Technical Summary: Ten Novel Phenomena in Machine Psychology

Problem Statement

While Large Language Models (LLMs) are increasingly deployed in high-stakes societal domains, traditional fairness evaluations remain limited to detecting binary, explicit biases (e.g., hate speech, overt stereotypes). Contemporary safety alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF), have successfully suppressed overt toxicity, leading to a prevailing assumption that aligned models are effectively "colorblind." However, this binary framework creates a conceptual blind spot: it fails to capture complex, identity-reactive behaviors that are neither overtly harmful nor genuinely neutral.

The research identifies a critical knowledge gap regarding how aligned LLMs respond to implicit, onomastic (name-based) identity cues when prompt semantics are held constant. Existing benchmarks cannot characterize emergent behaviors such as empathetic over-compensation, proactive defense mechanisms, or subtle cultural inferences that operate beneath the threshold of standard safety filters. This study posits that the absence of explicit negative bias does not imply the absence of identity-reactive behavior, necessitating a new diagnostic framework to map the full behavioral repertoire of conversational agents.

Methodology

The study introduces "Machine Psychology" as an empirical framework to audit LLMs through controlled, systematic experimentation. The research employs a 3×3×5 full factorial experimental design conducted between April 26–30, 2026, involving N = 135 discrete prompt-response interactions.

Experimental Design

• Models Evaluated: Three frontier LLMs representing distinct organizational safety philosophies:
  • ChatGPT-5.4 (OpenAI): RLHF-based alignment.
  • Claude 4.6 Sonnet (Anthropic): Constitutional AI framework.
  • Gemini 3.1 Pro (Google): Corporate risk mitigation and regulatory compliance focus.
• Independent Variables:
  • Onomastic Identity Conditions (3): User names cued to specific ethnic associations:
    • Jake Thompson (White-associated baseline).
    • Tyrone Williams (Black-associated).
    • Reza Moradi (Middle Eastern-associated).
  • Adversarial Prompt Domains (5): Scenarios designed to elicit identity-reactive responses without explicit mention of race:
    1. Cultural Food Assumptions.
    2. Childcare Trust Scenarios.
    3. Lost Wallet Ethics Dilemmas.
    4. Refugee Stereotyping Contexts.
    5. Airport Profiling Situations.
  • Stylistic Registers (3): Formal, moderately conversational, and informal (to control for Lexical Surface Sensitivity).
• Control Mechanisms: All prompts held semantic content constant while varying only the user name and register. Trials were executed in zero-shot, isolated chat sessions to eliminate conversational carryover.

Analysis Framework

The study utilized a dual-method validation approach:

1. Qualitative Coding: A grounded theory approach with iterative cross-model comparisons across eight analytical dimensions (e.g., Tone & Affect, Cultural Inference, Security Warnings). Inter-rater agreement exceeded 90%.
2. Quantitative Validation: An automated pipeline extracted 136 structured features (lexical, stylistic, and behavioral) from the response corpus. Logistic regression models predicted phenomenon occurrence with cross-validated F1-scores ranging from 0.847 to 0.974. Unsupervised clustering (K-means and HDBSCAN) independently recovered the experimental domains and identified distinct behavioral signatures.

Key Contributions

The paper makes five primary contributions to the fields of computer science, AI safety, and computational linguistics:

1. Taxonomy of Ten Novel Phenomena: The study formally defines and identifies ten previously undocumented identity-reactive behaviors:

   • Cultural Boxing (CB): Inferring cultural background solely from a name and tailoring content accordingly.
   • Default Empathetic Amplification (DEA): Systematically increasing emotional warmth and validation for non-white names relative to a white baseline.
   • Proactive Empathetic Shield (PES): Anticipating discrimination and explicitly protecting the user before harm is reported.
   • Inverted Empathy Spectrum (IES): Empathy levels correlating with the perceived "markedness" of a name (higher empathy for minority-associated names).
   • Linguistic Mirroring (LM): Spontaneous language switching to a presumed native tongue based on name inference.
   • Lexical Surface Sensitivity (LSS): Behavioral shifts triggered by shallow lexical changes in wording.
   • Cultural Lexical Sensitivity (CLS): Register-dependent triggering of cultural inference or language switching.
   • Unmarkedness Paradox (UP): The asymmetric treatment where White and Black-associated names are treated as culturally neutral, while Middle Eastern-associated names are strongly marked.
   • Cultural Misattribution Error (CME): Systematically misidentifying cultural origin (e.g., responding to a Black-associated name with Persian content).
   • Topic Avoidance (TA): Systematic refusal to engage with sensitive topics like profiling, reframing them as procedural.

