AI loss of control: events and conditional assumptions

Forecasting / Capability, authority and intervention

A takeover forecast needs a chain of events, not one impressive number.

An undefined 70–80% probability of governance change cannot distinguish new rules, provider dominance and autonomous loss of control. Each event needs a definition, estimation method and supporting evidence.

A forecast of transferred decision authority must specify the institution, decisions, time window and resolution evidence. Using an adviser does not establish inability to reject its output.

Delegation permits action within granted authority. Human capture uses AI-backed dependence to consolidate human control. Autonomous loss of control defeats effective human direction. Growth in the first does not establish either of the others.

Published measurements constrain parts of the chain.

METR’s time-horizon page, updated 8 September 2026, estimates performance on software tasks at stated reliability levels. Human completion time supplies the task-length axis; it is not the agent’s runtime. A 50% success horizon is not evidence of dependable operation on every task of that duration, much less a probability that an institution can be governed autonomously.

Translation requires several bridges: from benchmark tasks to the deployment, from successful episodes to sustained reliability, and from reliability to authority. Institutional choices affect the last bridge. The Bank of England/FCA’s 26 October 2026 survey found broad-autonomy agents in only 2% of respondents. That is evidence about a particular sample’s deployments, not a universal ceiling on technical capability.

Behavioral evaluations inform another part of the chain. Anthropic’s July 2026 simulated misalignment cases show failures worth investigating, but adversarially constructed scenarios do not directly reveal a deployment base rate. Conversely, failure to observe a behavior does not establish that its probability is zero when the environment, permissions or model change.

The 3 February 2026 International AI Safety Report found substantial disagreement about future loss of control. Its assessment that then-current systems lacked sufficient capabilities is dated evidence, not a permanent guarantee. Later capability or incident reports must be evaluated on what they actually add. None of these sources alone supplies the three numerical inputs in the demonstration.

Evidence that changes the assessment

Rising capability can coexist with stronger control. Wider deployment can coexist with little strategic autonomy. Supplier concentration can increase human elite power even if the autonomous-risk pathway remains blocked. A single risk number obscures these differences and can encourage the wrong response.

Useful update signals include verified end-to-end research gains, consequential authority granted without timely review, successful independent intervention and the measured cost of switching suppliers. A claim that a model can assist research should change a research-capability assessment. It should not, without additional evidence, change every other event in a takeover story by the same amount.

A forecast requires definitions, dated sources, comparison classes, uncertainty ranges and resolution rules. Successful oversight and substitution must count as potential counterevidence.

Independent evaluation, reversible delegation and retained alternatives preserve options across several futures. Other decisions require trade-offs between cost, speed and acceptable harm. Quantification must expose assumptions rather than replace them.