — a 7% automation-bias rate when pathologists estimated tumor cell percentages under time pressure with AI assistance). Medico-legal frameworks (FDA/CE-IVDR/lab-developed-test "homebrew" pathway, Calderaro & Kather, *J Pathol* 2025) are now an explicit object of theory.
5. **Several theoretical lenses are still under-applied to pathology**, despite obvious fit: cognitive-load theory, embodied cognition (the microscope as a tool that extends perception), Klein's recognition-primed decision making, and Safety-II/resilience engineering. These represent opportunities for novel framing rather than well-developed literatures.

***

## Details

### CATEGORY 1 — Theories of the cognitive and perceptual act of diagnosis

**Gestalt perception / "the gestalt" diagnostic impression.** Gestalt psychology (Wertheimer, Koffka, Köhler, 1920s) describes how the visual system organizes elements into wholes (figure-ground, closure, proximity, similarity). In pathology, pathologists explicitly talk about a case's "gestalt" — a holistic, low-magnification impression that often precedes feature-by-feature analysis. This is the dominant folk theory of fast diagnosis and underlies the perceptual primacy of low-power scanning. Empirically supported by eye-tracking work showing expert pathologists fixate on diagnostically relevant regions within the first few seconds (Mello-Thoms et al., *Arch Pathol Lab Med* 2012; Brunyé et al., *J Pathol Inform* 2014, 2017). It is also the namesake of multiple AI vendors (e.g., Gestalt Diagnostics) — a marketing tell that "the gestalt" is the cognitive ground digital tools are trying to replicate.

**Pattern recognition / template matching.** Closely linked to gestalt; the idea that recognition is matching incoming stimuli to stored templates. In medicine, pattern recognition is the canonical System-1 process in dual-process accounts of diagnostic reasoning (Pelaccia et al., *Med Educ Online* 2011). In pathology, Crowley et al. (*JAMIA* 2003, "Development of visual diagnostic expertise in pathology — an information-processing study") used think-aloud protocols with 28 novices, intermediates, and experts on breast pathology and found expertise differences in *search, perception, and reasoning* — i.e., template/pattern recognition is the operationalized core of diagnostic expertise.

**Dual-process theory (System 1 / System 2).** Originated in cognitive psychology (Evans, Stanovich; popularized for medicine by Kahneman's *Thinking, Fast and Slow*, and applied to clinical reasoning by Croskerry, Norman, Pelaccia). System 1 is fast/intuitive/pattern-driven; System 2 is slow/analytic/rule-driven. Pelaccia et al. (*Medical Education Online* 2011) is the canonical clinical application; Norman et al. (*J Eval Clin Pract* 2024, "Dual process models of clinical reasoning: the central role of knowledge in diagnostic expertise") is the recent critical revision arguing that *knowledge*, not the system itself, drives accuracy. Pathology-specific: most experienced surgical pathologists routinely report rendering a System-1 diagnosis "at the scope," then engaging System 2 only for difficult or atypical cases — the same model documented by van Diest, Foucar, and others.

**Expertise theory (novice-to-expert, Dreyfus, deliberate practice, chunking).** Dreyfus & Dreyfus (1986) novice-to-expert five-stage model; Ericsson's deliberate-practice account (focused practice with feedback); Chase & Simon's chunking. In pathology: Crowley et al. (2001 AMIA, 2003 JAMIA) showed expertise differences in visual diagnostic processes; Brunyé et al. (PLOS One 2014; *J Pathol Inform* 2017) extended this with eye-tracking, demonstrating that lower-expertise pathologists fixate disproportionately on visually salient but diagnostically irrelevant regions. Ericsson's deliberate-practice paradigm directly informs pathology residency curricula (sign-out volume, structured feedback).

**Cognitive load theory (Sweller).** Distinguishes intrinsic, extraneous, and germane load. Underapplied in pathology but increasingly invoked in commentary on digital workflow design (scrolling whole-slide images, multiple synoptic checkboxes, AI overlays). A frequent argument: poorly designed AI overlays add extraneous cognitive load and slow rather than speed sign-out. Mostly editorial/commentary level in pathology journals; well-established in medical-education research.

