How to govern AI — even if it's hard to predict | Helen Toner

Episode Summary

Executive Summary: Helen Toner argues that AI is advancing faster than our understanding of it, making traditional regulation difficult but not impossible. She urges policymakers and the public to prioritize adaptability over certainty through measurement, disclosure, external audits, and incident reporting, so society can steer AI toward broadly beneficial futures rather than leaving outcomes to a few powerful companies.

Main Topics: AI is hard to understand and predict (Priority: 5/5): Toner explains that even experts do not fully understand how modern AI systems work internally, which makes forecasting capabilities and risks difficult. The field lacks a clear definition of intelligence (Priority: 5/5): She argues that there is no consensus on what intelligence means, and old categories like narrow AI versus AGI no longer cleanly describe systems like ChatGPT. The 'black box' problem in neural networks (Priority: 4/5): Modern AI models contain enormous numbers of parameters, and current methods are not yet good enough to interpret what those numbers are doing. Do not be intimidated by AI or its builders (Priority: 4/5): Toner says AI is not magical and that non-experts, affected communities, and the public have a legitimate role in shaping AI governance. Policy should emphasize adaptability over certainty (Priority: 5/5): Rather than choosing between heavy-handed regulation and laissez-faire innovation, she advocates flexible governance that can respond as AI changes. Practical governance tools: measurement, disclosure, audits, incident reporting (Priority: 5/5): She proposes concrete mechanisms to better understand AI capabilities, require transparency from companies, and learn from failures in the real world. AI’s future should not be left to a few companies (Priority: 4/5): Toner warns that default incentives may concentrate power and narrow AI’s use cases, but society can still push for more beneficial applications.

Key Arguments: Experts can build and deploy AI systems, but they still do not deeply understand how these systems work internally, which limits prediction and governance. The old narrow-vs-general AI framework is no longer sufficient; systems like ChatGPT are general-purpose without being human-level intelligent. The main obstacle in understanding AI is not mystical complexity but the sheer scale of parameters in neural networks and the lack of interpretability tools. AI interpretability research is making progress and could significantly improve understanding within 5-10 years. Public participation in technology governance is essential; expertise matters, but affected groups should help shape rules and norms. AI policy should be built for changing conditions, using tools that improve visibility into capabilities and risks rather than assuming certainty. Governments should require capability measurement, company disclosure, independent audits, and incident reporting to create a clearer picture of AI development. Without intervention, AI companies may follow social-media-style incentives and concentrate power in a small number of firms or individuals. AI has enormous positive potential, including language translation, protein structure prediction, fusion energy, and agricultural innovation, so governance should enable beneficial uses while managing risks.

Data Points: Years of AI policy experience: About 8 years - Toner says she has worked on AI policy and governance for about eight years, first in San Francisco and now in Washington, D.C. Interpretability progress horizon: 5 to 10 years - She suggests that continued interpretability research could make AI systems much clearer to understand within five to ten years. Model scale: Millions, billions, or even trillions of numbers - She describes deep neural networks as containing enormous numbers of parameters that are difficult to interpret.

Pivotal Quotes: "I don't understand AI, and neither does anyone else." — Helen Toner: She uses this line to capture the surprising lack of internal understanding even among experts. "AI is already happening to us." — Helen Toner: She emphasizes that society cannot wait for perfect clarity before acting on governance. "We are users. We're workers. We're citizens." — Helen Toner: She underscores that everyone has a stake and a voice in shaping AI’s future.

Implications: Listeners should expect AI governance to rely on flexible, evidence-based oversight rather than perfect understanding. For industry and policymakers, transparency, audits, and reporting will be key to steering AI toward broad public benefit and away from concentrated power.