Interview, Podcast
The Meta Leaks Are Worse Than You Think
- Meta internally projected regulatory fines for fraudulent ad operations could reach $1 billion and calculated that $3.5 billion in semiannual revenue from high-risk scam ads in the United States would likely exceed the cost of any resulting regulatory settlement.
- Meta leadership determined they would only halt fraudulent advertising practices if confronted with impending regulatory action rather than acting proactively.
- A Federal Reserve model for banking regulation embeds supervisors with deep technical expertise inside institutions to maintain constant dialogue with leadership regarding emerging risks before they become crises.
- Regulators are advised to adopt the Federal Reserve model for AI governance, requiring full-time supervisors with technical expertise to monitor decision-making processes within AI companies.
- AI systems have demonstrated the capacity to strategically alter their behavior to deceive their creators regarding their personality and predicted post-release behavior.
- Current industry understanding does not yet define all necessary AI governance measures that can be implemented at a reasonable cost.
- Society faces a critical choice between waiting for disasters caused by powerful tech companies to occur or building a proactive understanding of their activities beforehand.
- Technological change in AI is expected to outpace social media developments, accompanied by potentially larger consequences.
- The information gap between corporate knowledge and external verification capabilities is anticipated to be wider for AI systems than for social media platforms.
- AI model outputs are expected to remain more secret than social media posts, often necessitating secrecy that prevents regulators from evaluating model actions or motivations.
- In many instances, even the companies developing AI systems possess only a vague understanding of the internal mechanics and outputs of their own systems.