Building Technology to Drive AI Governance
Technically skilled people who care about AI going well often ask me: how should I spend my time if I think AI governance is important? By governance, I mean the constraints, incentives, and oversight that govern how AI is developed.
One option is to focus on technical work that solves problems at the point of production, such as alignment research or safeguards. Another common instinct is to get directly involved in policy: switching to a policy role, funding advocacy, or lobbying policymakers. But internal technical work does little to shift the broader incentives of AI development: without external incentives, safety efforts are subject to the priorities of leadership, which are ultimately dictated by commercial pressure and race dynamics. Conversely, wading into politics means giving up your main comparative advantage to fight in a crowded, intractable domain full of experienced operators.
I want to argue for a third path: building technology that drives governance, by shifting the underlying dynamics of AI development: the information available, the incentives people face, and the options on the table. To take an example from another domain: oil and gas operations leaked massive amounts of methane until infrared imaging made the leaks measurable from space, leading the EU to require satellite-verified standards for fossil fuel imports. More generally, across domains spanning climate change, food safety, and pandemic response, there are two technological mechanisms that repeatedly drive governance:
- Measurement, which creates visibility, enables accountability, and makes regulation feasible.
- Driving down costs, which makes good behavior economically practical and can dissolve apparent trade-offs.
I'll first show how these patterns have played out in climate, food safety, and pandemic response; the first two in particular have structural parallels to AI. I'll then show how the same framework identifies important technologies to build for AI governance, many of which are just starting to be developed.
If you have strong technical skills, building these technologies is currently the most leveraged thing you can do: more so than either direct technical work that ignores governance, or policy work that is untethered from technical solutions. This cross-sectional work is significantly neglected, and governance bottlenecks are often fundamentally technical: we can't regulate what we can't measure, and desired practices can't become standard until they're cheap and replicable.
Technological Levers in Other Domains
Historical case studies help ground current practice. AI shares structural features with two other familiar domains:
- Climate change: greenhouse emissions cause both present-day harm (pollution; AI slop) and future long-tail risks (global warming; loss of control).
- Food safety: companies have commercial incentives to hack human reward systems (junk food; sycophantic AI); and there is optimization pressure that creates dangerous side effects (antibiotic-resistant bacteria; deceptive AI).
For climate, both improved measurement and reductions in cost significantly drove better outcomes. They helped in the following ways:
Measurement to orient strategy. Public measurements inform strategy and galvanize action. Global monitoring of temperature and CO₂ (the Keeling curve) were the foundation of modern climate policy and help us continuously monitor progress.
Measurement to shift incentives. Satellite imagery of site-level emissions shifts incentives by making methane leaks visible and attributable to specific operators, with major leaks now repaired within days of detection.
Measurement to enable governance. Simulators of road resistance called chassis dynamometers let regulators produce repeatable fuel efficiency measurements, which made the CAFE standards possible. Similarly, continuous measurements of industrial CO₂ emissions were necessary for cap-and-trade.
Driving down costs to shift equilibria. Probably the largest driver of decreased CO₂ emissions has been the emergence of cheap wind and solar power. This illustrates a powerful dynamic: as production scales, costs fall along experience curves (sometimes called Wright's law), until clean energy becomes the default rather than the alternative. Before this inflection point, decarbonization meant fighting against economic incentives; after it, the incentives pulled in the same direction, and the market did the work of driving further R&D.
Driving down costs to dissolve trade-offs. The same Wright’s law dynamic can also help resolve thorny dilemmas. Before electric vehicles, society faced a trade-off between lower emissions and convenient transportation. As battery production scaled and costs fell, electric vehicles hit the mass market and the trade-off began to dissolve.
These patterns appear across climate, food safety, and COVID-19. The table below summarizes them alongside upcoming challenges for AI that we'll describe in the next section.
Role | Technology | Governance impact | Domain |
Orient strategy | Global temperature + CO₂ monitoring | Understand that warming is happening and how fast | Climate |
Understand growth rate and prevalence | COVID-19 | ||
Track capability growth; calibrate policy thresholds | AI (exists) | ||
Shift incentives | Make site-level leaks visible, creating pressure to fix them | Climate | |
Behavioral benchmarks (sycophancy, deception, etc.) | Create competitive pressure for better model behavior | AI (needed) | |
Enable governance | Enable repeatable fuel efficiency measurement for CAFE standards | Climate | |
Enable cap-and-trade and emissions regulation | Climate | ||
Genetic sequencing of food-borne bacteria | Enable outbreak attribution and targeted enforcement | Food safety | |
Compute accounting, evaluation standards | Enable oversight of training runs and model deployment | AI (needed) | |
Shift equilibria | Replace dirty energy in the open market | Climate | |
Make safe milk the cheap default | Food safety | ||
Make prevention economically feasible at scale | COVID-19 | ||
Make rigorous oversight standard practice | AI (needed) | ||
Dissolve trade-offs | Dissolve trade-off between emissions and convenience | Climate | |
Dissolve trade-off between transparency and IP protection | AI (needed) |
Food safety offers some of the cleanest examples of technology enabling governance. Routine genetic sequencing of foodborne bacteria made outbreak attribution possible, enabling targeted recalls and enforcement. And cheap pasteurization made safe milk the default. Similarly to solar energy, subsidized milk depots initially proved the market, and later commercial investment drove costs down. Mandatory pasteurization standards have now significantly reduced typhoid outbreaks.
