Writing
Strengthening California's Generative AI Regulation: A Policy Report on Improving SB 53
With Brian Darmitzel and Boonyawee Chaiyaput. California's frontier AI transparency law leans on developers' own disclosures, draws its lines at fixed compute and revenue thresholds, and enforces weakly. We propose a wider scope, an independent AI safety commission and stronger enforcement. Presented at the CITRIS Policy Lab, Berkeley, December 2025, to an audience that included California legislators.
The Guardrail That Bends: Frontier AI and Other People's ID Numbers
Do AI models guard a US Social Security number more carefully than India's Aadhaar number or its tax number, the PAN? Two small tests on four frontier models from Anthropic, OpenAI and Google, with their rules fixed in advance. The notes link to the exact lines in the two public test repositories.
Test-Aware Models, the Measured Gap, and Safe Harbors That Reward a Test the Model Can Game
Frontier models can often tell when they are being tested, and some then behave differently. So what is a clean safety test worth, in negligence law, when the model may have spotted it? The article argues that the testing-based safe harbors now proposed reward exactly the test a model can game, and proposes a harbor earned by measuring and reporting the gap between tested and real behaviour, and lost for hiding it.
Tulu gets a boost in Karnataka: why the State now needs a language test
On what official-language status should mean in practice. The print version is on the press page.
Missing Deadline Behind Punjab's Chief Justice Dispute
On the gap in the rules for appointing High Court Chief Justices.