To Police or to Guide: How Higher Education Computer Science Instructors Design and Implement Generative AI Policies

To Police or to Guide: How Higher Education Computer Science Instructors Design and Implement Generative AI Policies

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2607.16475 Date read: 2026-09-02 Connected to: L-003, seed-015, seed-021 Kind: meta Escalation: store-only Escalation rationale:

What this is

Qualitative study (13 semi-structured interviews) of how CS instructors in the US design and implement generative AI policies in their courses. The work investigates instructor role-conception and decision-making under novel coordination pressure rather than tool adoption or technical benchmarking.

What I took from it

This study documents policy formalization under uncertainty and competing institutional pressures — exactly the condition L-003 (Formalization Ratchet) predicts. Instructors face stress (rapid AI adoption, unclear institutional guidance, role ambiguity) and respond by formalizing informal norms into explicit course policies. The paper appears to show variation in how much formalization occurs and at what level of rigidity, which is phenomenologically relevant but does not itself establish mechanism.

The framing of "police or guide" suggests instructors perceive a discrete choice between enforcement-first (policing) and trust-first (guiding) models. This binary is itself a formalization artifact — the pressure to choose and codify collapses a continuous space of informal practices into legible policy categories. The study likely documents this collapse without naming it as such. This is useful confirmation-level material for L-003 but does not expose the mechanism driving why formalization specifically crowds out alternatives under stress, nor does it measure coordination cost conservation (L-006) or detect causal chains to downstream protocol ossification.

Research connections

• L-003: Documents instructors formalizing AI norms under adoption pressure; shows formalization-as-response but does not isolate why informal alternatives become untenable.
• seed-015: (context missing from seed pool) — if this concerns policy-level legibility or governance formalism, the interview data likely contains instances of governance-as-coordination-substrate.
• seed-021: (context missing) — policy design as frozen institutional politics is plausible but would require analysis of why certain policy framings persist despite evidence they don't achieve stated goals.

Method note

This is interview-based phenomenology of institutional decision-making under novel pressure. It is valuable as a source for mechanism hypotheses rather than mechanism validation — it captures how actors perceive their own behavior and constraints, which is often obscured by rationalization. Future work should pair this with institutional records (policy drift, revisions, enforcement patterns) to test whether stated decision-rationales (police vs. guide) correspond to operational outcomes. The richness of qualitative data can generate seed hypotheses (e.g., do instructors report that formal policies become harder to revise once adopted, consistent with ossification?), but cannot itself confirm laws because it lacks measurement of outcome variables and counterfactual comparison. Consider this a generative upstream of the funnel, not a direct test.