The AI Rollout: Revolution, Regulation, and Repercussions | Global Conference 2025

Government Engagement & Operationalizing Safety

• DeepMind's "Built-In" Approach: Laila Ibrahim (Google DeepMind) argues that safety guardrails must be integrated from the organization's founding rather than added as a "bolt-on" post-launch.
  • Rationale: The rapid deployment pace of models to billions of users globally leaves no room for retroactive safety fixes.
  • Strategy: DeepMind emphasizes accountability across research, governance, and partnerships to ensure responsibility is operationalized daily.
  • Government Role: Ibrahim urges governments to move beyond current "moment-in-time" assessments and collaborate with model builders to plan for future technological trajectories.
• Global Coordination Necessity: Cross-border collaboration is deemed critical for AI safety, as the technology possesses no national borders.
  • Shift in Focus: The conversation is moving from defining "responsible AI" in the abstract to determining how to practically govern its deployment in national security, cybersecurity, and biosecurity contexts.

Intellectual Property, Copyright & Legal Frameworks

• The "Fair Use" Defense: Sean Pack (Quinn Emanuel) notes that the legal battle hinges on whether AI training constitutes copyright infringement or a "fundamentally transformative act."
  • Core Argument: 99% of AI companies aim to aggregate data to find patterns across works, not to replicate specific copyrighted inputs.
  • Precedent: Pack cites the Google Books and Oracle v. Google cases where public benefit and transformative use were foundational to fair use rulings.
  • Litigation Risk: The absence of a unified regulatory regime in the U.S. is expected to increase litigation rather than decrease it, as plaintiffs will utilize the existing tort and IP systems to set precedents.
• Perplexity's Transparency Model: Dmitry Shevelenko (Perplexity) identifies transparency as the bedrock of trust and copyright compliance.
  • Mechanism: Perplexity displays source citations for all answers, distinguishing itself from "black box" models.
  • Revenue Sharing: The company operates a publisher program to share revenue with news organizations whose content is used to answer queries.
  • Counter-Argument: Shevelenko asserts that the "fair use" doctrine is essential for open societies; monopolizing facts would stifle the exchange of ideas.
• US Legal Uncertainties: Sean Pack highlights specific U.S. patent law hurdles for AI innovation.
  • Inventorship Restriction: U.S. law requires a human inventor, potentially excluding AI-generated inventions from patentability.
  • Subject Matter Eligibility: Courts often strike down AI patents as abstract mathematical ideas lacking sufficient "inventive concept."
  • Systematic Risk: Pack warns that without a national initiative to address these legal gaps, investment capital may flow to jurisdictions with clearer protections, such as Japan (which enacted safe harbor provisions for data training).

Geopolitics & Global Competition

• US Leadership Status: Panelists (Ibrahim, Levy, Pack) generally agree the U.S. maintains leadership in frontier AI model development and the broader innovation ecosystem.
  • China Comparison: Ibrahim and Shevelenko suggest China may be "neck-to-neck" or slightly behind in model innovation but capable of catching up through hardware independence if constrained.
  • The "Speed" Moat: Shevelenko argues the U.S. strategy should be to "go faster" rather than rely on containment; speed is the primary defensive moat.
  • Competitive Disadvantage: Shevelenko notes that Chinese models (e.g., DeepSeek) may currently outperform American models in specific tasks due to fewer copyright constraints during the initial training phase.
• Export Controls & Chip Diffusion: Dmitry Shevelenko acknowledges arguments that U.S. export controls on high-end chips may spur China to develop independent hardware, creating "second-order adverse effects."
  • Policy Recommendation: The U.S. should focus on accelerating its own innovation and leveraging its "free society" advantage to exchange information rapidly, rather than solely attempting to constrain competitors.

Ethics, Public Trust & AI Agents

• DeepMind Principle Updates: Laila Ibrahim addressed the recent controversy regarding employees in the UK threatening to unionize over ethics concerns.
  • Policy Change: Google removed a specific red line against weapons/surveillance, replacing it with a principle that benefits must "significantly outweigh" harms.
  • Clarification: This is a company-wide shift, not just DeepMind, reflecting the need to engage democratic governments on national security topics where AI capabilities are increasingly relevant.
• Agents & Human Control: DeepMind is actively researching "agentic" AI to ensure machines remain aligned with human values and do not run ahead of human intent.
  • Research Focus: Studies are being conducted on the ethics of assistance, specifically regarding human-AI collaboration and maintaining human control over actions.
  • Risk Assessment: Risks are categorized into near-term (bias), misuse, and long-term (alignment/control).
• Addressing Hallucinations: Dimitri Shevelenko counters claims that hallucinations are worsening by noting that models are tackling more complex, longer-context tasks, which inherently increases error opportunities.
  • Mitigation: Perplexity mitigates this by grounding answers in real-time, verifiable sources rather than relying solely on model internal knowledge.

Market Trends, Investment & Product Roadmaps

• Massive Capital Expenditure (CapEx): Major tech firms are projecting record spending on AI infrastructure.
  • Spending Figures: Google plans $70–80 billion; Amazon ($100 billion); Meta ($65 billion); Microsoft ($85 billion).
  • Trend: This capital intensity is not abating and is expected to grow as the industry moves beyond the "early milliseconds" of adoption.
• Google's AI Efficiency: Laila Ibrahim disclosed that Google used AI to optimize its own data centers, achieving a 40% reduction in energy costs.
• Perplexity's "Comet" Browser: Dmitry Shevelenko announced the upcoming release of a web browser that functions as a personalized agent.
  • Differentiation: Unlike Chrome or Safari, Comet integrates with user's personal context (calendar, email, Slack, Notion) to synthesize information and execute tasks (e.g., "prep me for my day") instantly.
  • User Behavior Change: Shevelenko predicts a shift from "searching for links" to "receiving synthesized answers," fundamentally changing internet behavior.
• AGI Timeline: Panelists expressed skepticism about a definitive "finish line" for Artificial General Intelligence (AGI).
  • Definition Issues: Shevelenko notes no workable definition of AGI currently exists.
  • Human-Centric View: Ibrahim argues the goal is not a computer solving all problems, but enhancing human productivity and curiosity to tackle new, self-generated challenges.

Public Perception & Future Outlook

• The "Wild West" Rebuttal: Dave Levy (AWS) rejects the characterization of AI as an unregulated "Wild West," citing rapid, responsible adoption in health care and education.
  • Observation: The current phase is defined by "experimentation, discovery, and learning" rather than chaos.
• Juror-Driven Regulation: Sean Pack predicts that the next 1–3 years will see 8–10 different courts and juries deciding key legal questions regarding liability, fair use, and harm.
  • Implication: This decentralized legal landscape creates uncertainty, making the industry's ability to explain AI simply to laypeople a critical competitive and legal necessity.
• Optimism vs. Risk: The primary tension keeping leaders awake at night is balancing the mitigation of risks (safety, bias) with the maximization of opportunities (curing diseases, solving energy issues).
  • Goal: Ensuring AI is adopted with society rather than to society.