Interview
A.I. Policy and Public Perception - Miles Brundage and Tim Hwang
- Balancing the AI hype cycle with cryptocurrency developments is expected to yield positive outcomes for the field, while global investment, particularly from China, is projected to intensify.
- Malicious AI use, including fake news generation, drone-based attacks, and offensive cybersecurity, is anticipated to drive the creation of a comprehensive analysis report and establish norms like responsible disclosure for adversarial examples.
- Deepfake technology is expected to become democratized through cloud services, making the creation of believable fakes scalable and accessible to non-experts, potentially altering the media ecosystem.
- Autonomous drone technology is progressing, though end-users are not expected to easily repurpose them for tracking individuals in the near term due to complexity.
- AI tools are approaching a point of efficiency for wide-scale use by large populations, with virtual applications likely emerging before physical ones due to hardware costs and scalability challenges.
- The convergence of AI scalability with human intelligence in vulnerability discovery poses a risk of catastrophic cyber incidents, while adversarial offensive techniques are expected to continue defeating defense mechanisms.
- Society is predicted to remain largely unresponsive to data disclosures even if adaptive malware becomes a significant threat, and the public may not reject AI systems regarding news feeds due to misidentification of their usage.
- AI application in health tasks like dermatology and cancer detection will see continued progress, though widespread deployment is pending the resolution of interpretability and fairness issues.
- Regulatory approaches are expected to diverge cross-nationally, with the EU potentially favoring broad automated decision-making regimes while the U.S. adopts a case-by-case basis.
- Governments are predicted to increasingly view AI as a national security issue, influencing research priorities, while the Chinese government may prioritize deployment over academic interpretability research.
- Policy development is expected to gain methodological rigor, potentially mimicking climate change frameworks like the IPCC, as experts work to bridge gaps between technical research and liability discussions.
- Collaborative efforts between societal and technical researchers are anticipated to improve assumptions about capability timelines, with a focus on common causes in fairness, accountability, and security.
- The Future of Humanity Institute is expected to continue focusing on reducing existential risk probabilities, with short-term governance issues potentially setting a positive precedent for longer-term risks.
- Superhuman performance in StarCraft is predicted with 70% confidence within the upcoming timeframe, alongside significant, currently unexpected improvements in meta-learning.
- Corrigibility is believed to be a stable state for future powerful AI systems, and in the absence of catastrophic misuse, community norms are expected to remain pro-open publishing for realistic risk assessment.