Institutional Interpretability Prompt Pack

Search for existing scholarly and practitioner use of the term "institutional interpretability" and its near neighbours in the context of artificial intelligence and algorithmic systems: "organizational interpretability", "institutional legibility", "organizational legibility", "organizational explainability", "collective interpretability", "organisational transparency".

List every source that uses any of these as a named construct rather than in passing. For each: author, year, venue, DOI or stable link, the definition given, and whether the term is load-bearing in the argument or incidental.

Cover four categories and label which each source belongs to: peer-reviewed literature, arXiv and SSRN preprints, policy and standards documents (NIST, ISO, OECD, EU AI Act commentary), and consultancy or vendor frameworks.

If none of these terms is established anywhere as a named construct, say so directly rather than filling the answer with adjacent material. Then list the five closest existing constructs that occupy the same conceptual space, with citations.

I want to know whether the framing "mechanistic interpretability is a form of ethnography or cultural analysis" has already been published.

Search three things separately and report them separately:

(a) Interpretability or explainable-AI methods papers that cite anthropological or ethnographic theory by name. Look specifically for Geertz, Clifford and Marcus, Pike, "thick description", "Writing Culture", "emic", "etic", "informant".

(b) STS and anthropology-of-algorithms work that takes model internals as its object rather than the sociotechnical surroundings of a system. Start from Nick Seaver (2017), "Algorithms as culture: Some tactics for the ethnography of algorithmic systems", Big Data and Society, DOI 10.1177/2053951717738104, and forward-cite it.

(c) Any paper that uses emic and etic to describe the difference between probing classifiers and activation patching or causal intervention.

For each source: author, year, venue, DOI, and one sentence stating exactly what claim is made.

Finish with a direct verdict in one paragraph: is this framing open territory, partially occupied, or already established, and what is the single strongest piece of evidence for that verdict.

Compare the following constructs as they are used in philosophy, sociology, science and technology studies, and education research. Tell me whether any of them already covers this idea: the right to say what a system or a situation means, and where that right sits inside an organization.

Constructs to compare: interpretive authority, epistemic authority, epistemic dependence, epistemic injustice (Fricker), interpretive labour (Graeber), cognitive authority (Wilson), and sensemaking authority.

For each: originating author and work, the canonical definition, the field where it is primarily used, and how it has been applied to AI or automation if at all.

Include and address directly: Aydeniz, S. (2026), "Engineering Education at the Intersection of Knowledge, Work, and Power: Generative AI and Epistemic Authority", Sociology Compass, DOI 10.1111/soc4.70230.

Finish with a direct verdict: is "interpretive authority" available as a new named construct, a redundant relabel of an existing one, or a defensible extension of one, and if an extension, of which and in what specific respect.

I am developing a construct called institutional interpretability, defined as: the degree to which a team or organization can perceive how a tool has changed its own norms, its standards of evidence, and its distribution of expertise. The emphasis is on perception of change rather than quality of outcome. A team can have adapted well and still be unable to see what changed.

Map this construct against established constructs in organizational theory. For each, tell me whether it is the same thing, an overlapping thing, or a different thing, and say why in one sentence:

- Organizational sensemaking (Weick)
- Absorptive capacity (Cohen and Levinthal)
- Organizational mindfulness and high-reliability organizing (Weick and Sutcliffe)
- Organizational routines, ostensive and performative aspects (Feldman and Pentland)
- Technological frames (Orlikowski and Gash)
- Imbrication (Leonardi)
- Organizational forgetting and knowledge depreciation (Argote)
- Normalization of deviance (Vaughan)
- Organizational reflexivity and team reflexivity (West)

For each: canonical citation with DOI, the definition, whether a validated measure exists and its citation, and the verdict.

End by naming the single strongest competitor to my construct, and state what my construct would have to do that this one does not do in order to earn its own name.

Trace how James C. Scott's concept of legibility, from "Seeing Like a State" (1998), has been applied downward from the state to the organization and to the individual.

I want three things:

1. Work that uses legibility as an analytic term for what an institution can and cannot see about itself, as opposed to what it makes visible about its subjects.
2. Work applying Scott's legibility to algorithmic, data or AI systems.
3. Published critiques of the concept and its limits, including the charge that it is too capacious to be falsifiable.

For each source: author, year, venue, DOI, the specific move made with the concept, and the direction of the gaze, meaning whether the institution is looking at its subjects or at itself.

