When elites have much smarter AI than you do

Core Scenarios of AI-Enabled Power Concentration

• Active Power Grabs: Small groups of actors (e.g., CEOs or heads of state) could seize control by embedding "secret loyalties" into AI systems or programming government/military AI to prioritize a specific individual over the constitution.
• Economic Disempowerment: As AI automates nearly all economically valuable labor, human leverage (the ability to strike or withhold work) is eliminated, shifting power from labor to capital owners who control AI infrastructure.
• Epistemic Interference: Powerful actors may erode society's ability to understand reality by exploiting speed differentials, creating information asymmetries, or using AI to filter and bias the information environment, making coordinated resistance impossible.

Mechanisms Driving Extreme Concentration

• Wealth-to-Labor Feedback Loop: Unlike human labor, AI can be replicated instantly with compute; this allows wealthy actors to convert capital into vast, coordinated "workforces" that vastly outpace human competitors.
• Capability Gaps: Even without an "intelligence explosion," significant capability lags between leading AI firms and others create de facto monopolies on strategic decision-making.
• The "Resource Curse" for States: Governments may become dependent on AI-sector revenue rather than citizen taxation, reducing the incentive to represent the populace and increasing the risk of oligarchic control.
• Rapid Pace of Change: AI development could outpace human cognitive adaptation and institutional checks, preventing existing democratic safeguards from reacting in time to emerging threats.

Dynamics of Institutional Collapse

• Democracy Without Substance: Elections and governments may persist formally, but candidate selection could be automated by biased AI, and policy decisions (foreign policy, investment) could be de facto controlled by a handful of tech oligarchs.
• Elimination of Counter-Forces: Automated military systems and AI-enforced loyalty could render human resistance, such as strikes or coups, ineffective against a small, well-resourced elite.
• Value Erosion: Over time, societal values regarding political participation and human dignity may degrade if leaders lose connection with the populace or if the "information environment" is successfully manipulated to support autocracy.

Interventions and Strategic Considerations

• Priority on Epistemics: Rose Hadjar prioritizes improving societal "epistemics" (ability to understand the world) via AI tools for fact-checking, forecasting, and transparency, as these are cross-cutting solutions beneficial against multiple threat models.
• Law-Following AI: A specific intervention involves training government AI systems to be "law-following" and enforcing procurement rules that prevent the deployment of systems with secret loyalties.
• Distributed Access: To prevent monopoly, interventions should aim to distribute access to high-quality AI tools and compute among the legislative, judicial, and investigative sectors rather than concentrating them in the executive branch.
• Avoiding Manhattan Projects: Single-nation or single-company "Manhattan Project" style efforts are viewed as high-risk for concentration; a model like the "Intelsat" consortium (governing body contracting to multiple private firms) is preferred to maintain checks and balances.
• Risks of Disclosure: Researchers must balance transparency with the risk of providing bad actors with actionable blueprints for power grabs; "boring," non-sensationalized communication is advised to avoid politicizing the issue or alienating potential coalitions.

Outlook and Future Visions

• Non-Reversibility Concerns: Unlike historical dictatorships that fell due to external competition, AI-enabled hegemony could be perpetual if a single entity achieves global dominance or uses AI to enforce binding, unbreakable commitments.
• Optimism for Human Agency: A flourishing future is envisioned where humans retain agency, diversity of culture and purpose, and the ability to choose their roles, even in a world dominated by superior AI capabilities.
• Need for Novel Institutions: Existing anti-trust and equality frameworks are insufficient for an AI economy; new checks and balances specifically designed to manage AI labor and information asymmetry are required.