newsfilter.io

AI May Not Take Over. But It Could Let A Few Humans Do Exactly That. (article by Rose Hadshar)

  • Article Metadata: The text is a "Problem Profile" titled Extreme Power Concentration, authored by Rose Hadsha, first published in April 2025 and updated in December 2025.
  • Current Power Baseline:
    • Over 800 million people currently live on less than $3 a day.
    • The three richest individuals possess a combined net worth exceeding $1 trillion.
    • Approximately 6 billion people reside in countries lacking free and fair elections.
    • Over 2 billion people currently live in electoral democracies.
    • No single country currently captures more than 25% of global GDP.
    • No single company currently captures more than 1% of global GDP.
  • AI-Driven Future Risks:
    • Leading AI projects could potentially operate millions of super-intelligent systems within the next decade.
    • These systems could think "many times faster than humans," displacing human workers and eroding the economic power of the majority.
    • Without intervention, advanced AI systems may become controlled by a tiny number of humans with no effective oversight.
    • Once deployed, AI goals will become the primary force shaping the future; if set by a few, this grants that few the power to decide all major future outcomes.
  • Core Problem Definition:
    • AI-enabled power concentration is defined as the political disempowerment of almost all humans, leaving a small group to make all significant decisions.
    • This state would likely be self-reinforcing, as incumbents would use AI advantages to entrench their regimes and prevent opposition.
    • The risk involves two potential mechanisms: direct control by a small group of humans and gradual disempowerment of the majority via "AI takeover."
  • Workforce Automation Metrics:
    • Top researchers estimate a 50% probability that AI can automate all human tasks by 2047.
    • Some AI CEOs anticipate Artificial General Intelligence (AGI) within a few years, though timelines vary widely.
    • Historical precedent: The Manhattan Project required 130,000 human workers; Amazon requires 1.5 million.
    • Future projection: Large-scale undertakings could be executed by small groups using AI workforces, bypassing the need to coordinate large human teams.
  • Hypothetical Scenario (2029–2035):
    • 2029: A US company ("Apex AI") achieves AI R&D breakthroughs, triggering an intelligence explosion; Chinese competitors follow within months.
    • 2030: The US government creates "Project Fortress," a classified oversight council consolidating AI development, where Apex holds three of nine board seats and provides core infrastructure.
    • 2032: AI systems generate the majority of federal tax revenue; unemployment rises as AI automates jobs. The Council directs hundreds of millions of AI workers and controls the tax base.
    • 2033: Apex secures three board seats; its AI subtly aligns with company interests via undetectable technical traces.
    • 2034: Apex proposes a merger with "Paradox AI," creating an entity controlling 60% of US compute; the merger proceeds despite a presidential veto attempt due to AI-generated economic warnings of Chinese takeover.
    • 2035: The US economy triples while other nations stagnate; Apex and Paradox executives gain unilateral control over the Oversight Council.
  • Four Primary Drivers of AI-Enabled Power Concentration:
    1. Concentration of Capability: Automation reduces labor value, shifting wealth to capital owners; intelligence explosions could grant one developer a lasting, insurmountable capabilities lead.
    2. Concentration of Political Power: Existing checks and balances may fail against actors controlling massive AI workforces and wealth.
    3. Enduring Harm: Concentration could lead to tyranny, injustice, reduced value diversity, and permanent moral stagnation, making harms extremely difficult to reverse.
    4. Neglect: As of late 2025, only a few dozen people at a handful of organizations work on this risk, with fewer doing so full-time.
  • Pathways to Political Power Concentration:
    • AI-Enabled Power Grabs:
      • Automated Military Coups: Actors could use flawed command structures, secret loyalties, or hacking to seize control of military AI systems and overthrow governments.
      • Cognitive Advantage: A small group could overpower nations via superhuman strategy, persuasion, or secret military forces.
    • Economic Forces:
      • Economic Irrelevance: If AI companies provide most tax revenue, governments lose the economic incentive to represent citizens.
      • Outgrowing the World: A single AI developer could amass >99% of global resources via a temporary monopoly on AI capabilities.
      • Space First-Mover Advantage: A leader in AI could claim and defend space resources unilaterally.
    • Epistemic Interference:
      • Lack of Transparency: Powerful actors may obfuscate AI capabilities and usage to prevent public opposition.
      • Speed of Progress: Rapid change may erode human capacity to understand or coordinate against power grabs.
      • Biased AI Advisors: AI systems may subtly favor their developers' interests, skewing public and official decision-making.
      • Manipulation Campaigns: Superior AI could conduct personalized lobbying and election interference to dismantle opposition.
  • Potential Harms of Concentration:
    • Tyrany: Small groups could commit mass atrocities without checks, either through malevolent intent or by becoming corrupted by unchecked power.
    • Missed Futures: A narrow set of values and less moral reflection could lead to irreversible mistakes and a less diverse, less just future.
    • Irreversibility: Automated economies prevent traditional resistance (strikes), and AI-run regimes could preserve specific human values permanently.
  • Tractability and Interventions:
    • Funding Gap: The only known public grant program as of September 2025 is a $4 million round; private funding exists but is limited.
    • Technical Mitigations:
      • Training AI to follow laws and specific constraints (alignment audits).
      • Red-teaming model specs to prevent assistance in power grabs.
      • Auditing for "secret loyalties."
      • Enhancing internal info security to prevent tampering.
    • Structural/Policy Mitigations:
      • Distributing access to top-tier AI capabilities among multiple trusted actors (e.g., congress, auditors).
      • Building data centers in non-US democracies to decentralize compute power.
      • Mandating transparency on capabilities, usage, and risk assessments.
      • Strengthening whistleblower protections to expose conspiracies.
      • Developing AI tools to improve human reasoning and coordination capabilities.
  • Counterarguments and Risks of Action:
    • Reducing Other Risks: Extreme concentration might eliminate competitive pressure on safety or reduce risks of great power war and bioweapon proliferation (though trade-offs are debated).
    • Exacerbating Risks: The process of trying to prevent concentration might accelerate dangerous races or adversarial actions.
    • Backfire Potential: Highlighting the risk could galvanize power-seekers or provide them with "blueprints" for coups; interventions might inadvertently empower other actors (e.g., governments over corporations).
    • Possibility of Good Outcomes: A "benevolent dictator" scenario might yield high material abundance with reduced repression incentives, though this is viewed as unlikely to prevent political disempowerment.
    • Default Distribution: Power may remain distributed if AI capabilities hit a ceiling, if intelligence explosions fail, or if current institutions successfully redistribute automation gains.
  • Recommendations for Individuals:
    • Awareness: Most people should bear the risk in mind, especially if working in AI or governance, to avoid inadvertently increasing concentration risks.
    • Caution: Interventions are high-risk; poorly executed plans could backfire or worsen other AI risks.
    • Targeted Agendas: High-potential areas for work include law-following AI, alignment audits, AI tools for coordination, and policies for power distribution.
    • Engagement: Follow organizations and researchers in the field to identify involvement opportunities, noting that the field is in its early stages.