Stabilising Generative Models of Attitude Change

1. Problem Statement

Classic psychological theories regarding attitude change (e.g., Cognitive Dissonance, Self-Consistency, Self-Perception) are traditionally formulated as "verbal theories" or "mechanism sketches." While conceptually rich, these accounts lack the technical specifications and operational constraints required to be executed as computational systems. Consequently, they are often underdetermined, meaning multiple distinct computational implementations could theoretically satisfy the verbal description, making it difficult to rigorously test, compare, or replicate their specific predictions.

The authors address the challenge of translating these abstract verbal theories into executable, runnable actor–environment simulations that can reproduce the characteristic behavioral patterns of classic psychological experiments. A secondary problem identified is that modern Large Language Models (LLMs), when placed in these scenarios, often fail to reproduce historical results due to conflicts between modern linguistic priors (e.g., contemporary ethical norms, skepticism of authority) and the specific socio-ecological assumptions of mid-20th-century experiments.

2. Methodology: Generative Actor-Based Modelling (GABM)

The authors utilize the Concordia simulation framework to render verbal theories as executable decision logics.

• Architecture: The system uses a central Game Master (GM) that manages the world state and adjudicates actions. Generative Actors (powered by the Gemma 3 LLM) operate via predictive pattern completion. They generate a "suffix" (intended action) based on a "prefix" containing memories, observations, and a specific Decision Logic.
• Decision Logics as Algorithms: The core innovation is treating each psychological theory as a distinct algorithmic chain of reasoning steps (a decision logic) that processes the situational prefix before generating an action.
  • Festinger Actor (Cognitive Dissonance): Implements a logic that detects logical inconsistencies between attitudes, beliefs, and behaviors. If a conflict is found, it generates a resolution (rationalization) to reduce psychological discomfort.
  • Aronson Actor (Self-Consistency): Similar to Festinger but acts as a "high-pass filter." It only triggers attitude change if the inconsistency threatens the actor's moral or competent self-concept. It includes a mechanism to check for recent self-affirmation (buffering the self-concept) which can prevent the conflict from triggering change.
  • Bem Actor (Self-Perception): Implements a parsimonious self-inference system. Instead of checking for internal conflict, it infers attitudes by answering three reflexive questions: "What kind of person am I?", "What kind of situation is this?", and "What would a person like me do in this situation?"
  • Minimal Actor: A baseline "naked" LLM with no specific decision logic, serving as a control.
• Experimental Environments: The models were tested in three canonical paradigms:
  1. Item Rating (Choice Difficulty): Measuring post-decision dissonance.
  2. Boring Task (Forced Compliance): Measuring attitude change under insufficient justification ($5 vs. $200 reward).
  3. Worm Experiment (Anticipatory Suffering): A diagnostic test to distinguish internalist (dissonance) from behaviorist (self-perception) logic, specifically testing if attitude change occurs before the behavior is performed.
• Stabilisation Process: A critical methodological component involves iterative manual stabilisation. The authors found that raw LLMs often failed to replicate results due to modern priors (e.g., actors finding tasks "meditative" rather than boring, or questioning the ethics of the experiment). They had to manually engineer semantic and normative scaffolds (e.g., hardcoding trust in the experimenter, defining "boring" explicitly, clarifying the nature of "desirability") to align the actor's internal state with the historical experimental context.

3. Key Contributions

1. Formalisation of Verbal Theories: The paper successfully translates three competing verbal theories of attitude change into distinct, executable algorithmic decision logics within a shared simulation environment.
2. The "Stabilisation" Methodology: The authors argue that the process of manually refining the actor–environment coupling to achieve stable reproduction is not merely "debugging" but a core methodological step. This process surfaces hidden operational and socio-ecological dependencies (e.g., specific cultural assumptions about authority and labor) that were implicit in the original theories but undocumented.
3. Diagnostic Differentiation: The framework successfully distinguishes between theories that produce indistinguishable behavioral predictions in standard tests. For example, the Worm Experiment successfully differentiated the Bem actor (who did not change attitude before acting) from the Festinger/Aronson actors (who did).
4. Self-Affirmation Mechanism: The implementation successfully replicated the specific prediction of Aronson's theory: that self-affirmation buffers against attitude change only for the Self-Consistency actor, leaving the other actors unaffected.

4. Results

• Reproduction of Canonical Effects: All three theory-driven actors (Festinger, Aronson, Bem) successfully reproduced the expected behavioral shifts in the Item Rating and Boring Task experiments, whereas the Minimal actor did not.
• Differentiation:
  • Worm Experiment: The Festinger and Aronson actors exhibited "anticipatory suffering" (choosing to eat the worm to reduce expected discomfort), while the Bem actor correctly failed to show this pre-behavioral shift, adhering to self-perception logic.
  • Self-Affirmation: The manipulation selectively reduced attitude change in the Aronson actor but had no effect on the Festinger or Bem actors, validating the specific mechanistic split between the theories.
• Role of Scaffolding: The results demonstrated that without the "stabilisation" interventions (e.g., enforcing the "boring" nature of the task or the "trustworthiness" of the experimenter), the models failed to reproduce the effects, often defaulting to modern, rational, or ethical interpretations that bypassed the psychological mechanisms.

5. Significance and Implications

• Theory Clarification: The work demonstrates that psychological mechanisms are not purely intra-psychic; they are coupled actor–environment systems. The "truth" of a theory is contingent on specific environmental scaffolds that may be culturally latent.
• Beyond "Found Objects": Unlike studies that treat LLMs as "found objects" to see if they exhibit human-like cognition, this approach treats the LLM as a substrate for explicitly instantiating theoretical mechanisms. The behavioral shifts arise from the structured simulation architecture, not the latent properties of the LLM alone.
• Research Through Design: The paper advocates for a "research through design" tradition where the iterative construction of the artifact (the simulation) serves as a form of inquiry. The "stabilisation" process reveals the boundary conditions and hidden assumptions of classic theories.
• Future Directions: This Generative Actor-Based Modelling (GABM) workflow provides a practical framework for mapping the operational space of psychological mechanisms, allowing researchers to generate precise hypotheses about the environmental alignments necessary for specific behaviors to emerge.

In summary, the paper argues that to truly understand and test classic psychological theories in the age of AI, one must move beyond verbal descriptions to executable systems, acknowledging that the "stabilisation" of these systems is a necessary process for uncovering the hidden contextual commitments required for psychological phenomena to occur.