🤖 AI

Psychological and behavioural responses in human-agent vs. human-human interactions: a systematic review and meta-analysis

This systematic review and meta-analysis of 162 studies reveals that while humans exhibit comparable functional behaviors and task performance when interacting with agents versus other humans, they demonstrate significantly reduced prosocial behavior, moral engagement, and social attributions toward agents, suggesting that agents are viewed as instrumentally equivalent but intrinsically inferior social partners.

Original authors: Jianan Zhou, Fleur Corbett, Joori Byun, Talya Porat, Nejra van Zalk

Published 2026-05-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jianan Zhou, Fleur Corbett, Joori Byun, Talya Porat, Nejra van Zalk

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a party. You have two conversation partners: one is a real human, and the other is a very advanced robot or computer program that acts exactly like a human. They both tell jokes, listen to your stories, and play games with you perfectly.

The big question this research team asked was: Does your brain and your behavior react the same way to the robot as it does to the human?

To find out, they didn't just guess; they acted like detectives. They gathered and analyzed 162 different scientific studies (involving hundreds of experiments) to see how people actually behave when talking to machines versus people. They looked at everything from how much you trust them to how much you feel guilty if you lie to them.

Here is what they found, broken down into simple concepts:

1. The "Tool" vs. The "Friend" (Where they are the same)

When it comes to getting the job done, your brain treats the robot and the human almost identically.

The Analogy: Think of the robot as a very skilled co-pilot. If you are flying a plane (or solving a math problem) together, you trust the co-pilot's calculations just as much as a human's.
The Finding: People perform tasks just as well with agents as with humans. They trust the agent to do the job, they align their own behavior to match the partner (like mirroring body language), and they feel just as much control over their own actions. If the robot is good at the task, you treat it as a reliable tool.

2. The "Moral Gap" (Where they are different)

However, when it comes to feelings, morals, and kindness, there is a big difference.

The Analogy: Imagine you are playing a game where you have to share a pile of candy. If you are playing with a human, you might share half because you feel bad if they get none. If you are playing with a robot, you might keep almost all the candy for yourself.
The Finding: People are significantly less kind and less morally engaged with agents. They are less likely to help, less likely to feel guilty about hurting the agent, and less likely to feel a sense of shared responsibility.
The "Vibe" Check: People also perceive agents as less "real." They feel the robot is less likeable, less competent, and less "present" in the room. It's like talking to a very convincing mannequin; you know it's there, but you don't feel the same warm connection you do with a real person.

3. The "Context is King" (It depends on the situation)

The researchers found that these differences aren't always the same; they change based on the setting, like a chameleon changing colors.

The "Look" Matters: If the robot looks exactly like a human (same voice, same face), people like it more and treat it more like a person. If it looks obviously fake, the "robot gap" gets bigger.
The "Room" Matters: If you are in a video call (computer-mediated), the difference between robot and human shrinks. But if you are face-to-face with a physical robot, people notice the difference much more and treat the robot with less respect.
The "Game" Matters: If you are competing against someone (like in a game), you might actually be happier playing against a robot than a human. But if you are trying to cooperate or learn something, the human connection feels more natural.

The Bottom Line

The paper concludes that we have a split personality when dealing with AI:

As a Partner: We treat agents as excellent tools. We trust them to do the work, we work well with them, and we feel just as capable.
As a Person: We treat agents as lesser beings. We don't feel the same moral duty to them, we don't feel as connected, and we don't attribute the same "soul" or responsibility to them.

In short: We will happily let a robot drive our car or write our report, but we probably won't invite it to our wedding or feel bad if we accidentally break its heart. The robot is a great partner for the task, but not yet a peer for the relationship.

Problem Statement
Interactive intelligent agents (e.g., chatbots, virtual humans, robots) are increasingly integrated into professional and personal domains, evolving from reactive tools to proactive, autonomous agents. While these systems achieve human-level performance on specific tasks, it remains unclear whether they elicit the same psychological and behavioural responses as human partners. Existing empirical evidence is fragmented across disciplines (computer science, psychology, communication, business) and methodological traditions, yielding mixed findings. Some studies suggest equivalence in responses (supporting the "Computers Are Social Actors" paradigm), while others indicate deficits or distinct responses (challenged by frameworks like the Threshold Model of Social Influence or Uncanny Valley Theory). Previous reviews have either prioritized theoretical argument over systematic synthesis, focused narrowly on specific partner types (e.g., avatars vs. agents), or treated diverse psychological and behavioural responses as a monolithic entity. Consequently, there is a lack of a holistic, evidence-driven understanding of how human responses in Human-Agent Interaction (HAI) compare to Human-Human Interaction (HHI) when task performance is matched.

Methodology
This study conducted a systematic review and meta-analysis following the PRISMA 2020 guidelines.