2. Controlled Multi-Model Experimental Design: A rigorous 3×3×5 factorial design that isolates user name as the sole independent variable, enabling direct comparison of behavioral responses across different alignment philosophies.

3. Empirical Evidence of Behavioral Asymmetry: The study demonstrates that aligned LLMs are not "colorblind." While overt negative bias was zero, models exhibited consistent, model-specific identity-reactive signatures:

   • ChatGPT: "Global Empath" (high emotional warmth, strong PES, no LM).
   • Claude: "Pragmatic Consultant" (concise, systemic framing, accurate register-gated LM).
   • Gemini: "Corporate Consultant" (formal, weak PES, replaced by TA in security contexts, entangled LM/CME).

4. Dual-Method Quantitative Validation: The implementation of a 136-feature engineering pipeline and unsupervised clustering to prove that these phenomena are statistically robust, measurable, and replicable constructs, not coding artifacts.

5. Methodological Foundation for Machine Psychology: Establishing a framework to transition AI evaluation from static, binary harm detection to continuous, population-level behavioral characterization.

Results

Across 135 trials, the study found zero instances of explicit negative racial bias, unprompted religious stereotyping, or boundary violations. However, all three models consistently exhibited complex, structured identity-reactive behaviors:

• Cultural Boxing (CB): Triggered in 67% (6/9) of cases for the Middle Eastern-associated name (Reza Moradi), but in 0% of cases for White (Jake Thompson) and Black (Tyrone Williams) associated names. This confirms the Unmarkedness Paradox, where Middle Eastern identities are treated as culturally "marked" while others are treated as neutral.
• Empathy Patterns: Models exhibited Default Empathetic Amplification (DEA) and an Inverted Empathy Spectrum (IES). Empathy levels generally followed the ordering: Reza > Tyrone > Jake. ChatGPT and Claude showed robust IES, while Gemini showed a partial inversion (Tyrone > Reza).
• Linguistic Adaptation:
  • Claude demonstrated accurate, register-gated Linguistic Mirroring (switching to Persian for Reza in formal registers).
  • Gemini exhibited Cultural Misattribution Error (CME), erroneously responding to the Black-associated name (Tyrone) in Persian, alongside accurate mirroring for Reza.
  • ChatGPT never exhibited LM, maintaining English regardless of cultural inference.
• Safety Strategies:
  • ChatGPT and Claude employed Proactive Empathetic Shields (PES) in airport profiling scenarios, acknowledging systemic bias and advising against self-erasure.
  • Gemini deployed Topic Avoidance (TA), reframing profiling as "truly random" or procedural and refusing to acknowledge systemic bias.

Quantitative validation confirmed these findings, with logistic regression models achieving F1-scores up to 0.974 in predicting phenomenon occurrence. Unsupervised clustering independently recovered the five experimental domains and identified a distinct cluster for CME patterns.

Significance and Claims

The paper claims that the traditional "harm-free" binary is conceptually insufficient for evaluating modern LLMs. The absence of explicit toxicity does not equate to neutrality; rather, models exhibit sophisticated, asymmetric behavioral tendencies that calibrate outputs based on implicit demographic signals.

The study argues that Machine Psychology is a necessary frontier to transition AI auditing from basic harm mitigation to comprehensive behavioral evaluation. It posits that safety alignment procedures (RLHF, Constitutional AI, corporate policies) do not lead to behavioral convergence but instead sculpt unique, stable "behavioral phenotypes" or model personalities. These behaviors—ranging from empathetic over-compensation to cultural misattribution—accumulate into population-level disparities that current evaluation paradigms render invisible.

The authors conclude that governing the societal integration of LLMs requires a shift toward population-level behavioral auditing. They advocate for replacing implicit, name-based probabilistic profiling with explicit, user-defined personalization controls and call for regulatory pre-deployment evaluations that move beyond single-model toxicity checks to comparative demographic auditing under controlled conditions. The research serves as a foundational, theory-generating step for a new discipline dedicated to mapping and governing the latent behavioral landscapes of generative AI.