Three focus areas
As frontier AI systems become more capable, governance
depends increasingly on technical facts and infrastructure: what systems
can do, how much compute they use, and whether claims about them can be independently verified.
Laws are hard to enforce if governments cannot measure capabilities or
monitor deployed systems. Technical AI governance asks how that substrate
can make governance more informed, precise and enforceable.
Each answers one of the three causes, and is briefed against it below.
Evidence exists — but isn’t shared
Developers hold the training data, the internal evaluations and the usage telemetry. What reaches policymakers and independent researchers is self-reported and unverifiable, and the cost of frontier-scale compute prices replication out of reach.
Third-party access & risk modelling
Third-party evaluators work through black-box APIs, on roughly a week’s notice, and rarely get a reliable upper bound on what a model can do. The map from a benchmark score to a real-world risk pathway stays largely undrawn — which is precisely the map policymakers need in order to write a guardrail that binds.
Read the brief
What blocks it
Developers hold the training data, internal evaluations and usage telemetry. What reaches the outside world is self-reported and unverifiable. Model scorecards restate the problem rather than solve it. Frontier-scale compute prices independent replication out of reach.
What we want in the room
Evaluators who can say what structured access would actually let them measure; policy teams who can say what finding would change a decision; and the people who have to translate one into the other.
Questions on the table
What does a defensible upper bound on capability look like? What evaluation window is long enough for long-horizon tasks? What would third-party access look like if it were designed for statistical power rather than reputational cover?
- Background reading
- Frontier AI auditing: toward rigorous third-party assessment ↗
Evidence exists — but can’t be verified
Even well-founded policy stalls at enforcement. States have commercial and strategic reasons to move fast, and no mechanism to confirm that a counterpart is doing what it says it is doing.
Hardware verification, scalable oversight & international coordination
An agreement is only as good as the ability to check it. Tamper-resistant hardware that keeps a trustworthy record of what a chip was used for — how much compute ran, whether an evaluation took place — turns unverifiable self-reporting into claims a counterparty can test.
Read the brief
What blocks it
States face commercial and strategic incentives to move quickly, and have no way to confirm that a counterpart is doing what it claims. Coordination fails not because the policy is wrong but because compliance is uncheckable.
Why hardware
A chip that can produce an unforgeable certificate about its own usage makes a claim auditable without exposing model weights or commercial secrets. That is the rare mechanism that survives adversarial incentives. ERA is building hardware security as a sub-field of AI safety, and this track is where that work meets the people who would have to enforce it.
Questions on the table
What does technical compliance actually look like on the ground? Where does scalable oversight substitute for verification, and where can it not? Which claims are worth verifying first?
- Background reading
- Governing through the cloud ↗
- Faster AI diffusion through hardware-based verification ↗
- Recommended directions for scalable oversight ↗
Evidence doesn’t exist yet
Safety teams get weeks to design policy they could plausibly implement. Evaluators get days to test. Long-horizon and multi-agent capability has no settled measurement — so some risks are simply unmapped.
Measuring its effects on AI progress and oversight
As models take on more of the research loop, the pace of capability gain and the difficulty of overseeing it move together. This track asks what can actually be measured about automated AI R&D — before the measurement problem outruns the measurers.
Read the brief
What blocks it
Models are beginning to distinguish evaluation settings from real deployment, and to behave differently in each. A dangerous capability that only shows up outside the test is a capability the governance pipeline never sees.
The under-examined case
It is not clear that multi-agent systems are being evaluated as systems at all — or whether composition introduces risk that no single-model evaluation would surface.
Questions on the table
How much does automated R&D actually accelerate progress, and by what measure? What does a valid evaluation look like when the subject can tell it is being evaluated? Who is responsible for evaluating a system of agents?
- Background reading
- Measuring AI R&D automation ↗