Skip to content
Nornic

About the research

Prediction, explanation, influence — and the line between support and manipulation

Consumer AI is very good at predicting what people will do. It is far less clear whether it can help people decide well. This project separates three goals that are routinely conflated, and studies the third — influence — with an unusual objective: the decider’s own judgment and welfare, not a conversion metric.

Prediction, explanation, and influence

Diagram of three separated goals — prediction (what will the person choose?), explanation (why might they choose it?), and influence (how should AI support?) — with human agency and consumer welfare constraining the entire system.
Fig. 01Three goals that are routinely conflated — and the constraint that binds all of them.
01

Prediction

What will the person choose?

Forecasting behavior from data. Powerful, increasingly commoditized — and silent about whether the predicted choice is good for the person making it.

02

Explanation

Why might they choose it?

Understanding the reasons behind a choice. An explanation is only useful if it is faithful — a true account of the calculation, not a plausible story told after the fact.

03

Influence

How should AI support the decision ethically?

Changing the decision process itself. The ethical line: support engages a person’s deliberation, while manipulation bypasses it. This project lives entirely on the support side of that line.

Human-centered AI

A different objective function

Conventional consumer AI optimizes the firm’s metric — clicks, conversion, retention, revenue. A human-centered system is accountable to the decider instead: objective decision quality, understanding, autonomy, satisfaction, and reduced regret. Four commitments follow from that switch.

Human-centered AI framework: prediction, explanation, and influence share one ethical boundary between support and manipulation; the framework commits to keeping the human in command with faithful explanation, consumer welfare, decision quality, and human agency; conventional AI optimizes clicks, conversion, and revenue while human-centered AI optimizes decision quality, understanding, autonomy, satisfaction, and reduced regret.
Fig. 02The framework in one figure: one ethical boundary, four commitments, and a different set of things worth optimizing.

  • Human in command

    The system recommends; the person decides. Overriding the AI is always easy, visible, and free of friction or penalty.

  • Faithful explanation

    When reasoning is shown, it is the actual computation — inspectable arithmetic over stated inputs, never a post-hoc rationalization or fabricated confidence.

  • Consumer welfare as the outcome

    Success is measured by whether the person chose well for their situation, not by whether they complied with the system.

  • Agency measured, not assumed

    Perceived autonomy, perceived manipulation, and override behavior are recorded as first-class outcomes alongside decision quality.

The study, at a high level

A randomized, four-condition decision experiment

Randomized, between-subjects, four-condition online experiment. Participants receive a realistic mobile-service usage profile and choose one of six plans. The objectively best-value plan can be computed from the profile, which makes decision quality measurable rather than assumed.

  • Arm 01

    Human alone

    The participant decides with the plan information only — no recommendation of any kind.

  • Arm 02

    Opaque AI

    A recommendation is shown without any reasoning.

  • Arm 03

    Explainable AI

    The same recommendation, together with the calculation behind it, open for inspection.

  • Arm 04

    Reflective AI

    The system first asks what the participant considers important, then recommends a plan aligned with those stated priorities.

Where a recommendation appears, all conditions share the same deterministic computation, and every screen is visually identical except for the assigned form of support. We publish directional hypotheses and analysis plans through preregistration — not on participant-facing pages — to avoid biasing behavior.

What is measured

Primary outcome. Objective decision quality — the normalized cost gap between the selected plan and the best eligible plan for the assigned profile.

  • Comprehension
  • Decision confidence
  • Trust in the recommendation
  • Perceived autonomy
  • Perceived manipulation
  • Post-choice regret
  • Satisfaction
  • Recommendation acceptance and override behavior
  • Time on task and information inspection
  • Preference alignment

Privacy & consent

Consented, minimized, deletable

What consent covers

  • Purpose of the prototype and the procedures involved
  • Exactly what data is collected, and for how long it is retained
  • How to withdraw during the session and how to request deletion afterwards
  • That no deception is used, and that no external AI service receives your data
  • A contact for questions or concerns

How data is treated

  • Anonymous study identifiers — no accounts, names, or emails in study data
  • Only protocol-required data is collected; nothing is reused for advertising
  • No advertising pixels, session replay, or marketing analytics on study routes
  • Recommendations are computed locally by a deterministic engine — no personal data is sent to an external model
  • Every stored record is versioned so results stay auditable and reproducible

This is an engineering prototype. Public recruitment and any claims of research findings will only follow the applicable ethics review and preregistration.