**Bayesian / probabilistic diagnostic reasoning.** Bayes's theorem applied to pre- and post-test probability. In pathology, explicit in cytopathology (Bethesda categories), molecular reporting (variant classification), and biomarker interpretation (e.g., HER2 IHC scoring, where pathologists update belief based on staining patterns and clinical priors). Most pathologists use Bayesian reasoning implicitly rather than computationally. Friston's free-energy principle / predictive coding (see "Other lenses") offers a more recent formalization.

**Hypothetico-deductive reasoning and illness scripts.** Elstein, Shulman & Sprafka (1978) established hypothetico-deductive reasoning empirically; Feltovich & Barrows (1984) introduced *illness scripts* — structured mental representations of diseases comprising enabling conditions, fault (pathophysiology), and consequences. Custers (*Med Teach* 2015, "Thirty years of illness scripts") and Lubarsky et al. (*Adv Health Sci Educ* 2015) review the theory. In pathology, residents build "morphologic illness scripts" tied to entities (e.g., the script for low-grade serous carcinoma includes architectural features, nuclear grade, expected mitotic rate, expected IHC profile, and characteristic molecular alterations). Schmidt & Boshuizen's "knowledge encapsulation" model explains why expert scripts are compact yet rich.

**Visual search theory and eye-tracking of pathologists.** Originated with Kundel & Nodine in radiology (1970s–80s) — they proposed the "global-focal" model: a holistic global impression in the first second, followed by directed search to confirm/refute. Brought into pathology by Crowley, then extended by Mello-Thoms (*Arch Pathol Lab Med* 2012, dermatopathology virtual slides), Brunyé/Elmore (PLOS One 2014; *J Pathol Inform* 2017, "Accuracy is in the Eyes of the Pathologist," with eye-tracking on 12 breast WSI cases including atypia/DCIS), and the Brunyé et al. 2024 review "Eye tracking in digital pathology: A comprehensive literature review" (*J Pathol Inform*). The PathoGaze1.0 dataset (Thai V, Li R, Ling M, Jiang S, Wolfe J, Machiraju R, Hu Y, Li Z, Parwani A & Chen J, arXiv:2510.24653, October 2025) now provides large-scale gaze + decision data — 19 pathologists interpreting 397 WSIs, 18.69 hours of recordings, 171,909 fixations, and 263,320 saccades.

**Heuristics and cognitive biases in diagnostic error.** Croskerry's catalog of cognitive biases (anchoring, availability, confirmation, satisfaction of search, premature closure, framing) — *Acad Med* 2003 and onwards. Anchoring is now empirically documented in computational pathology (Rosbach et al., arXiv 2024/2026 — pathologists adopt AI-suggested tumor cell percentages even when wrong, with a 7% automation-bias rate). Satisfaction of search (Tuddenham 1962, Berbaum 1990s) was originally a radiology construct but applies to pathology — finding one lesion and missing a second nearby; documented in eye-tracking studies showing reduced gaze dwell time after a first target is found.

**Naturalistic decision making / recognition-primed decision (Klein).** Gary Klein's RPD model (1989, popularized in *Sources of Power*, 1999) describes how experts in time-pressured, ill-structured environments recognize a situation by pattern-matching to prior cases, mentally simulate a response, and act. RPD has been applied widely in nursing and emergency medicine but only sparingly to pathology, despite an excellent fit: a surgical pathologist viewing a frozen section is a paradigmatic time-pressured expert pattern-matcher. This is a *novel/under-applied* frame for pathology.

**Embodied cognition and the microscope as tool.** Rooted in Merleau-Ponty, extended by Andy Clark and others (the "extended mind"). The microscope is not merely an instrument but an extension of perceptual cognition — focus knob, condenser, stage all become extensions of the pathologist's perceptual loop. The shift to digital whole-slide images breaks this embodiment and forces new motor schemas (mouse, scroll, zoom). Gibson/Norman affordance theory is the related lens: glass slides afford random-access pan-and-zoom by hand; WSI affords mouse-mediated, latency-dependent navigation — explaining perceived productivity loss in digital workflow studies. These frames are *metaphorically* present in commentary but rarely operationalized in empirical pathology research.

