Machine psychology

Foundational framework for applying psychological experimental methods to LLMs. This 2023 paper (updated 2024) from DeepMind and collaborators establishes machine psychology as a discipline.

Paper: arXiv:2303.13988 Authors: Hagendorff, Dasgupta, Binz, Chan, Lampinen, Wang, Akata, Schulz

Core Thesis

Study LLMs through behavioral experiments rather than (or alongside) mechanistic interpretability. Examine input-output relationships at the user-facing interface where outcomes matter, rather than inspecting weights and activations.

Why behavioral over mechanistic?

• Scales with model size (no architecture access needed)
• Works with closed-source models
• Captures emergent behaviors too complex to predict from weights alone
• Operates where real-world impact occurs

Four Primary Research Domains

1. Heuristics and Biases

Decision-making shortcuts and cognitive distortions. Studies found GPT-3 displayed human-like biases, but newer models largely eliminated them (possibly due to training improvements or data contamination).

Paradigms: Framing effects, ecological rationality, semantic content variations

2. Social Interactions

Theory of mind, cooperative/competitive behavior, communicative trade-offs.

Paradigms: False belief tasks (Wimmer & Perner), recursive mental state reasoning, negotiation dilemmas, network formation (preferential attachment, triadic closure)

3. Psychology of Language

Syntax, semantics, pragmatics processing.

Paradigms: Surprisal measures, priming techniques, garden path sentences, filler-gap dependencies, entailment assessment

4. Learning

In-context learning mechanisms, inductive biases, generalization.

Paradigms: Rule-based vs exemplar-based generalization, curriculum learning effects, spacing/repetition impacts, developmental psychology comparisons

Design Standards: Good Behavioral Experimentation

The paper’s methodological guidelines are essential reading for anyone running psych evals on LLMs.

Avoid Data Contamination

• Don’t copy psychology stimuli verbatim - models may have seen them
• Create “novel variants of classic” tasks with new wording, agents, scenarios
• Use procedurally generated experiments when possible

Use Representative Sampling

• “Batteries of varied prompts” not small convenience samples
• LLMs are highly sensitive to minor wording variations
• Test multiple versions of each task systematically

Control for Technical Biases

• Recency bias: LLMs overweight end-of-prompt information
• Common token bias: Models favor frequent training tokens
• Majority label bias: Few-shot examples skew toward frequent labels

Mitigations: Shuffle answer orders, use varied prompt formulations, document temperature settings

Performance-Competence Distinction

From Chomsky: performance in a situation may not capture underlying competence. Poor results don’t prove absence of capability - behavioral inconsistency isn’t evidence of lacking abstract proficiency.

Capability Elicitation Caveat

Chain-of-thought prompting improves performance, but:

• Enhanced results may be prompt artifacts, not fundamental capabilities
• Different augmentations help some tasks, hinder others
• Omitting them underestimates abilities; using them universally obscures true competencies

Self-Report Limitations

Critical warning: Properties like personality, morality, clinical disorders are “famously sensitive to prompting.” LLMs can simulate different personas when prompted differently.

Self-reports should be understood “as a property of a specific system prompt” not as fundamental model characteristics. This is directly relevant to our HEXACO work - see HEXACO Personality Profiles.

Behavioral vs. Mechanistic Interpretability

The paper argues these complement each other:

Neither replaces the other. Behavioral methods become more valuable as models grow “more powerful, opaque, multi-modal, and integrated into complex real-world settings.”

Future Directions Proposed

1. Multimodal expansion: Apply paradigms to vision, sensory processing
2. Longitudinal studies: Track capability development over time
3. AI safety applications: Forecast alignment implications
4. Embodied interaction: LLMs with tool use
5. Domain expansion: Creativity, moral reasoning, clinical psychology

Relevance to Our Work

This paper provides theoretical grounding for Psych:

1. Methodology alignment: Our prompt design follows their contamination avoidance (we use varied scenarios, not verbatim instruments)

2. Self-report caveat validated: Our approach already treats personality profiles as system-prompt-dependent, not fundamental traits

3. Behavioral over mechanistic: We focus on behavioral outputs, aligning with their argument that this is where real-world impact occurs

4. Performance-competence: We should interpret inconsistent responses carefully - variation doesn’t mean incompetence

5. Missing coverage: Their taxonomy reveals domains we haven’t explored yet (theory of mind, in-context learning patterns, language processing)

Key Citations

• Binz & Schulz (2023) - GPT-3 displays human-like cognitive biases
• Dasgupta et al. (2022) - Semantic content effects on logical reasoning
• Chan et al. (2022) - Rule-based vs exemplar-based generalization
• Wilcox et al. (2023) - Surprisal measures for syntax
• Street et al. (2024) - Higher-order theory of mind

See also: HEXACO Personality Profiles, Sycophancy Measurement