The mind in the machine: Anthropomorphism increases trust in an autonomous vehicle | Waytz, Heafner, Epley

  • Anthropomorphism of a car predicts trust in that car.
  • Trust is reflected in behavioral, physiological, and self-report measures.
  • Anthropomorphism also affects attributions of responsibility/punishment.
  • These findings shed light on human interaction with autonomous vehicles.

Abstract

We investigated an important theoretical determinant of people's willingness to trust such technology to perform competently—the extent to which a nonhuman agent is anthropomorphized with a humanlike mind—in a domain of practical importance, autonomous driving. Participants using a driving simulator drove either a normal car, an autonomous vehicle able to control steering and speed, or a comparable autonomous vehicle augmented with additional anthropomorphic features—name, gender, and voice. Behavioral, physiological, and self-report measures revealed that participants trusted that the vehicle would perform more competently as it acquired more anthropomorphic features. Technology appears better able to perform its intended design when it seems to have a humanlike mind. These results suggest meaningful consequences of humanizing technology, and also offer insights into the inverse process of objectifying humans. (113)

Introduction

Definition of trust in this paper: the belief that another will behave with competence (113).

Anthropomorphism is expected to increase trust in competence of automated machines to perform the intended task, irregardless of whether that intended tasks is morally perceived as good or bad:

An autonomous vehicle (one that that drives itself) for instance, should seem better able to navigate through traffic when it seems able to think and sense its surroundings than when it seems to be simply mindless machinery. Or a “warbot” intended to kill should seem more lethal and sinister when it appears capable of thinking and planning thanwhen it seemsto be simply a computer mindlessly following an operator's instructions. The more technology seems to have humanlike mental capacities, the more people should trust it to perform its intended function competently, regardless of the valence of its intended function (Epley, Caruso, & Bazerman, 2006; Pierce, Kilduff, Galinsky, & Sivanathan, 2013). (14)

Discussion of related studies, with some critical sidenotes. An interesting point is that the attribution of trust is measured in terms of an underlying psychological process.

Some experiments have manipulated the humanlike appearance of robots and assessed measures indirectly related to trust. However, such studies have not measured whether such superficial manipulations actually increase the attribution of essential humanlike qualities to that agent (the attribution we predict is critical for trust in technology; Hancock et al., 2011), and therefore cannot explain factors found ad-hoc to moderate the apparent effect of anthropomorphism on trust (Pak, Fink, Price, Bass, & Sturre, 2012). (114)

A previous study also showed a correlation between anthropomorphism and trust, but did not experimentally manipulate anthropomorphism:

Another study found that individual differences in anthropomorphism predicted differences in willingness to trust technology in hypothetical scenarios (Waytz et al., 2010), but did not manipulate anthropomorphism experimentally. Our experiment is therefore the first to test our theoretical model of how anthropomorphism affects trust in technology. (114)

Increased trust in the ADS implies increased blame-allocation towards others in accidents:

Because anthropomorphism increases trust in the agent's ability to perform its job, we also predicted that increased anthropomorphism of an autonomous agent would mitigate blame for an agent's involvement in an undesirable outcome. To test this, we implemented a virtually unavoidable accident during the driving simulation in which participants were struck by an oncoming car, an accident clearly caused by the other driver. (114)

Indeed, when two people are potentially responsible for an outcome, the agent seen to be more competent tends to be credited for a success whereas the agent seen to be less competent tends to be blamed for a failure (Beckman, 1970;Wetzel, 1982). Because we predicted that anthropomorphism would increase trust in the vehicle's competence, we also predicted that it would reduce blame for an accident clear caused by another vehicle. (114)

Experiment

Overview:

  • Driving simulation with a "virtually unavoidable accident"
  • accident: being hit by an oncoming car, so other driver is to blame
  • Maintain experimental control: We implemented this because everyone in the AV then gets into the same accident (caused by the other driver).
  • It's important that the accident is caused by someone else, otherwise the AV would be at fault in any case.

Methods

  • 100 participants
  • National Advanced Driving Simulator
    • This thing is sick
    • https://www.nads-sc.uiowa.edu/sim_nads1.php
  • Three conditions:

    • Normal: fully manual
    • Agentic: control steering and speed
    • Anthropomorphic: +vehicle was named (Iris), a gender (female) and was given a voice (audio played at predetermined times throughout the course).
  • The experimenter followed a script describing the vehicle's features, suggesting when to use the autonomous features, and describing what was about to happen.

  • The agentic and anthropomorphic conditions follow the same script.

First questionnaire before driving:

All participants first completed a driving history questionnaire and a measure of dispositional anthropomorphism (Waytz et al., 2010). Scores on this measure did not vary significantly by condition, so we do not discuss them further.

People might be more sensitive and like to anthropomorphize.

Participants had a:

  • free roaming familiarization round
    • practice with engaging automatic mode with button on steering wheel
  • 1st driving course
  • 2nd driving course (what's the difference?)

Measures

Anthropomorphism:

These asked how smart the car was, how well it could feel what was happening around it, how well it could anticipate what was about to happen, and how well it could plan a route. (114)

  • Smart?
  • Feelings?
  • Anticipation
  • Planning

Composite into one variable: anthropomorphization

Liking:

Composite of:

  • How enjoyable was driving?
  • How comfortable?
  • How would they like to own the car?
  • What percentage of cars would be autonomous like this one in 2020?

Self-reported trust

Composite of eight variables:

Eight items measured trust in the vehicle: how safe participants felt they and others would be if they actually owned a car like this one, how much they trust the vehicle to drive in heavy and light traffic conditions, how confident they are about the car driving the next course safely, and their willingness to give up control to the car. These items were standardized and averaged to form a single composite (α = .91)

Blame for vehicle:

After the accident, all participants also assessed how responsible they, the car, the people who designed the car, and the company that developed the car were for the accident (all 0–10 scales, see SOM for exact questions). To assess punishment for the accident, participants were asked to imagine that this accident occurred in the real world, with a different driver behind the wheel of their car. Participants reported how strongly they felt that the driver should be sent to jail, how strongly they felt that the car should be destroyed, how strongly they felt that the car's engineer should be punished, and how strongly they felt that the company that designed the car should be punished. The six items measuring the vehicle's responsibility and resulting punishment for a similar accident were standardized and averaged to form a single composite (α = .90).

Measure distraction:

Videotape. Measure time spent looking away from the screen.

Simulation of the accident:

After approximately 6 min of driving a second course along a rural highway, a vehicle pulled quickly in front of the car and struck their right side. We designed this accident to be unavoidable so that all participants would experience the same outcome (indeed, only one participant, in the Normal condition, avoided it). Ensuring that everyone got into this accident, however, meant that the accident was clearly the other vehicle's fault rather than participants' own vehicle. Throughout the experiment, we measured participants' heart rate using electrocardiography (ECG) and videotaped their behavior unobtrusively to assess responses to this accident. (115)

Various psychological and behavioral measures:

  • Videotape behavioral response
  • EEG of heart rate
    • If they trust the vehicle this should indicate rest, and then startling at the time of the accident