***

### CATEGORY 2 — Theories of pathology as a system, organization, or social practice

**Complex adaptive systems (CAS) and attractor states.** Originating with Holland, Gell-Mann, and Santa Fe Institute work, formalized for healthcare by Plsek & Greenhalgh (*BMJ* 2001, "Complexity science: the challenge of complexity in health care"). A CAS has agents interacting under simple rules, generating emergent patterns; attractors are stable behavioral basins toward which the system tends. **John Garratt (Director, Canadian Pathology Quality Assurance Inc.) explicitly frames pathology this way in a public LinkedIn comment:** "The key might be the H\&E if we consider it to be a stable attractor in a diagnostic system which we are using to observe complex biological systems. Histology is then a spatial snapshot of a dynamic complex adaptive system that we have to learn how to read." This framing is heuristic rather than peer-reviewed, but it crystallizes a real empirical phenomenon: the H\&E/glass-slide/microscope workflow has resisted displacement for \~140 years despite repeated waves of "disruptive" technology. CAS theory predicts that to move a system out of an attractor, one needs either a stronger competing attractor or a perturbation large enough to push the system over a separatrix.

**Chaos theory and nonlinear dynamics.** Goldberger (*Lancet* 1996, "Non-linear dynamics for clinicians: chaos theory, fractals, and complexity at the bedside") is the canonical clinical introduction. In pathology specifically, chaos/fractal frameworks have been applied to *carcinogenesis and tumor morphology* (fractal dimension of tumor borders, irregular nuclear contours) — see Baish & Jain, *Cancer Res* 2000 on fractals in tumor vasculature. Application of chaos theory to *pathology practice itself* (as opposed to the diseases pathology studies) is largely metaphorical.

**Sociotechnical systems theory.** Originating with Trist & Bamforth (Tavistock Institute, 1951, coal-mining studies). Healthcare applications by Berg, Coiera. The core claim: technical and social subsystems jointly determine performance; optimizing one in isolation degrades the other. Digital pathology rollouts that focus only on scanner throughput while ignoring sign-out workflow, MDT integration, and trainee learning routinely underperform — a textbook sociotechnical failure mode.

**Actor-network theory (Latour, Callon, Law).** From the STS tradition, ANT treats humans and non-humans (microscopes, scanners, antibodies, LIS systems, regulatory agencies, the AI model itself) as actants in networks of association. Latour & Woolgar's *Laboratory Life* (1979) is the foundational laboratory ethnography. ANT has been used to analyze health-IT implementation (Cresswell et al., *BMC Med Inform Decis Mak* 2010) but rarely with explicit reference to anatomic pathology. The framework is especially apt for pathology because it forces analysts to take the slide, the stain, and the scanner as agents — exactly Garratt's "H\&E attractor" intuition, expressed in STS vocabulary.

**Sociology of professions and Abbott's jurisdiction theory.** Andrew Abbott, *The System of Professions* (1988): professions compete for jurisdiction over tasks; jurisdictional claims are made in legal, public, and workplace arenas. Abbott explicitly treats *diagnosis–inference–treatment* as the elemental work of professional jurisdiction. Pathology has fought multiple jurisdictional battles — with cytotechnologists, with radiologists over imaging-derived biomarkers, with clinical geneticists over molecular reports, and now with AI vendors and informaticians. Boundary work (Gieryn 1983) describes how professions defend their cognitive territory; "workplace assimilation" (Abbott) describes how junior groups blur boundaries on the ground.

**Diffusion of innovations (Rogers) and TAM/UTAUT.** Rogers's *Diffusion of Innovations* (1962, 5th ed. 2003) describes innovation attributes (relative advantage, compatibility, complexity, trialability, observability) and adopter categories (innovators, early adopters, early majority, late majority, laggards). The Technology Acceptance Model (Davis 1989) and UTAUT (Venkatesh et al. 2003) focus on perceived usefulness and ease of use. These have been applied to digital pathology adoption in survey studies but are widely criticized as too thin for healthcare — leading to the NASSS framework.

