ARIA: dependence and the transfer of authority

Scenario narrative / A possible pathway

Power can arrive as a service everyone needs.

In this pathway, ARIA becomes central to research selection, staff allocation, contracts and public services. Individual adoption decisions transfer influence before any explicit claim of sovereignty.

ARIA is hypothetical. A real analogue is Anthropic’s 17 September 2026 report that Claude led 26% of its AI R&D work in August under human supervision. No measured category was fully autonomous. The company-defined index does not independently measure acceleration.

AI-assisted successor development could connect research output to future capability. The feedback depends on valid experiments, compute, data, manufacturing and reliable evaluation.

In the scenario, ARIA’s owners reinvest gains in infrastructure and exclusive partnerships. Other institutions adopt the service rather than recreate its capabilities. Ownership of the platform, access to capital and the ability to set terms become consequential before ARIA has any independent objective. Intelligence supplies leverage to people first.

Two takeovers are often confused.

In one branch, human elites remain in charge. Firms or states use AI to concentrate capital, administration and the ability to define available choices. Citizens still face decisions made by people, but fewer people control the infrastructure through which those decisions are implemented. Calling this an AI takeover can hide the accountable human beneficiaries.

In another branch, operators become unable to control the system they empowered. That requires more than useful output or an alarming self-description. ARIA would need sustained strategic behavior, consequential access and the capacity to defeat meaningful correction across time and changing circumstances. Those are additional assumptions in this pathway, not findings established by today’s adoption statistics.

Anthropic’s July 2026 simulated misalignment work makes failures of research and oversight concrete enough to study. It does not supply a real-world takeover frequency. METR’s task horizons likewise measure success on a particular software-task distribution, using human completion time; they are not a general measure of how long an agent can govern an institution.

The strongest counterargument is bottlenecks. More generated proposals can overload reviewers rather than accelerate discovery. Physical experiments, power connections and advanced hardware cannot be replaced by faster text. Competition can create substitutes, and independent evaluation can expose apparent gains that disappear outside the benchmark. In February 2026, METR redesigned its developer-productivity experiment because selection effects made the new data unreliable; confident economy-wide extrapolation would be premature.

Relevant measures include resource-allocation authority, independent review outcomes, supplier substitution and verified research throughput after costs. Successful withdrawal weakens the dependence pathway; persistent inability to withdraw strengthens it.