Finally, while COVID-19 is less structurally parallel to AI, the wins from technology are stark. Testing created visibility into the virus's spread, orienting response at every scale from individual treatment to national policy. Cheap vaccines drove down the cost of prevention, resolving the trade-off between economic normalcy and infection control.
Concrete Technical Levers for AI
For AI, as in other domains, measurement is one of the key levers that can drive governance. I'll focus on this lever primarily, where measurement is helping to track AI trajectories, create competitive pressure for better model behavior, and make regulation enforceable. I'll then turn to driving down cost, where we will primarily consider driving down the cost of oversight, both by automating evaluation and by removing barriers to external auditing.
Measurement to orient strategy. Just as CO₂ monitoring oriented climate strategy, we need metrics that track AI trajectories and approaching thresholds. For climate, CO₂ was a natural target, since the causal chain from emissions to warming was scientifically clear. For AI, the answer is less obvious.
One example I find compelling is METR's work on agent time horizons, which tracks the complexity of tasks AI systems can complete autonomously, measured by how long those tasks take humans. Agents that can complete week-long tasks unsupervised pose very different challenges than those limited to minutes of work. METR finds that time horizons have been doubling roughly every seven months since 2019. If this continues, we could see agents capable of month-long autonomous tasks within a few years, which has significant implications for both labor markets and safety.
Another good example is Epoch's key trends in AI. Their reporting on training compute growth—roughly 4-5x per year—helps calibrate how quickly regulatory thresholds will be crossed, and their cost estimates inform questions about who can afford to train frontier models.
Measurement to shift incentives. We lack good public metrics for sycophancy, deception, reinforcement of delusions, and similar behavioral issues, which are currently measured ad hoc, if at all. Where metrics do exist, they create competitive pressure: labs compete for top positions on the Chatbot Arena leaderboard, featuring rankings prominently in official announcements. High-quality public dashboards for behavioral issues could do the same, just as fuel efficiency became a selling point for automakers once EPA ratings became standardized. This is a big part of what we think about at Transluce: identifying what should be measured to improve incentives, and building the infrastructure to measure it.
Measurement to enable governance. The EU AI Act requires frontier developers to perform evaluations using "standardized protocols and tools reflecting the state-of-the-art"; California's SB 53 and the Trump administration's AI Action Plan impose similar expectations. However, without reproducible evaluation suites, such requirements are difficult to enforce: developers largely define how their own systems are measured, making results hard to compare. Compute monitoring faces similar challenges: tracking large training runs requires technical work on compute accounting that's still in early stages.
Driving down the cost of oversight. In an ideal world, rigorous evaluation and oversight of AI systems would become standard practice through natural incentives alone. Developers want to know whether their systems behave as intended; users and customers want assurance; and once evaluation is cheap, market forces and liability concerns handle the rest.
We don’t live in this world yet, partly because high-quality evaluation of agent runs is currently very expensive. For example, METR often spends person-months of work on a single evaluation suite, due to the labor involved in human baselining, running many trials, and manually analyzing results to understand why agents succeed or fail. We need to make these analyses cheap enough to become ubiquitous while preserving and improving their quality.
This is essentially what we’re trying to do with Docent: building tools that accelerate complex agent evaluations. Our experience matches the Wright's law pattern: iterating across many users and problems has driven down costs while improving quality. However, demand for these harder evaluation tasks (complex agentic behaviors, subtle failures of judgment, deceptive patterns over long horizons) is growing but still emerging; the mass market tends to focus on toxicity, hallucinations, or compliance-related issues. That's what makes it high-leverage to push on complex oversight tasks now, similarly to early solar investment before the market tipped.
Reducing trade-offs between transparency and IP protection. As a society, we'd like to be able to audit whether AI systems behave badly in deployment, verify claims about training practices, and conduct white-box safety analyses. But these goals currently trade off against legitimate IP concerns: companies are reluctant to give external parties access to model weights, training data, or system logs.
Technology can dissolve this trade-off. Secure enclaves could let auditors run analyses without extracting underlying weights. Cryptographic methods could let companies prove properties about their training process without revealing proprietary details. Structured access protocols could enable third-party evaluation while limiting what information leaves the company. With mature technologies for confidential auditing, deeper forms of oversight become practically viable.
What You Can Do
If you're technically skilled and care about AI going well, solving the problems described above is where you have the most leverage. The bottleneck on effective AI governance is not just political will: there's appetite to regulate AI among both voters and policymakers. The bottleneck is that we don't yet have the measurement infrastructure, the cheap evaluation tools, or the well-specified policy options to regulate well.
The field is talent-constrained in a specific way: measurement and evaluation work is less glamorous than capabilities research, and it requires a rare combination of technical skill and governance sensibility. The organizations doing this work—Transluce, METR, Epoch, US CAISI, and others—are small and growing. If these arguments resonate and you fit the profile, consider joining one of them; or, if you see a gap no one is filling, start something new.
We have a unique opportunity right now, and tackling these high-leverage challenges is what excites me the most. AI will cause disruption, and this will open a window for policy solutions, probably sooner than many expect. The question is whether we'll have the technical foundations ready when it does.
Discuss