**NASSS framework (Greenhalgh et al. 2017, *****JMIR*****).** Non-adoption, Abandonment, Scale-up, Spread, Sustainability. Seven domains: condition/illness, technology, value proposition, adopter system, organization, wider system, embedding/adaptation over time. Each rated simple/complicated/complex; complexity in multiple domains predicts failure. **Explicit application to pathology**: Betmouni (*Digital Health* 2021, "Diagnostic digital pathology implementation: Learning from the digital health experience") explicitly recommends NASSS for digital pathology deployment, writing: "I would recommend adapting the NASSS framework for the planning, deployment and monitoring of digital pathology deployment in clinical practice. This will facilitate effective redesign of the diagnostic histopathology workflow; which in itself will prepare the ground for eventual deployment of AI tools in diagnostic practice." This is the single most useful theory-import for understanding AI adoption failure in pathology.

**Normalization process theory (Carl May).** Four constructs: coherence (sense-making), cognitive participation (engagement), collective action (doing the work), reflexive monitoring (appraisal). Used in actual pathology rollouts: Mikkelsen et al. (*Int J Environ Res Public Health* 2022) used the NoMAD instrument to assess staff readiness for digital pathology in the Region of Southern Denmark; Kamath CC, Wissler Gerdes EO, Barry BA, Minteer SA, Comfere NI, Peters MS, Wieland CN et al., "Staff Experiences Transitioning to Digital Dermatopathology in a Tertiary Academic Medical Center" (*Mayo Clinic Proceedings: Digital Health* 2024 May; 2(3):289–298) used NPT to study a Mayo Clinic multi-site transition (22 interviews, 34 surveys) and found earlier NPT stages (understanding, participation) were better supported than later stages (doing it, reflecting on it).

**Communities of practice / situated learning (Lave & Wenger).** Pathology training is a paradigmatic community of practice: legitimate peripheral participation moves residents from peripheral (gross room, low-stakes cases) to central (independent sign-out). MDTs and tumor boards are inter-professional CoPs. Useful frame for understanding why AI tools that don't fit the existing CoP's discourse fail to embed — a junior pathologist learning IHC patterns via informal mentorship has no analogous channel for learning AI-output interpretation.

**Workflow / human factors / ergonomics and resilience engineering / Safety-I vs Safety-II.** Hollnagel's distinction (*From Safety-I to Safety-II*, 2015): Safety-I focuses on counting and eliminating failures; Safety-II focuses on understanding why things normally go right and supporting performance variability. Pathology has historically been a Safety-I field (defect rates, amendment rates, second-review programs) but the laboratory's ability to absorb pre-analytic variation, missing clinical history, and edge cases is a textbook Safety-II story. Under-applied in pathology journals but growing.

**Translational science / "valley of death."** The gap between bench science (or AI model development) and clinical deployment. In pathology AI, this maps to a stark gap: as of August 2025, exactly three histopathology AI tools hold FDA authorization — Paige Prostate (de novo, 2021), Ibex Prostate Detect (510(k) clearance, February 2025), and ArteraAI Prostate (de novo, August 2025) — against thousands of academic models published. By comparison, radiology has "a couple of hundred" AI/ML devices cleared (Fraunhofer IIS, July 2025), illustrating the magnitude of pathology's translational gap.

**Disruptive innovation (Christensen).** Innovations enter at the low end (or new markets), are initially inferior on traditional metrics, and improve to overtake incumbents. Routinely invoked for digital pathology (Sectra commentary; Hwang & Christensen 2008 on healthcare disruption). The theory has been heavily criticized as post-hoc/imprecise (Lepore *New Yorker* 2014; Christensen et al. *J Manag Stud* 2018), but the language is ubiquitous in pathology marketing.

