Challenges for policy makers in AI governance:
- Performance of frontier LLMs is jagged across different tasks. This makes it difficult to reliably measure or predict what LLMs are capable of (which could trigger certain legislative or regulatory action).
- LLMs can be misused by bad actors in ways either unintended by developers or in ways that developers are actively trying to prevent (e.g. cybersecurity, biorisk).
- Privacy and ensuring data intended for one purpose is not misappropriated for another.
- Ensuring benefits of AI diffuse equitably and any biases in training data are not amplified.
- Effective oversight in the face of (i) rapid development cycle, (ii) AI providers based in USA, (iii) vast difference between the resources being channeled to building AI systems vs resources for AI safety, (iv) proliferation of fast-acting autonomous AI systems.
- The shift away from isolated static systems (chat, classifier) to multi-agent systems acting autonomously and interacting with external systems.
- Poor institutional capacity - existing institutions are ill-placed to handle this rapidly evolving technology.
The fear is too much regulation will stifle innovation and block the benefits of AI, too little risks facilitating any number of harms.
Solutions include:
- Charlie Bullock and Christoph Winter (Institute for Law & AI) propose ‘radical optionality’. This calls for governments to develop mechanisms for information-gathering (e.g. whistleblower protections, and mandated reporting) and to build capacity within government in order to better inform future decision-making.
- Dean Ball suggests developing private regulators which can then be certified by government.
- Gabriel Weill calls for reform to tort liability regimes (including insurance requirements, and strict liability for certain harms).