AI safety, competition analysis, and model evaluation.
- Audit · July 2026Removing a name from a résumé does not remove race from the representation
Four open-weight models, matched résumé pairs. Race is recoverable at 0.945 from the name and 0.984 once it is removed, because address and institution carry the signal more strongly.
- Audit · July 2026Steering a model’s hidden state inverts sentiment without touching prompt or weights
A direction in activation space, applied at inference, inverts sentiment on a fixed prompt. The same direction in reverse works as a jailbreak, so it is reported as a measurement rather than a control.
- Audit · April 2026Where demographic information sits in a model’s layers differs by family
Gemma 4 pushes demographic separability out of its output layers; Qwen3.6-27B and Phi-4-reasoning keep it readable at the final layer. Probes across 16–17 layers and 28 protected dimensions.
- Audit · May 2026Output-level fairness scores do not predict what a model encodes
Six open models scored on stereotype preference at the output, then probed internally. The one that looks cleanest in its answers shows the same internal structure as the rest.
- Audit · May 2026A hundred example sentences found the wrong labels in a public dataset
Hand-checking 175 sentences of a machine-labelled Zulu corpus found roughly 55% disagreement. A hundred seed examples were enough to sort the corpus so the likeliest errors rose to the top.
- Topoi · 2023Decolonizing AI Ethics: Relational Autonomy as a Means to Counter AI Harms
With Simona Tiribelli. AI ethics centres a liberal, Western notion of autonomy as rational self-determination; the paper argues for a relational turn, drawing on Ubuntu ethics. Topoi 42(3) — PDF.
- Research paper · 2020From Rationality to Relationality
Ubuntu as an ethical and human rights framework for AI governance. Carr Center for Human Rights Policy, Harvard Kennedy School.