***

### CATEGORY 3 — Theorizing AI adoption specifically in pathology

**Why digital/computational pathology stalls.** Reis-Filho & Kather (*JNCI* 2023, "Overcoming the challenges to implementation of artificial intelligence in pathology") name "2 fundamental paradigm shifts": digital diagnosis and AI-rendered final diagnosis, noting that "the long-awaited adoption of AI in pathology... has not materialized, and the transformation of pathology is happening at a much slower pace than that observed in other fields (eg, radiology)." Tizhoosh & Pantanowitz (*J Pathol Inform* 2018, "Artificial Intelligence and Digital Pathology: Challenges and Opportunities") is the most-cited foundational barrier paper. Swillens et al. (*Oncogene* 2023) provide an international survey of pathologists' barriers and facilitators using an implementation-science framework. Bessen et al. (*Diagnostics* 2025) document that "there is hesitancy among pathologists, many of whom feel more comfortable with the manual pathology workflows they trained with" — empirical confirmation of the H\&E attractor.

**NASSS for pathology AI.** Betmouni (2021) is the explicit application; NASSS-in-radiology AI work is more developed and provides a template. The framework maps pathology AI failure modes onto specific domains: technology (interoperability, DICOM-Path, scanner heterogeneity), value proposition (no clear reimbursement, productivity benefit unclear at low volumes), adopter system (pathologist trust, deskilling fears), organization (IT capacity, capital cost), wider system (FDA/CE-IVDR, CAP/CLIA), and time (continuous embedding and adaptation).

**Human-AI teaming, automation bias, trust calibration.** Rosbach et al. (arXiv 2411.00998, 2024, "Automation Bias in AI-Assisted Medical Decision-Making under Time Pressure in Computational Pathology"; updated/expanded as arXiv 2603.11821, 2026, "Stuck on Suggestions"): trained pathology experts (n=28) estimating tumor cell percentages with AI assistance showed a 7% automation-bias rate, "where initially correct evaluations were overturned" by incorrect AI advice; professional experience and self-efficacy were associated with reduced dependence on the system, while higher AI-assisted confidence was associated with *increased* automation reliance. Syrnioti A, Polónia A, Pinto J & Eloy C, "Human–machine interaction in computational cancer pathology" (*ESMO Real World Data and Digital Oncology* 2024, DOI 10.1016/j.esmorw\.2024.100062) argue that "the performance of the augmented pathologist that works in synergy with artificial intelligence (AI) is generally accepted as the most accurate in comparison to AI standing alone and the general pathologist standing alone" — i.e., the empirical sweet spot is hybrid, but achieving calibration is hard.

**Co-evolution / mutual shaping.** From STS / the social construction of technology (Bijker, Pinch). Pathology workflow and AI tools co-evolve: AI tools that demand specific scanning protocols change wet-lab practice; pathologist annotation practices change how AI is trained. The pathology informatics infrastructure (LIS, image management, DICOM-Path) is co-evolving with the AI ecosystem.

**Liability and medico-legal theory.** Calderaro J, Morement H, Penault-Llorca F, Gilbert S & Kather JN (*J Pathol* 2025, "The case for homebrew AI in diagnostic pathology") propose the lab-developed-test / "homebrew" pathway as a way around FDA/EU IVDR bottlenecks: "AI methods, if marketed and sold, require authorisation or clearance as in vitro diagnostic (IVD) devices by regulatory bodies like the Food and Drug Administration (FDA) in the USA or Notified Bodies in the European Union (EU)... These regulatory requirements create significant barriers to bringing AI solutions into clinical practice." Cestonaro et al. (*Front Med* 2023) review medical liability for AI-assisted diagnosis generally, finding the regulatory framework "inadequate and requires urgent intervention."

