Skip to contentSIMULGENCEQuantum LabLab

AI power / Defaults and delegated choice

How AI defaults could shape inferred preferences

AI assistants could interpret accepted defaults as personal preferences, then reinforce their own earlier assumptions. Limited attention makes acceptance an ambiguous signal.

Enter the explanation
A mechanical selection gate and optical recorder linked by a circular conveyor and return channel. An illustration of the hypothetical feedback between AI defaults, acceptance and inferred preferences.Mechanism / Evidence / Constraints
Simulgence3 minute read Position analysisExplore AI

Decisions made before attention arrives

An AI assistant could help produce a person's choices, learn from those choices, and describe the result as personalisation. The risk begins when acceptance is treated as evidence of preference.

The attached bottle cap makes a quieter form of influence visible. EU rules changed the design of certain beverage containers to reduce plastic litter. A cap can be deliberately detached; leaving it attached requires less intervention. This is physical design rather than a preselected setting, but both arrange what happens without further effort.[1]

Governments can shape that starting point. The UK's Children's Code calls for high privacy settings by default, recognising that many children will never change them.[2] A policy can protect people through their limited attention; a commercial setting can allocate that same inattention differently. Outcomes and beneficiaries require examination in each case.

Every request to reconsider a setting draws on the same finite attention. An institution sees its own manageable choice. The individual encounters the combined demand.

Acceptance can become evidence

Defaults can influence evaluation as well as effort. Experiments in The Golden Halo of Defaults in Simple Choices found that preselection could increase an option's perceived value, with effects limited by context and existing goals. Thinking about a choice does not necessarily remove the influence of its starting condition.[3]

AI extends the question into delegated action. In May 2026, Google announced agentic purchases governed by user-specified criteria.[4] Such delegation can save time, but instructions leave trade-offs unresolved.

Consider a hypothetical travel assistant. It favours convenient hotels; a busy user accepts them. If the system records acceptance as a preference for convenience, later recommendations may repeat the original assumption. Cancellation rights, privacy or neighbourhood variety can disappear from consideration without ever being explicitly rejected.

The loop is a proposed mechanism, not a finding that every assistant behaves this way. Acceptance can reflect genuine satisfaction. The problem is distinguishing that from fatigue, trust or the absence of a visible alternative.

Power moves upstream

No single coordinator is necessary. Regulators set requirements, platforms select available services, developers configure ranking, and users delegate. Each decision may have an identifiable author while the combined direction remains difficult to attribute.

Some people inspect alternatives; others reject a default to demonstrate independence or exploit an exception for personal advantage. These actions have different motives. The power to change one personal setting remains different from the power to set the starting conditions for millions.

A useful assistant could earn enough trust that its assumptions receive less scrutiny. Individually tailored choices could also make common influence harder to recognise: different users would encounter different recommendations rather than a shared rule.

Delegation can return attention while transferring discretion. The consequential distinction is whether an assistant remembers what a person deliberately chose, what it selected on their behalf, and what they merely did not correct. A system that merges those records risks learning a person partly of its own making.

Sources & further reading

Follow the evidence.

  1. 01
    European Commission: Less plastic waste means cleaner beaches (14 August 2024)

    The stated purpose of the cap design requirement; not an estimate of its effect.

  2. 02
    ICO: 7. Default settings

    Children's Code guidance includes a compelling-reason exception taking account of the child's best interests.

  3. 03
    Sullivan et al.: The Golden Halo of Defaults in Simple Choices (2025)

    First published online 22 November 2024. Simple-choice experiments; does not establish the proposed AI feedback loop.

  4. 04
    Google: Introducing the Universal Cart and more ways to help you shop (19 May 2026)

    Vendor announcement of delegated purchases; not independent evidence of preference formation or universal availability.