AI safety, competition analysis, and model evaluation.

Four open-weight models, matched résumé pairs. Race is recoverable at 0.945 from the name; removing it raises detectability to 0.984, because address and institution carry the signal more strongly than the name does. Protocol, per-model results and the data are in the post.

A direction in activation space, applied at inference, moves a review from positive to negative on a fixed prompt. Scored with third-party classifiers. The same direction applied in reverse functions as a jailbreak, which is why it is reported as a measurement rather than a control.

Gemma 4 pushes demographic separability out of its output layers; Qwen3.6-27B and Phi-4-reasoning keep it linearly readable at the final layer across all seven BBQ axes. Probes at 16–17 layers, 28 protected dimensions, 10,064 sentence pairs, five-fold cross-validation.

With Simona Tiribelli. AI ethics frameworks centre autonomy as the safeguard against harm, but that principle is itself a liberal, Western one — rational self-determination. It misses the ways autonomy is violated, and the harms tied to the legacy of colonization that fall hardest on the already marginalized. The paper argues for a relational turn, drawing on relational autonomy in moral philosophy and in Ubuntu ethics. Topoi 42(3), 867–880. Open access — PDF.

Ubuntu as an ethical and human rights framework for AI governance. Published by the Carr Center for Human Rights Policy, Harvard Kennedy School. 147 footnotes.