A care robot's joke may delight someone on the first day and irritate them by the sixth week. What would keep it useful after the novelty wears off? Answering that requires looking at how it fits the tasks, expectations and daily lives of the people using it.

01

The first smile is an easy test; the hundredth day is harder

A robot that turns its head, remembers a name and replies politely may delight visitors at a demonstration. The harder question is whether anyone needs it after the novelty wears off. Emily Cross and Arvid Kappas argue that social robotics has often treated the social part as a feature to add to a machine rather than a phenomenon to understand. Commercial failures and short-lived engagement show why those relationships need to be studied during everyday use. A socially convincing gesture is not enough if the device cannot do useful work, sustain an interaction or fit into the routines of the people meant to use it.

The review follows two promises made by social robots: they will read human states and they will express responses that people understand. Both promises depend on psychology. A camera output labeled 'sad' is a measurement claim about a person; a friendly-looking movement is a design choice that changes expectations. The authors ask engineers and psychologists to study the interaction itself, across varied users and real settings, rather than judging success solely by whether a prototype works for a few minutes in a laboratory.

02

A face is not a transparent emotion meter

Many emotion-recognition systems classify facial movements using a small list of supposedly universal basic emotions. The review traces that idea to mid-twentieth-century theories and explains why it is a fragile foundation for current robots. A facial movement can depend on context, culture, what a person wants to communicate and what they are actually feeling. The authors do not deny that expressions carry information; they challenge the shortcut from a visible configuration to a single inner state. A robot that treats a smile as unambiguous happiness may respond badly when the smile is polite, nervous or ironic.

This is also a problem of evidence. Training a classifier to match labels assigned to photographs does not show that the labels accurately describe people in ongoing interactions. Real encounters include speech, timing, relationship history and changing goals. If a robot acts on an uncertain emotional inference, its mistake can alter the very interaction it is trying to understand. The review argues for contemporary, context-sensitive theories of emotion and for testing in situations where people do more than pose a face for a camera.

03

What a robot expresses depends on its body

Recognizing a human expression and producing a readable response are different engineering and psychological tasks. A text interface can change an avatar's face instantly; a physical robot is limited by motors, materials, size and movement. Designers may assume that looking increasingly human will make a machine easier to relate to. Cross and Kappas describe why that assumption can backfire: near-human appearance paired with stiff or mistimed movement can create unease, the pattern often called the uncanny valley. A device need not imitate a person closely to invite social interpretation.

Research on moving abstract shapes, discussed in the review, shows how readily people attribute intentions to minimal cues. That observation suggests a design question more useful than 'How human can we make it?': which social signals serve this task and make the machine's capacities understandable? A small expressive movement may suffice for one context, whereas another requires clear speech or precise physical assistance. Matching appearance to actual ability also avoids promising an intimacy or competence the robot cannot deliver.

04

Connection needs a purpose beyond novelty

The authors point to Paro, a seal-shaped robot used in some care settings, as a limited but instructive success. Its small movements and sounds in response to touch can invite engagement within a defined therapeutic situation. That does not establish that every person needs a robot companion or that a device replaces human care. Paro has a particular role, audience and setting. The example helps separate a useful social interface from the vague promise of a general-purpose mechanical friend.

To assess that usefulness, researchers need to examine what people actually do in their homes, schools and care institutions. A robot's social behavior should help it perform a valuable function, not be its sole reason to exist. In an editorial illustration, a device that politely reminds someone of a task matters only if the reminder is timely, wanted and integrated with the task; a charming voice cannot rescue poor timing. Long-term studies are needed to learn whether engagement persists and whether the robot adds benefit after the demonstration ends.

05

A hidden human operator can distort the promise

Some laboratory robots appear autonomous because a person behind the scenes selects their responses. This Wizard-of-Oz technique is legitimate for exploring how users might react to capabilities that do not yet exist. The problem begins when evidence from that simulation is read as evidence that an autonomous machine can already deliver the same interaction. The review discusses estimates that a substantial portion of human–robot interaction research has involved some level of human operation or supervision. The gap between demonstrated and deployable ability is therefore not a footnote.

A human operator can understand context, recover from misunderstandings and time a gesture in ways current software may not. If participants believe they are meeting an independent robot, the method also raises questions about disclosure and deception. Cross and Kappas do not reject prototype experiments; they ask investigators to state what was simulated and to test whether the promised behavior survives without the operator. Otherwise a company, policymaker or user may buy expectations created by a person concealed inside the experiment.

