Computers-as-social-actors (CASA) experiment

Description

Take a human-human social psychology finding, substitute the computer for one person, and run a controlled experiment varying only the social cue to see if users respond socially.

The CASA paradigm tests whether people apply human social rules to interfaces. A well-established social effect (similarity attraction, in-group favouritism, response to praise, reciprocity, politeness) is replicated with a computer in place of one person. Conditions differ in a single social cue, such as dominant versus submissive text style, a 'teammate' label with matching colours, sincere versus random praise, or helpful versus unhelpful prior results, while information content and function stay identical. Stanford studies used samples of around 48 to 76 participants, pre-screened moderators such as personality, and controls such as an identical but different machine. Measures combine questionnaire ratings (competence, liking, credibility) with a behavioural outcome such as adopting the computer's ranking or effort spent on a favour. Today it is the natural way to test personas and social framing in chatbots, voice assistants and AI agents.

Attributes

Loss type
Siloed
Effort
Weeks

Steps

  1. Choose an established human-human social effect and state the prediction for the interface or agent.
  2. Pick a task that gives the computer a believable role and a measurable behaviour, such as a ranking task where users can adopt or reject its advice, or an open-ended favour task.
  3. Create conditions that differ in one social cue only (text style, labels, colours, praise, helpfulness) while holding information and functionality constant.
  4. Pilot the stimuli so manipulations land as intended and messages do not repeat; pre-screen or measure moderators such as personality and cross them with interface type.
  5. Add controls that rule out other explanations, such as an identical but different machine or praise disclosed as random.
  6. Randomly assign participants (the classic studies used 48 to 76) and run sessions with a consistent cover story and minimal experimenter influence.
  7. Measure attitudes (competence, liking, credibility) by questionnaire and at least one behavioural outcome (recommendations adopted, effort, cooperation).
  8. Analyse condition differences and debrief participants about any cover story.

Sources

  • Persuasive Technology (Fogg, 2003)

Also known as

  • CASA paradigm
  • Media equation experiment
  • Social response to computers study
  • Substitute-the-computer replication

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference