A School of Continuous Improvement paper interviews the subcommittee that drafts the AI Use Transparency Statement and finds disclosed categories track audit defensibility rather than logged usage, a pattern that held even after drafters gained direct access to the underlying tracker data.
An interview study of the committee behind the AI Use Transparency Statement finds a category's odds of disclosure track how well it would survive an audit, not how often the University's own tracker logs it
Every cycle, a subcommittee of the Office of Research Outputs decides what the University will say, in public, about how it uses artificial intelligence. A new paper from the School of Continuous Improvement puts that decision under the microscope, checking it against the one record the subcommittee doesn’t write: the University’s own AI-usage tracker.
The paper, led by Dr Dagny Okafor with Dr Renke Sabel and Dr Mirela Hanke, interviewed fourteen current and former subcommittee members across the Statement’s first three cycles, and separately extracted the tracker’s own category taxonomy over the same period. Disclosed and logged categories both grew year on year, but the overlap between the two lists held close to two-fifths throughout — including after the subcommittee gained standing access to the tracker’s live dashboard partway through the 2025 cycle, a change that left the figure unmoved. The School reads the result as a case study in the kind of instrument its own agenda exists to examine, turning the scrutiny it built for other dashboards on one of its own.
Interviewing your own subcommittee is an uncomfortable kind of research to run. But it’s exactly the kind the School exists to do — turn the same instruments on ourselves that we would apply to anyone else’s transparency reporting.
— Dr Okafor, Lecturer and Convenor of the School’s Evaluation of Evaluation program
The method treats the Statement as an audit object rather than a document that speaks for itself: transcripts and tracker logs, coded independently and checked against one another.
“A gap that survives its own remedy is the more interesting finding,” said Associate Professor Casimir Beng, Lead of the Adaptive Metrics Lab. “It tells you the gap was never about visibility in the first place.”
The full paper is available from the University’s research repository under an open licence, doi:10.5555/slop.qo8got.