Skip to content
PaperFren

Can an AI teammate guide you without making you feel bossed around?

Open paper intelligence

An agent that hinted at the best goal through its own movements helped people pick the better target as often as an agent that told them outright, while leaving them feeling more in control.

Source

Balancing Performance and Human Autonomy With Implicit Guidance Agent

Nakahashi R, Yamada S · Frontiers in artificial intelligence · 2021

doi.org/10.3389/frai.2021.736321Read the full paper ↗1 citationscc by

Study at a glance

Design
Human experiment — Within-subject online experiment: each participant played a pursuit-evasion maze task with a supportive, an explicit-guidance and an implicit-guidance agent in randomised order.
N
N=100 · 100 crowdsourced adults recruited; 3 excluded for failing dummy tasks. Each did 15 regular tasks (5 per agent).
Population
Adult crowdworkers in Japan (Yahoo! Crowdsourcing)
Outcome
Rate of capturing the best target; 7-point ratings of perceived autonomy, ease of collaboration and intention inference

Structured fields used in claim comparison tables when every cited study has a complete layer.

What they did

The authors modelled a human-agent game as a POMDP in which the hidden state is the human's target. They built three agents: a supportive one that simply follows the human's goal, an explicit one that shows the best target, and an implicit one that moves so the human, reasoning via a Bayesian Theory of Mind, can infer the better target. Online participants played maze tasks where the participant and the agent had to trap a fleeing object from both sides, rating each agent afterwards.

What they found

Agent type strongly affected how often participants captured the best object (F(2, 968) = 79.9); both guidance agents far outperformed the supportive agent, and implicit and explicit guidance did not differ significantly. Perceived autonomy also differed by agent (F(2, 192) = 36.4), with participants feeling more in control with the implicit agent than the explicit one. Explicit guidance was also rated harder to collaborate with, possibly because its interface showed extra information.

The limits

What it doesn't show

The mazes were tiny, with only two possible targets and a small discrete action space that made the agent's intentions easy to read, so the advantage may not hold in complex or continuous tasks. The agent assumes every human uses the same fixed, rational cognitive model and trusts the agent. Autonomy was measured by a single general, subjective Likert item, and the crowdsourced online sample was mostly male.

Key terms

POMDP
Partially observable Markov decision process: a planning framework where the agent must act while some of the state, here the human's goal, is hidden.
Bayesian Theory of Mind
Modelling how an observer infers another agent's goal by Bayesian updating on its observed actions.
Implicit guidance
Steering a partner by acting so they can infer the better plan themselves, rather than telling them directly.
Boltzmann rationality
An assumption that agents choose better actions more often, with probability rising exponentially with an action's value.
Within-subject design
Every participant experiences all conditions, so comparisons are made within the same person.

Flashcards

1 / 10

0 of 10 answers reviewed

Research intelligence for this paper

See its role on concept claims, tensions it is part of, placement history, and related discoveries.

Open paper intelligence

Quiz yourself

1 / 5

Which agent assumed the human would infer its target from its movements?

Common questions

If implicit and explicit guidance gave the same success rate, why prefer implicit?

Because participants felt more autonomy with the implicit agent, so it achieved similar performance at a lower cost to the human's sense of control.

Did participants know the agent was hinting?

No; implicit guidance was not explained to them, yet they inferred the agent's intended target from its movements anyway.

Why was the supportive agent so poor?

It just followed the human's chosen target, and the tasks were designed so people often could not tell which target was actually catchable.

More on Human-AI interaction