Interview
A.I. Policy and Public Perception - Miles Brundage and Tim Hwang
- Shift in Technological Hype: The cryptocurrency boom has displaced AI as the primary subject of public fascination, inadvertently relieving pressure on AI researchers from constant scrutiny regarding an "AI bubble."
- Core Research Focus (Miles): A comprehensive analysis titled "The Malicious Use of Artificial Intelligence" categorizes AI as a "dual-use" or "omni-use" technology, distinguishing intentional misuse (e.g., fake news generation, drone-enabled terrorism, offensive cybersecurity) from unintentional harms like algorithmic bias.
- Proposed Norms: The researchers argue for the adoption of "responsible disclosure" norms in AI similar to those in computer security, where vulnerabilities (e.g., adversarial examples capable of fooling driverless cars) are reported to developers before public release.
- Disinformation Research (Tim): Work at Oxford focuses on the intersection of AI techniques and disinformation, specifically analyzing the scalability and "democratization" of creating deepfakes (e.g., face-swapping pornography) which lowers the expertise barrier for malicious actors.
- Prediction Methodology: Risk assessment relies on extrapolating technical trends from academic literature and identifying when technologies become efficient enough for mass deployment by non-experts, rather than waiting for actual incidents.
- Virtual vs. Physical Threats: Near-term AI risks are perceived to be concentrated in the digital realm (e.g., scalable cyberattacks, hacking) rather than physical harms (e.g., autonomous drones), due to the lower cost, higher scalability, and greater difficulty of hardware integration and environmental perception in the physical world.
- Public Perception Gaps: Society often conflates specific AI applications (e.g., newsfeed algorithms) with the general concept of "AI," while simultaneously anthropomorphizing or fearing physical robots that may not utilize advanced machine learning, complicating public policy responses.
- Top Policy Concerns: Two dominant issues drive policy discussions:
- Geopolitics: Intense international competition, particularly regarding China's investments in AI, framing the technology as a national security imperative.
- Interoperability (Interpretability): The inability of complex systems to explain their decisions, a critical barrier for adoption in regulated fields like healthcare and finance.
- Technical Robustness: Neural networks remain highly vulnerable to "adversarial examples" and "backdoors" (e.g., the "Bad Nets" study), where minor pixel manipulations or hidden triggers cause catastrophic failures, currently outpacing defense mechanisms.
- Positive Applications in Health: AI is seeing rapid pilot deployment in medical imaging (dermatology, radiology) and predictive analytics (hospital readmission, diagnosis), though widespread adoption is currently hindered by interpretability and fairness requirements.
- Interpretability Debate: A divide exists between machine learning practitioners who prioritize empirical performance ("it works") and computer science/legal frameworks demanding step-by-step explainability, with the field potentially "over-indexing" on transparency in some areas.
- Regulatory Divergence:
- Europe: Adopts broad, proactive frameworks (e.g., GDPR's "right to an explanation") focusing on automated decision-making as a category.
- United States: Utilizes a reactive, sector-specific "patchwork" approach (e.g., specific regulations for medical AI) rather than general legislation.
- AI Governance Methodology: Researchers are advocating for more rigorous policy tools akin to the IPCC climate reports, utilizing scenario planning and formal modeling to address deep uncertainty regarding AI timelines and impacts.
- Institutional Collaboration: Effective governance requires a feedback loop between academia, industry, and government; industry focuses on immediate product fairness, academia on fundamental research, and government on broad mandates, necessitating cross-sector dialogue.
- FHI Mission: The Future of Humanity Institute addresses both immediate societal issues and long-term existential risks, positing that reducing the probability of human extinction (even by 0.1%) preserves massive future value.
- Value Alignment Continuity: Near-term fairness and accountability challenges are framed as "value alignment" problems; the methods developed now (e.g., ensuring systems seek critical feedback) could establish precedents for controlling future Artificial General Intelligence (AGI).
- Open vs. Closed Research: The default norm is open publishing to ensure realistic risk assessment and public policy, though exceptions exist for high-stakes domains (e.g., adversarial vulnerabilities in millions of driverless cars) where controlled disclosure is prioritized.
- Future Predictions (2017–2020 Horizon):
- Miles: 70% confidence in AI achieving superhuman performance in StarCraft by the end of the forecast period.
- Tim: Significant advancements in "meta-learning," where AI systems automatically design and optimize their own neural network architectures.