From Policing Provenance to Designing for Evidence: Testing Four Propositions on Generative AI Against Research and Institutional Policy

Synthesis: Ansari tests four propositions about generative AI and assessment against two coded corpora — 41 research and sector sources and 55 university and sector policy documents — and finds they hold unevenly. Redesign over detection is well supported, but as redesign plus assurance: 4 of 55 documents prohibit AI by default, and of the 23 with a detection position, 16 reject detectors and one requires them. Disclosure is the proposition the evidence complicates most, because institutions value declarations while students demonstrably do not complete them. From that reading comes the paper's contribution, a nine-step decision chain in which the role assigned to AI is the seventh question rather than the first, following from the evidence an outcome requires.

Key Findings

1. Redesign beats detection, in policy as well as argument. Only 4 of 55 documents prohibit AI by default, and of the 23 that take a position on detection, 16 reject detectors while one requires them.
2. Graded permission is the emerging form. 31 of 55 documents define named categories of use rather than a binary allow/prohibit rule, which matches the permission scales the assessment literature recommends.
3. Disclosure is valued by institutions and poorly observed by students. Policies treat declarations as valuable, but the limited available evidence suggests many students do not complete required disclosures — so the proposition that disclosure yields low-value information is partly vindicated, for a different reason than the one given.
4. Testing unaided ability is supported as one lane, not a return to invigilation. The paper accepts controlled conditions for outcomes that require unaided performance while rejecting a program built entirely around them.
5. The peer-review extension is plausible but under-examined. The corpus supports judging a review by its quality rather than by how much a human typed, on very little direct evidence.
6. The design procedure is a nine-step chain, with AI's role seventh. It asks for the intended outcome, whether AI can already do that work, whether unaided demonstration is required, whether AI use is itself the capability, what evidence of thinking is needed, and only then what role AI should have.
7. Friction and assurance are different instruments. Requiring local data, reflections or in-class material raises the cost of an AI answer and decays as models improve; supervised conditions, vivas and program-level validity depend far less on current model limits.

Why validity, not provenance

The paper's framing is that provenance questions — did a human type this? — can no longer be answered, so an assessment procedure built around them certifies nothing. What a mark should certify is that the student performed the intended intellectual work, and if a universally available tool can produce the artifact, the assessment has to be redesigned until the evidence it collects still supports that claim.

That reframing has a practical consequence the paper draws out: a no-AI rule enforced by friction is weak not only because it can be broken but because marks are then read on the false assumption that it was kept. The alternative is to design tasks where the evidence of thinking is collected by construction — process artifacts, oral verification, supervised performance.

The assurance distinction

The most useful analytic move is the split between friction and assurance. Friction is a design choice that makes an AI-produced answer worse or costlier: personal data, local context, reflection on one's own error, a problem the student poses rather than answers. It works against today's models and decays as they improve, so a course built on friction alone is running a race it will lose.

Assurance is a choice that collects evidence of the student's reasoning regardless of what the tool can do — an unpredictable viva mapped to outcomes, an observed practical, supervised conditions, or program-level design in which a capability is demonstrated across several assessments. The paper argues these are more durable and explicitly not permanent: validity is a claim argued from evidence, and assurance methods invite their own workarounds once they become standard.

What the corpora can and cannot show

The evidence base is two purposive samples, coded by a single coder, with a policy corpus weighted toward the United Kingdom. Counts like 4 of 55 are descriptions of that sample, not of the sector, and the author says so at the outset. The literature corpus includes six methodology references retained in the count but not coded on the propositions.

The method is a pragmatic exploratory review rather than a systematic one, with no pre-registered protocol or screening tally. Its value is in the design framework the reading produced and in the shape of the policy landscape, not in prevalence estimates.

What this means for practice

• Start from the evidence the outcome needs, then decide AI's role. The chain's ordering is the actionable part: naming a permission level first locks in a design that may not collect evidence of the intended learning.
• Prefer assurance to friction where the outcome is load-bearing. A viva or supervised component survives model improvement; a "use your own data" requirement does not.
• Do not build assessment around disclosure. Policies value declarations, but students do not reliably complete them, so a grade that depends on an accurate declaration rests on an unevidenced assumption.
• Treat peer review the same way and say so. If a review is judged on quality, the policy should not imply that human authorship of every sentence is the criterion.

Limitations

• Two purposive corpora, single coder. 41 literature sources and 55 policy documents, coded by one person, with no second coder or inter-rater check reported.
• UK-weighted policy sample. Counts describe the sample rather than the sector, and the author states this bound explicitly.
• Pragmatic, not systematic, review. No pre-registered protocol or logged screening tally; sources were gathered by targeted searching and citation chasing.
• The anchoring position is an opinion piece. The four propositions are restated from a non-peer-reviewed post and tested as claims, which shapes what counts as support.
• Capability claims date quickly. The step that asks whether AI can already perform the intended work must be re-run as models change, which the paper acknowledges.

Citation

Ansari, S. (2026). From Policing Provenance to Designing for Evidence: Testing Four Propositions on Generative AI Against Research and Institutional Policy. arXiv:2611.01330 [cs.CY].