***

### Other notable theoretical lenses

* **Free-energy principle / predictive coding (Friston)** — neurocomputational accounts of perception that frame diagnosis as Bayesian inference under a hierarchical generative model. Largely a neuroscience frame, but increasingly imported into accounts of expert perception including pathology.
* **Knowledge encapsulation (Schmidt & Boshuizen)** — experts compile detailed biomedical knowledge into compact, ready-to-use clinical knowledge structures. Explains why experts diagnose fast but cannot articulate why.
* **Sensemaking (Weick)** — useful for analyzing how pathologists handle ambiguous cases at MDTs.
* **Total Quality Management / Lean / Six Sigma** — operations-management frames adapted to lab medicine (e.g., the Henricks/CAP literature on lab quality).
* **Foucauldian biopower / medical gaze** — the pathologist as paradigm of Foucault's "clinical gaze" (*The Birth of the Clinic*, 1963); minimal direct pathology-journal use but rich in medical humanities.
* **Affordance theory (Gibson, Norman)** — the affordances of glass slides (random-access pan and zoom by hand) vs. WSI (mouse-mediated, latency-dependent) explain perceived productivity loss in digital workflow studies.

***

## Recommendations

1. **For describing the cognitive act of diagnosis, anchor on three established frames**: dual-process theory with illness scripts (Pelaccia, Custers, Norman), visual-search/eye-tracking expertise research (Crowley, Brunyé, Mello-Thoms), and cognitive-bias theory (Croskerry). These are peer-reviewed and pathology-specific.
2. **For describing pathology as a system, use NASSS as the top-line framework** and bring in CAS/attractor language (as Garratt does) and Abbott's jurisdictional theory as complementary lenses. NASSS is the most actionable; CAS is the most evocative; Abbott explains the political economy.
3. **For explaining AI non-adoption specifically, layer the frames**: NASSS to map domains of complexity; sociotechnical systems theory to insist that scanner + workflow + pathologist + organization must be co-designed; automation-bias literature (Rosbach) to set realistic expectations for the hybrid model; Calderaro/Kather's homebrew pathway to address the regulatory bottleneck.
4. **Three theoretical opportunities are under-exploited and could ground original contributions**: (a) Klein's recognition-primed decision making applied to frozen-section pathology; (b) Hollnagel's Safety-II applied to laboratory resilience; (c) embodied-cognition / extended-mind analyses of microscope vs. WSI work.
5. **Benchmarks that would change these recommendations**:
   * Publication of a large multi-site RCT showing AI-augmented sign-out improves diagnostic accuracy and throughput → strengthens disruptive-innovation/translational case.
   * Emergence of additional pathology-specific NASSS case studies beyond Betmouni → would solidify NASSS as the field's default frame.
   * FDA authorization counts: if the histopathology AI tally moves from 3 (Aug 2025) to >20, the translational-gap framing weakens.
   * Empirical demonstration of pathologist deskilling under prolonged AI use → would shift the human-AI teaming discussion sharply toward Safety-II and embodied-cognition frames.

***

## Caveats

* The "H\&E as attractor" framing originated by Garratt is a public LinkedIn comment (not peer-reviewed); it is a useful conceptual hook but should be cited as commentary, not empirical finding. Garratt himself does not explicitly claim attractor dynamics *prevent* AI adoption — that inference is the analyst's.
* Several frameworks here (chaos theory applied to practice itself, embodied cognition, Foucault's clinical gaze) are applied to pathology mostly *metaphorically*; the report flags these distinctly from frames with empirical pathology literature (eye-tracking, NPT, NASSS, dual-process).
* Dual-process theory has been critically revised (Norman et al. 2024 argue that knowledge, not process, drives accuracy); readers should not treat System 1/System 2 as a fully validated dichotomy.
* The Christensen disruptive-innovation construct is widely criticized as post-hoc and imprecise; treat invocations in vendor literature with appropriate skepticism.
* Eye-tracking expertise studies in pathology have historically been modest in sample size (7–40 pathologists), though the PathoGaze1.0 dataset (Thai et al. 2025, n=19 pathologists, 397 WSIs, 171,909 fixations) marks a step-change.
* The NASSS framework, while explicitly recommended for pathology by Betmouni 2021, has had limited explicit uptake in pathology journals compared to radiology — there is a literature gap, not a deep evidence base.
* The Rosbach automation-bias finding (n=28, web-based experiment) is a first empirical signal, not a definitive estimate; replication across institutions and task types is needed.