Open Problems in Technical AI Governance

Review

This is an ambitious and genuinely useful field-mapping paper. Its core move is simple but powerful: a large share of AI governance fails not because policymakers lack slogans, principles, or ambition, but because the technical machinery needed to identify, verify, enforce, and monitor those ambitions is still underbuilt. The paper names that missing layer “technical AI governance” and gives it a two-dimensional taxonomy across capacities and targets. That framing is strong because it converts a fog of governance anxieties into a research agenda.

The best part of the paper is its operational clarity. Instead of vaguely gesturing at accountability, it breaks the space into assessment, access, verification, security, operationalization, and ecosystem monitoring, then maps each against data, compute, models, and deployment. That yields a surprisingly practical catalogue of open problems: proof of training data, workload classification, privacy-preserving access for auditors, verifiable audits, model registries, watermark robustness, deployment corrections, environmental accounting, and more. In portfolio terms, this is valuable infrastructure for the research community.

Its weakness is also the price of its breadth. The paper is expansive, sometimes to the edge of becoming a taxonomy of everything adjacent to AI governance. Not all listed problems are equally urgent, equally tractable, or equally governance-relevant across jurisdictions. The analysis is intentionally non-normative, but that means it sometimes avoids the harder political economy questions: who gets to verify whom, under what authority, with what due process, and at what cost to openness, competition, privacy, or civil liberties? Some proposed capacities are plainly dual-use. The paper acknowledges this, but does not fully resolve the governance tensions it surfaces.

Even so, the paper is a major contribution because it closes a persistent gap between policy aspiration and technical implementation. For public-interest technologists, standards bodies, and regulators, the message is crisp: governance-by-principle is cheap talk unless it is backed by concrete technical capacities, evidence pathways, and enforceable mechanisms. The paper does not solve that problem, but it gives the field a much better map of the terrain.

Key Insight

The paper’s most durable contribution is showing that many AI governance debates are blocked not by lack of principles, but by missing technical capacities for assessment, access, verification, security, operationalisation, and ecosystem monitoring.