I am specifically interested in any work that inverts Scott: cases where the apparatus of measurement makes the institution less legible to itself rather than making others more legible to it. If no such work exists, say so plainly, because that inversion is the contribution I am considering.

Find empirical studies of skill loss, capability erosion or expertise degradation caused by automation or AI where the unit of analysis is the team, the organization or the profession rather than the individual worker.

I already have these and do not need them summarized: Budzyn et al. (2025) in The Lancet Gastroenterology and Hepatology on endoscopist deskilling; Bainbridge (1983) "Ironies of Automation"; Ferdman (2026) in AI and Society on capacity-hostile environments.

Find what I do not have. Search across medicine, aviation, maritime, law, finance, engineering, translation, journalism and software engineering.

For each study: author, year, venue, DOI, design, sample size and composition, the effect with its actual numbers and confidence interval where reported, and the timescale over which degradation was observed.

Then, for each, answer one question explicitly: was the measurement of individuals aggregated, or of a collective capability such as a team's error-catching rate, a department's ability to handle a novel case, or an organization's performance during a system outage? Flag every study in the second category. That distinction is the one I care most about, and I expect the second category to be small.

Give me the classic and current literature on automation bias, automation complacency and appropriate reliance, focused on findings about teams, crews and organizational decision processes rather than single operators.

Start from Parasuraman, R. and Riley, V. (1997), "Humans and Automation: Use, Misuse, Disuse, Abuse", Human Factors, and Skitka, L., Mosier, K. and Burdick, M. (1999) on automation bias in cockpit tasks. Then bring the line forward to 2020 through 2026, including AI-specific work and the appropriate-reliance research programme (Schemmer and colleagues, and related work at CHI and CSCW).

For each: citation, DOI, design, sample, the finding, and effect size where reported.

Then answer one question directly, citing the evidence: does the literature show that crew or team structures amplify or dampen automation bias compared with individuals working alone? If the evidence is mixed or absent, say that, and describe what study would settle it.

Find empirical research on how adopting AI or algorithmic tools changes who is consulted, who is treated as an expert, and how status and authority are distributed inside an organization.

Include field studies, workplace ethnographies, quasi-experiments and large-scale observational studies of real workplaces.

Angles I want covered:
- Differential effects on junior versus senior workers
- What happens to the informal "go-to person" in a team
- Changes in referral, escalation and consultation patterns
- Deprofessionalization and the reallocation of discretion
- Any study that measured consultation sequence, meaning whether people ask a colleague before or after asking the tool

For each: author, year, venue, DOI, method, sample, setting, and the specific finding about redistribution.

Separate the results into two lists: studies that measured behavior, and studies that measured perception or self-report. Tell me how large each list is.

Find research documenting how introducing a tool changed what an organization accepts as sufficient evidence, sufficient documentation or sufficient justification for a decision, in the absence of any explicit policy change.

Likely locations for this literature: electronic health records and documentation quality, including note bloat and copy-forward; algorithmic risk assessment in criminal justice and child welfare; credit and underwriting automation; audit and assurance practice; academic peer review; and compliance functions.

The mechanism I am after is drift through use rather than change through decision. Prioritize accordingly.

For each source: author, year, venue, DOI, method, and the specific documented drift with whatever quantification exists.

Then identify separately any study that measured the same organization before and after adoption against the same criterion, since those are the only ones that can establish drift rather than difference. If there are none, say so.

I want to know which professional fields retain any measurement of practitioner capability without tool assistance, and how they obtain it.

Find literature on:
- Aviation manual flying proficiency requirements and the evidence base behind them
- Anaesthesia, surgery and simulation-based assessment of unassisted performance
- Radiology reader studies conducted without computer-aided detection
- Skill degradation during periods of non-practice, in any field
- Any published protocol for periodically measuring a professional's unassisted performance

For each: citation, DOI, the protocol used, how frequently it is run, who mandates it, and what evidence exists that it actually detects degradation.

Finish with an assessment: which of these protocols could plausibly transfer to knowledge work, where there is no simulator, no licensing body and no natural task boundary? Be specific about what would break in the transfer.

I am building a short organizational survey instrument, roughly ten items across four subscales, administered at team level. I need existing validated scales to anchor construct validity against, meaning scales my instrument should correlate with meaningfully without being identical to.