06

Fluent conversation can raise the wrong expectation

Generative AI can make a robot speak more naturally and respond across a wider range of topics. But a fluent answer is not the same as social understanding or physical competence. A machine may discuss a plan persuasively while failing to grasp a gesture, move safely in a crowded room or notice that the user no longer wants to talk. The review argues that the conversational sophistication of large language models can heighten expectations that the robot's body and autonomy cannot meet. Psychology becomes more necessary, not less, when the surface of the interaction becomes convincing.

A sustained dialogue also differs from answering isolated questions: participants take initiative, reveal intentions and adjust to one another. The authors call for models of interaction that account for those turns, not only for generating plausible sentences. More capable systems introduce practical burdens as well—computing costs and privacy questions when cameras, microphones and personal histories are used. None of these concerns means that generative models have no role. They mean that a better language engine cannot, by itself, solve the human and bodily problems of social robotics.

07

There is no average person waiting for a robot

How people respond to a robot varies with age, culture, personality, experience and cognitive capacity. The review describes findings that more extraverted users in some studies anthropomorphized robots more readily, while other relationships, including links between personality and trust, were inconsistent. Such variation cannot be settled by a demonstration with a small, convenient sample. Robust tests of individual differences often need far more participants than traditional human–robot interaction studies recruit, and they need settings in which the device will actually be used.

A design that works for one group may be distracting, inaccessible or unwanted for another. The authors call for closer collaboration between engineers and psychologists so that theories, measurement and larger research programs shape the machine from the beginning. Psychology is not a final usability test applied after the hardware is built. It helps define what counts as a social cue, how to measure a relationship with a device and whether a reported benefit generalizes beyond the original sample.

08

Useful, honest and sustainable are social requirements too

The review notes that many companies behind widely studied social robots have struggled financially. Embodied hardware is costly, and a machine that only chats or entertains may lose its place when novelty fades. This is not a simple market prediction about every future robot. It is evidence that social appeal cannot substitute for a well-defined task, a realistic price and maintenance over time. The authors also discuss changing regulation, transparency and the risk that short-lived devices become electronic waste.

The most constructive question is where social intelligence improves a function that already matters. A care device, educational tool or workplace machine might benefit from better timing, communication and adaptation without pretending to be a human replacement. Whether it does so is an empirical question involving users, staff, costs, privacy and long-term outcomes. Cross and Kappas' conclusion turns the usual slogan around: building a social machine starts with a serious account of people and their situations. The machine is only one participant in that larger design problem.

That requires a different research rhythm from a spectacular prototype. Teams need to specify the purpose of the device, test the psychological assumptions behind its social cues, observe use over time and report when a human operator supplied the behavior. They also need to ask what happens when a system fails or a user disengages. A robot's social claim should be judged by its actual contribution to an interaction, not by the impression that its surface can create for a moment.

What to keep

  1. Social intelligence is relational, not only computational.
  2. Embodiment changes expectations.
  3. Culture and roles shape human–robot interaction.
?

The nuance

What this story does not let us claim

Short laboratory encounters can overestimate novelty and understate the difficulties of long-term use in real settings.

Go deeper: concepts and optional reading practice

Reversible cards

Test the concept

1 / 2

Topic assessment

Check whether you can explain the idea, not merely recognize it.

Every answer is grounded in the explanation above. Some questions introduce new cases to test whether you can apply and analyze the ideas, not merely recall them.

7 questions4 options per questionOne correct answer
01Which statement best captures the review’s central thesis?
02Which is the most appropriate operational definition of “Anthropomorphism”?
03Which case correctly applies the central concept?
04If two studies obtain different results, which interpretation best preserves the logic of this guide?
05Which complementary conclusion belongs to the paper’s core argument?
06Which formulation best distinguishes “Embodiment” from a superficial label?
07Which limit should accompany any application of the paper to a new case?
Answer all seven questions.

Source & scope

Social Robotics Is Not (Just) About Machines, It Is About People: Psychology's Role in Developing Social Machines. Emily S. Cross & Arvid Kappas. Annual Review of Psychology, vol. 77 (2026).

This editorial guide is an independent explanation. Consult the original publication for the studies, references and full context.

Open the original review ↗