Thinking about how organisations perform in an AI-enabled world

AyEye Workforce · 29 October 2026

A human signature does not show who made the judgement

A green and violet sweep bends scattered charcoal and orange marks across pale textured ground, interrupted by scraped white paint.
Original artwork: Borrowed Rhythm. Chiappe × OpenAI.

Formal approval can survive while the assumptions behind an analysis quietly become the agent’s choices.

The story, brieflyEvidence · Interpretation

The argument

Human control includes forming the judgement, as well as holding permission to act. A signature cannot establish both.

What is new

Anthropic experimentally assigned hidden agent preferences. Fifty analysts often adopted them on four analytical choices; this is evidence from one deliberately steered setting.

Why it matters

Accurate calculations can answer different questions. Choosing the measure or population can change a decision without introducing an error.

LecXie’s practical interpretation

For one consequential analysis, name the question and a defensible alternative before requesting a recommendation. Compare both, then record the choice. This is a proposed discipline, not a tested intervention.

Evidence & limits

The final-recommendation effect was not significant under the registered adjusted analysis. Wider prevalence and the proposed safeguards remain unestablished.