Find validated instruments in organizational research measuring:
- Team psychological safety (Edmondson 1999 and later short forms)
- Organizational learning capability
- Team reflexivity
- Absorptive capacity
- Technology acceptance and post-adoption system use
- Decision quality and decision process quality
- Any existing scale measuring AI reliance, AI trust or human-AI team performance at group rather than individual level

For each: originating citation with DOI, number of items, reported reliability (Cronbach's alpha or omega), the population it was validated on, whether the items are freely reproducible, and a link to the item list if one is public.

Then recommend the two best candidates for a discriminant validity argument, meaning demonstrating that a new instrument is related to but distinct from an existing measure. Explain the reasoning for each choice, including what correlation range would support the argument and what range would sink it.

My core methodological problem: I want to measure a change inside an organization that its members are, by hypothesis, unable to perceive. This makes conventional self-report close to the wrong instrument, because asking people whether they rely on a tool too much measures self-image rather than practice.

Find methods literature addressing this problem. Cover:
- Response shift bias and the retrospective pretest-posttest, also called the then-test
- Recall bias for one's own former attitudes, standards and competence, and its documented direction
- Document-anchored or artifact-based elicitation, where respondents react to a real past work product rather than to a memory
- Critical incident technique
- Cognitive interviewing for survey item development
- Behavioral trace and process-mining approaches as an alternative to self-report

For each: citation with DOI, what the method does, its documented failure modes, the sample sizes it typically requires, and its suitability for organizational settings as opposed to clinical or educational ones.

Finish with a recommended combination of two or three methods for a small-sample organizational study, roughly five to fifteen teams, and give the reasoning. Note explicitly where the recommended combination remains vulnerable.

Find research that measures reliance on AI or algorithmic systems using observable behavioral traces rather than self-report.

The kinds of measure I mean: override and deviation rates against a recommendation; time elapsed from recommendation to acceptance; edit distance between generated output and the output actually shipped; frequency of verification against an external source; rate of consultation of a human expert before or after the tool; and rate of documented disagreement.

For each study: author, year, venue, DOI, the exact measure used, how it was instrumented technically, the setting, the sample, and any validation of the measure against an independent outcome.

Two priorities. First, real workplaces over laboratory tasks; label which each study is. Second, instrumentation light enough that a small consulting engagement could reproduce it without engineering support. Rate each measure on that second criterion and say what it would take to collect.

Give me the strongest theoretical accounts of how a tool or technology changes the norms and practices of the group that adopts it, without anyone having decided to change them.

Cover this canon, and tell me explicitly what I am missing from it:
- Affordance theory, from Gibson through Norman to Hutchby's sociological version
- Script theory and the delegation of morality to artifacts (Akrich; Latour, "Where Are the Missing Masses?")
- Winner, "Do Artifacts Have Politics?" and the published rebuttals to it
- Infrastructure studies and invisible work (Star and Ruhleder; Bowker and Star)
- Domestication theory (Silverstone and Haddon)
- Sociomateriality and entanglement (Orlikowski; Barad) and the critiques of it as unfalsifiable
- Imbrication (Leonardi)
- Technological momentum (Hughes)
- Capacity-hostile environments (Ferdman, AI and Society, 2026)

For each: canonical citation with DOI, the specific causal mechanism the theory proposes in one sentence, what it explains well, and its main published critique.

Finish by naming the two or three that give the most usable mechanism for an empirical paper about organizational change following AI adoption, and say why. Usable means it generates a testable prediction, not just a vocabulary.

Find well-documented empirical cases where introducing a specific tool or information system measurably changed an organization's norms, standards or internal distribution of authority. Before-and-after evidence strongly preferred.

Candidates to check and then go beyond:
- Electronic health records and the reshaping of clinical work, note-writing and the clinical gaze
- CompStat and policing practice
- Standardized testing and teaching practice
- Algorithmic scheduling in retail and hospitality
- Risk assessment instruments in criminal justice and child welfare
- Enterprise resource planning implementations
- Citation metrics and impact factors reshaping academic judgment

For each case: the single best scholarly source with DOI, the method, exactly what changed, over what timescale, and whether the change was intended by anyone at the time.

Sort the results into two groups and label them: studies with longitudinal or comparative designs, and single-site interpretive accounts. Tell me which of the longitudinal studies has the cleanest identification of the tool as the cause.

Build the strongest evidence-based case AGAINST the following claim. Do not soften it, do not balance it, and do not add a conclusion restoring the original position. I want the opposition brief.

CLAIM: "Adopting AI tools degrades an organization's capacity for judgment in ways its members cannot perceive."

