Overview
Most reviews of university AI governance ask whether a policy exists and what
it says. This project asks a harder question: when AI use moves beyond the
form the policy-writers had in mind, does the university’s documentation still
connect the situation to someone with the authority, criteria, safeguards and
review process needed to handle it responsibly? I built a method to test that,
scenario by scenario, against real institutional documentation — and applied
it across a set of Western Australian universities.
Problem
Universities have produced AI principles and policies at speed. But a
published principle can be defensible on its own terms while still leaving
basic governance questions unanswered for an unfamiliar case: who is allowed
to decide, what criteria they should use, what process an affected person can
use, and where review or escalation sits. Those gaps matter more as AI moves
from bounded, visible assistance towards more autonomous or delegated forms of
use. Reading policies for content cannot surface this; the documentation has
to be tested against realistic pressure.
My role
Sole researcher, working under the supervision of Dr Jon Mason at Charles
Darwin University: designed the method, built the documentary corpus and
scenario set, carried out the coding, and led the analysis.
Research question
As the type of AI use in a scenario becomes less familiar and more autonomous,
does a university’s published governance still connect it to an accountable,
sufficiently specific institutional response — and where does that connection
break down?
Method
I collected the complete set of publicly available governance-relevant
documents for each university in the study — policies, principles, procedures
and guidance, not just the primary academic-integrity policy — and built a
fixed, quality-screened set of scenarios spanning a deliberate range of AI
involvement, from routine, bounded assistance through to more autonomous,
delegated forms of action. Each scenario was applied to each university’s
documentation using a structured response chain that checks, in order,
whether authority, procedure, decision criteria, safeguards and review remain
connected as the scenario gets harder. A non-compensatory decision rule means
a strong showing on one dimension cannot cover for a missing one — a
university either has a workable pathway for a given scenario, or it does not.
Findings
The analysis is complete and reported in a manuscript currently under peer
review, so I am not publishing specific figures here yet — the exact results
belong in the published version, not a preview of it. In general terms:
governance held together reasonably well for the most familiar, bounded forms
of AI assistance, and became markedly less reliable as scenarios moved toward
more autonomous or delegated AI action. The most common weak point was not a
complete absence of governance — most universities named someone responsible
even for unfamiliar cases — but a lack of sufficiently specific criteria,
safeguards or review attached to that role. Full results, tables and the
coding protocol will be added to this page once the manuscript completes
review.
Outputs
- A manuscript reporting the full method and results — currently under peer
review. A preprint of the method is available now on arXiv (see Related
publication below); the peer-reviewed version and full results will be
linked here once review completes.
- The scenario-design and coding protocol underlying this study also informs
the fictional Case Portfolio published elsewhere on this site.
Impact
Too early to report a verified, external impact — the manuscript has not yet
completed review. The clearest impact so far is internal: the method now
structures how I evaluate governance documentation across my other projects.
Limitations
- Decisions are made from published documentation alone; real institutional
decision-making also draws on unwritten precedent, discretion and context
the method cannot observe.
- The scenario set is deliberately focused and cannot cover the full space of
possible AI use.
- A second coder checked a portion of the classifications for reliability, but
the bulk of the coding was carried out by a single analyst; the protocol is
documented so the study can be independently replicated.
Lessons learned
- The hard part was never detecting whether a governance document existed —
it was tracing whether authority, criteria, safeguards and review actually
stayed connected to each other as a scenario got less familiar.
- Forcing every judgement to cite the specific clause it relied on turned out
to be the single most useful design decision: it makes “the policy doesn’t
actually decide this” a visible, countable finding rather than something
that gets quietly smoothed over.
A preprint reporting the method is available on arXiv (arXiv:2608.28925). The
peer-reviewed journal submission is still under review, so the figures and
university-by-university breakdown above are held back until that process
completes — see the Publications page for the preprint.
Downloadable materials
Not yet published while the manuscript is under review. The coding protocol
will be made available here after publication.