Independent applied AI research group
AIthena studies how machine learning behaves once it is installed inside a public institution, whether that is a benefits program, a clinic, or a policy model. We work on the failures an average score hides: who a system is quietly worse for, and whether anyone is in a position to notice.
A model can look equally accurate for two groups and still leave one of them measurably worse off. Switch the view.
Schematic of the result reported in Access Distortion in Economic World Models. Bars are illustrative, not measured values.
Research
Every one of our lines of work starts the same way: a system reports that it is doing fine, and the people inside it are not. We formalise the gap, then build the instrument that can see it.
Who a system actually reaches, as opposed to who it formally admits. We study access barriers, participation debt, and the distance between a policy on paper and the outcome it produces once it meets real constraints.
Fairness / access / policy evaluation
Evaluation frameworks have blind spots by construction. We prove what a given framework cannot see, and work out what it takes for a model's output to be auditable and justified in a regulated setting.
Evaluation limits / auditability / justification
High-stakes clinical AI, where an average score hides exactly the cases that matter: Alzheimer's pathology, MRI reconstruction failure modes, and the limits of aggregate evaluation in medical imaging.
Medical imaging / neurodegeneration / failure modes
Publications
Six papers across public policy, algorithmic fairness, trustworthy ML, and medical AI, all accepted for 2026.
Chicago Booth WM 2026
Formalises access distortion: a failure mode where a world model's aggregate accuracy looks acceptable while decision quality is systematically worse for people with weaker information access. Two access tiers can be indistinguishable on accuracy while differing roughly twofold in realised decision utility.
Brown University MLSJ 2026, Oral
Introduces participation debt: the gap between formally granted opportunities and opportunities a person can realistically reach. Shows that standard fairness metrics can be fully satisfied while participation debt grows, particularly across healthcare and public policy.
EIML @ ICML 2026
Proves that evaluation frameworks are bounded by construction: entire classes of failure are invisible by design when they fall outside the evaluation space. The result puts a hard limit on what any single evaluation regime can certify.
PhilML @ ICML 2026
Formalises the justification deficit: the measurable gap between predictive success and epistemic warrant. A model can be accurate for the wrong reasons, and standard metrics cannot detect the difference.
Stanford PAI 2026
A controlled evaluation framework grounded in k-space physics, showing that the structure of a corruption, not its severity, determines the qualitative character of the failure. Structured line dropout produces catastrophic collapse (normalised SSIM drop 0.877) with high cross-subject variance, a pattern aggregate PSNR cannot distinguish. Presented to over 400 students, researchers, and industry practitioners.
CVPR 2026, Subtle Visual Computing Workshop
Multimodal unsupervised methods for surfacing candidate structural-functional dissociation patterns in Alzheimer's MRI. Accepted to present at CVPR 2026; the preprint has been read over 600 times.
Team
A small group. We pick problems that are unglamorous and consequential, and we keep working them until there is something publishable and something usable.
Collaborate
If there is a real decision riding on a model and the usual evaluation is not telling you enough, write to us. We are open to research collaborations on any of the problems listed here.
What we take on