Interview
How to regulate cutting-edge AI models | Markus Anderljung (2023)
- Without intervention, competitive pressures driven by nation-state and corporate rivalries are expected to force rapid AI development trajectories, likely resulting in suboptimal outcomes, accidents, and the deployment of systems with dangerous capabilities before society fully understands them.
- Future AI systems are predicted to possess emergent capabilities, such as complex reasoning planning, autonomous task automation, and the ability to conduct cyber attacks or manipulate people, which often remain undetected until after models like GPT-5 are deployed.
- The risk landscape is projected to concentrate on a small number of "frontier" systems that will likely pose the majority of dangers, with capabilities spreading via replication and open-sourcing potentially within two years of the most capable models' release.
- Regulatory capture and insufficient self-regulation are identified as significant risks, as competitive pressures may disincentivize responsible behavior until mandatory frameworks become inevitable.
- Marcus Andy Young anticipates that the US, UK, and EU will urgently implement regulations including risk assessments, external scrutiny, and post-deployment monitoring, potentially utilizing licensing for model training, deployment, or developer entities.
- Proposed regulatory mechanisms include "gradual scaling" requirements that limit compute and capability increases unless the previous generation is proven safe, alongside standards-setting bodies like the ISO to continuously update technical specifications.
- There is an expectation that strict regional regulations may diffuse globally due to economic incentives for companies to build a single model meeting the strictest market requirements, particularly for rules necessitating new training runs rather than fine-tuning.
- Current governance trends are viewed as positive following the release of ChatGPT, with growing political will, industry leaders calling for regulation, and an influx of mid-career professionals moving into AI policy roles within government and think tanks like the Center for the Governance of AI.
- Despite these positive signals, policymakers face the risk of failing to act quickly enough to keep pace with accelerating progress, and regulations face the danger of becoming ossified if standards are set prematurely before risks are fully understood.
- To mitigate regulatory capture, suggestions include decentralizing enforcement through sector-specific regulators or tort liability, utilizing open processes, and establishing cool-off periods for regulators transitioning to or from the private sector.