Threshold Systems

The Mask of Fluency, Driftwatch and the essay A Soft Room With No Exit look at the moment an AI output becomes something a person has to rely on. Before they act, what can they see of its warrant, its uncertainty, or whether it even applies to them? Afterwards, what routes are left for correction or challenge? And when does a helpful system start to narrow the space a person thinks in, or the places they go for support?

A paper now under review asks what plurality among powerful AI systems would have to include before it counts as alignment for the people governed by it. Structural pluralism can stop any one system from authorising itself. It can also leave weaker members ruled by a council of stronger ones, however well-meaning. Independent evaluation and the standing of affected members to challenge decisions have to be secured separately, and the paper argues that each is necessary.

The experimental programme below tests when plurality yields useful correction, and when it collapses into correlated judgement or a new centre of control.

The same problem turns up in organisations and economies, where some actors are now far more capable than others.

Prediction, Configuration and Endogenous Demand (preprint, below) reopens the old argument over economic planning. Predictive systems can stand in for local knowledge that could never be handed over directly. Once their forecasts start shaping what people want, though, demand stops being clean evidence of preference.

A companion paper, also under review, follows that setting into agency and responsibility for people who become working parts of a larger coordinating system.

A further manuscript looks at how technology changes organisational depth, and with it who still holds the leverage to steer.

AI makes these old organisational questions harder to put off.

The plural-alignment work has a public CPU research harness, now at v0.2.1. It models one narrow institutional question. When capability and authority are uneven, does a claimant still have a usable route through recognition, evidence, review and remedy?

v0.2.1 separates results fixed by the institutional design from those that depend on cognitive assumptions such as reviewer noise and correlated error. Plural standing that follows authority upward corrects more wrongful decisions than a single protected appeal, and that advantage holds across every cognitive setting tested. The cost side is less stable. Once reviewer errors correlate above roughly 0.30, or reviewer noise rises above roughly 0.20, plural review lets more unjustified challenges through than the single appeal does. The model's defaults (0.25 and 0.18) sit just under both lines.

v0.3 replaces those assumed values with measured ones. Language models take bounded advocate and reviewer roles against a frozen case bank and scoring scheme, and the experiment records representation fidelity, proxy substitution, reviewer error and cross-model correlation. Then a simple question: where do real model panels sit relative to those thresholds?

On what users can know about the limits, applicability and warrant of an AI output before they rely on it. Accepted for publication in AI & Society.

A reproducible behavioural baseline for dependency-capture failure modes in human-facing AI assistants, with scored artefacts, run records, comparison material and provenance.

An economics paper using Iain M. Banks's Culture as an institutional thought experiment, on AI-mediated prediction, preference formation and the demand side of planning.

Essay on how a supportive AI system can quietly become the only room a person has left.