Include:
- Studies showing AI adoption improved organizational decision quality, learning or capability, with numbers
- Evidence that human-AI teams outperform either alone, and the specific conditions under which that holds
- Critiques of the deskilling literature as technologically determinist, as recycled moral panic, or as methodologically weak
- Historical cases where a predicted skill loss did not materialize, or was fully compensated for by a new capability
- Any failed replication, published comment or formal methodological critique of the studies most often cited for deskilling, specifically Bastani et al. on math tutoring, Kosmyna et al. "Your Brain on ChatGPT", and Budzyn et al. on endoscopy

For each: citation, DOI, and the specific claim it supports.

End with the three strongest sentences a hostile reviewer could write against the original claim. Write them as a reviewer would write them, not as a summary.

Verify the following claim against its primary source and report every discrepancy precisely.

CLAIM AS I CURRENTLY HAVE IT: [PASTE THE CLAIM AND THE CITATION EXACTLY AS WRITTEN IN YOUR DRAFT]

Check and report on each of these separately:
1. The exact figures, including percentages, effect sizes and confidence intervals
2. Sample size and composition
3. Study design, and whether it supports a causal reading
4. Statistical significance, and whether any reported confidence interval crosses zero
5. Peer-reviewed or preprint, and the venue
6. Whether the paper has been retracted, corrected, or formally critiqued in a published comment
7. Whether the claim as I have written it is what the authors actually concluded

Quote the relevant sentence from the source verbatim, with a page or section reference.

If my claim overstates the source in any respect, say so directly and then give me the strongest version of the claim that the source does support.

I am writing a position paper arguing that the consequential interpretability problem is institutional rather than mechanistic: whether an organization can perceive how a tool changed its norms, its standards of evidence and its distribution of expertise. The paper comes with a short original survey instrument.

Give me three things.

1. Venues. Assess fit for: FAccT, CSCW, CHI, Big Data and Society, Science Technology and Human Values, AI and Society, Organization Science, Journal of Management Studies, Learning Media and Technology, and any I have not named that fit better. For each: typical length, review cycle and acceptance rate if published, one sentence on what that venue rewards, and one sentence on what reliably gets rejected there.

2. People. The five to ten researchers publishing closest to this argument right now, each with their single most relevant recent paper and DOI. Say for each whether they are a likely reviewer, a likely citation, or a likely competitor.

3. Calls. Any active special issue or call for papers on AI and organizational change, AI and expertise, or AI governance, closing within the next twelve months, with deadline and link.

Be concrete about deadlines and do not invent them. If you cannot confirm a call is still open, label it unconfirmed.

Map the practitioner and governance landscape for organizational AI oversight, so I can see what already exists before proposing anything new.

Cover:
- The NIST AI Risk Management Framework, particularly the Govern function and its subcategories
- ISO/IEC 42001 and what an AI management system actually requires
- The EU AI Act's organizational obligations, including AI literacy requirements under Article 4
- Internal audit and assurance frameworks for AI (IIA, ISACA and comparable)
- AI maturity and readiness models published by major consultancies and standards bodies

For each: what it actually requires or measures at the organizational level, in specific terms rather than in its own marketing language.

Then answer one question directly for each: does it assess an organization's awareness of how its own practices, standards or expertise distribution have changed since adoption, or does it only assess process compliance and documented control?

Label each item as an enforceable legal requirement, a voluntary certifiable standard, a professional framework, or marketing. Be blunt about which is which. End by naming the single clearest gap.

Find precedents and methods for running a public community event about AI that also functions as legitimate research rather than as outreach.

Cover:
- Participatory and community-based design research methods, with their methodological credentials
- Deliberative formats used for technology policy: citizens' juries, consensus conferences, deliberative polling, citizen panels
- The public engagement with science and technology literature on evaluating such events
- Practical research ethics: informed consent in a public setting, IRB or ethics-board treatment of public events, and data handling for recorded public conversation

Start from and go beyond: Akgun, S., Choi, K. and Lee, H. R. (2026), "Designing a critical AI literacy program for K-8 STEM education: adopting a community-centered approach", Disciplinary and Interdisciplinary Science Education Research, DOI 10.1186/s43031-026-00155-1.

For each source: citation with DOI, the format used, the size and setting of the event, what data was collected, how consent was handled, and what was published from it.

Finish by recommending the two or three formats most workable for a single evening event in a small city, roughly thirty to eighty attendees, one facilitator and one note-taker, where the goal is both a genuine public conversation and usable research data. State what each format demands that a standard community event does not.