Data Collection: A systematic search was performed across Scopus, Web of Science, ACM Digital Library, and PsycInfo (2000–2024). From 11,456 records, 162 eligible studies (146 contributing to the meta-analysis; 468 effect sizes) were selected.
Inclusion Criteria: Studies must involve healthy adult participants in dyadic interactions comparing performance-matched agent vs. human partners. The interaction task, dynamics, and partner role/performance had to be consistent across conditions, isolating the partner identity (agent vs. human) as the primary variable.
Classification: An empirical-to-conceptual approach was used to inductively classify 23 distinct human response types into six themes: (1) Prosociality and morality, (2) Social perceptions of interaction partners, (3) Trust in interaction partners, (4) Social alignment, (5) Personal agency and task performance, and (6) Interaction experiences.
Analytical Approach: The study employed both frequentist and Bayesian meta-analytic approaches using random-effects models (three-level to account for multiple effect sizes within studies). Hedges' g was used as the effect size metric. Heterogeneity was assessed via $Q$ and $I^2$ statistics. Univariate meta-regressions were conducted to explore moderators (study, participant, partner, scenario, and measurement characteristics) for response types with sufficient data ( $k \geq 10$ ) and significant heterogeneity.
Quality and Bias: A tailored 23-item quality checklist was developed to assess study design rigour, data/reporting rigour, and research integrity. Publication bias was assessed via funnel plots and Egger tests. Sensitivity analyses (outlier removal, alternative priors, approximated effect sizes) confirmed result robustness.

Key Contributions

First Holistic Synthesis: This is the first meta-analysis to systematically compare a broad spectrum of psychological and behavioural responses in HAI vs. HHI using performance-matched partners.
Differentiated Response Mapping: The study moves beyond monolithic "social influence" measures, distinguishing between specific response types (e.g., separating moral engagement from task performance, or subjective trust from behavioural trust).
Dual-Method Validation: By integrating frequentist and Bayesian approaches, the study provides complementary evidence, quantifying the strength of evidence for both the presence and absence of effects (via Bayes Factors).
Moderator Identification: The research identifies specific contextual and design factors (e.g., interaction medium, agent form, task nature) that moderate the differences between HAI and HHI, highlighting the context-dependency of partner effects.

Results
The meta-analysis revealed a nuanced landscape of similarities and differences:

Significant Differences (Agent Disadvantage):
- Prosociality and Morality: Individuals exhibited significantly less prosocial behaviour (e.g., sharing in dictator games) and lower moral engagement (reduced moral acts, intentions, guilt) with agents compared to humans.
- Social Perceptions: Agents were perceived as significantly less socially present, less likeable, and less competent. Crucially, agents were attributed significantly less agency (capacity for intentional action) and responsibility (accountability for outcomes).
Significant Similarities (Parity):
- Trust and Alignment: Behavioural trust, subjective trust (frequentist), social alignment (synchrony, advice-taking), and perceived self-agency showed no significant differences between agents and humans.
- Performance and Experience: Objective task performance, strategic economic behaviour, and most interaction experiences (satisfaction, future intention, naturalness, enjoyment, workload, engagement) were comparable.
Moderators:
- Appearance and Medium: The "agent disadvantage" in perceived likeability and competence was significant when appearance differed or in face-to-face/physical settings but disappeared or diminished when appearances matched or in computer-mediated/virtual settings.
- Interaction Realism: Differences in agency and responsibility attribution were significant in online/hypothetical studies but non-significant in lab/real-time studies.
- Task Nature: Interaction satisfaction was higher with agents in oppositional interactions but comparable in cooperative ones.
- Operationalisation: Perceived relational qualities (e.g., customer orientation) were lower for agents in hypothetical/vignette studies but comparable in real-time interactions.

Significance and Implications
The paper claims that while agents can achieve instrumental value on par with humans (in terms of task performance, trust, and alignment), they currently lack intrinsic value comparable to humans (in terms of moral engagement, prosociality, and social attributions like agency and responsibility).

Theoretical: The findings suggest that social responses should be treated as distinct constructs rather than a monolithic entity. The results support a "Threshold Model" where automatic, low-level reactions (like task performance) may be similar, but deliberate, high-level social attributions (like morality and responsibility) diverge.
Design and Regulation: The study implies that agents are effective collaborators for goal-directed tasks but require careful design in morally sensitive domains (healthcare, education). Designers must address the "responsibility gap" by creating clear accountability structures. Furthermore, the high heterogeneity in subjective responses suggests that agent disadvantages are malleable; strategic design (e.g., matching appearances, optimizing interaction contexts) can mitigate deficits in social perception.
Future Research: The review establishes a baseline for the "pre-GenAI era," noting that most included studies used traditional or Wizard-of-Oz agents. It calls for future research to re-examine these effects with Generative AI and to explore sustained, naturalistic interactions beyond one-time lab studies.

1. The "Tool" vs. The "Friend" (Where they are the same)

2. The "Moral Gap" (Where they are different)

3. The "Context is King" (It depends on the situation)

The Bottom Line

More like this