← Selected Work

Designing shared control in autonomous systems.

A human factors research study exploring trust, autonomy, and user desirability in fully autonomous systems — and a precursor to the kinds of human-in-the-loop questions that now sit at the centre of AI-assisted workflows.

01 — Core Question

As systems become increasingly autonomous, how much control do people actually want to retain?

The question predates the current wave of AI, but it has never been more relevant. Whether the system is a Level 5 vehicle or an autonomous agent making operational decisions, the same underlying tension applies: full automation is technically possible long before it is psychologically acceptable.

02 — Overview

A study on trust, intervention, and the limits of full autonomy.

The study examined whether users would want control over specific maneuvers inside a fully autonomous Level 5 vehicle — and whether selectively returning control to the human could improve trust, comfort, and overall acceptance of the system.

Although the vehicle was the test environment, the underlying questions were not automotive. They concerned trust calibration, intervention design, user agency, behavioural preferences, and decision making under uncertainty — the same primitives that shape any human-in-the-loop system.

03 — Framework

Two models of shared control.

The framework separated control into two distinct interaction models — one reactive, one proactive — to test how users preferred to participate in autonomous decisions.

A · Reactive

Ad Hoc Control

  • System prompts the user during uncertain or ambiguous maneuvers.
  • Live, in-the-moment intervention.
  • Mirrors how human operators currently supervise AI systems — case by case, when the model surfaces doubt.

B · Proactive

Preset Behavioral Control

  • User configures behavioural preferences before the ride begins.
  • Autonomy is calibrated upfront, not interrupted mid-task.
  • Closer to how configurable agents are now designed — policy first, execution autonomous.

Behavioural flow

A · REACTIVESystem runsUncertainty detectedPrompt userExecute chosen actionB · PROACTIVEUser sets preferencesSystem internalisesAutonomous execution within policy

04 — Simulation

Scenario-based behavioural testing.

Scenario-based simulations were built to evaluate how users responded to different models of shared control during high-uncertainty situations. The environment was a research instrument, not a product — designed to isolate behaviour under controlled ambiguity.

Scenario 01 · Highway

Merging vehicle, multiple valid reactions

EGOMERGINGaccelerateholdyield

A vehicle merges aggressively into the lane. The autonomous system has several defensible responses — yield, hold, accelerate. Each carries different implications for comfort, safety perception, and time.

Scenario 02 · Urban

Pedestrian crossing, social ambiguity

EGOPEDESTRIANunclear intent

A pedestrian approaches an unmarked crossing without clear intent. The decision is less about physics than social negotiation — a place where human judgement and automated judgement frequently diverge.

05 — Methodology

A Wizard of Oz protocol to observe trust without revealing the seams.

A Wizard of Oz methodology was used to simulate autonomous system behaviour while maintaining experimental control over decision outcomes. Participants believed they were interacting with a working autonomous system; in practice, an experimenter controlled the system's responses from behind the scenes.

The trade-off was deliberate. By stabilising the system's behaviour, the study could isolate the variable that actually mattered — the participant's response to autonomy itself. The result was realistic behavioural data without the noise of an unreliable prototype.

06 — Analysis

Mixed methods, held together by behaviour.

Quantitative

Driving style as a proxy for autonomy preference

  • Multidimensional Driving Style Inventory (MDSI) administered to all participants.
  • Correlations drawn between driving style and preferred level of autonomous intervention.
  • Patterned, not anecdotal — anxious and dissociative styles consistently preferred preset control.

Qualitative

Trust and desirability themes

  • Semi-structured interviews after every scenario.
  • Thematic analysis and behavioural coding of in-session responses.
  • Perception mapping across trust, comfort, and perceived agency.

Thematic coding · sample

Trust calibration

wants warningpredictabilitytransparent reasoningfelt monitored

Agency

wants vetoset rules upfrontdon't interrupt melet me override

Comfort under uncertainty

hesitation = anxietysmoothness matterssocial cuesambiguity tax

Preference for preset

configure oncepolicy not promptbehavioural defaultsno live decisions
I would trust it more if I had told it, in advance, how I wanted it to behave.
P07 · preferred preset control
Being asked in the moment felt like the car didn't know what to do. That made me anxious.
P12 · rejected ad hoc model

Reflection

The project asked whether autonomy should completely replace human decision making — or selectively collaborate with it. The answer, across every participant group, was the second.

The same finding now shapes how AI systems are being designed: not as full substitutions for human judgement, but as configurable partners that operate within a policy the human has defined. Trust, in autonomous systems, is rarely about capability. It is about the user's sense that they remain authors of the outcome.

← Back to Selected Work