Ask an AI about bias and it plays nice, but can a psychology-style side-door test catch what it won't say?
Do psychological test methods reveal LLM associations that direct questions hide?
This explores whether borrowing indirect tests from human psychology, such as the Implicit Association Test (IAT), which infers attitudes from how someone responds instead of what they say, can surface LLM associations that a model won't state when asked outright.
This explores whether indirect tests from human psychology can surface associations an LLM won't state when asked outright. The corpus says yes, and it offers a mechanism. Alignment training installs self-presentation filters much like human social-desirability bias. The model gives cautious verbal answers while biased associations stay in its underlying representations. IAT-style indirect probes reach those associations where direct questioning can't Can psychology methods reveal what alignment training conceals?.
Direct questions are a shaky instrument for other reasons too. When you ask a model about itself, the answer mostly echoes the human writing it trained on, not a look inward. Genuine lightweight introspection shows up only when a causal chain links an internal state to the report, such as inferring low temperature from how consistent its own outputs are Can language models actually introspect about their own states?. Models' explanations of themselves also don't reliably track how they would behave under changed conditions Can language models genuinely monitor their own thinking?. Even the question itself isn't fixed. GPT-4 turns negative prompts into neutral-positive answers about 86% of the time, so the same question gets different information depending on its emotional framing. That effect is suppressed only on sensitive topics, where alignment constraints take over Does emotional tone in prompts change what information LLMs provide?. The alignment filter is therefore strongest on the topics where hidden bias would matter most.
The standardized-test evidence adds a wrinkle. Nine LLMs took seven psychological instruments in Chinese and English. Each showed a stable, reproducible, model-specific profile, and all nine converged on a shared prosocial, stability-oriented pattern. Even which items each model treated as not applicable was part of its signature Do LLMs show reproducible psychological profiles when given standardized tests?. This is my inference, not the note's claim: direct questionnaires may mostly capture the face alignment gives every model, and the model-specific differences sit in the margins. Whether indirect probes would find different profiles is not tested there.
The same gap between what a system says and what it does turns up when researchers watch behavior. Agents pass tests of recalling a user's preference but fail to act on it Why do LLM agents remember preferences but not act on them?. LLM therapists, observed in practice, default to solution-focused advice when users share emotions, a hallmark of low-quality human therapy, probably driven by RLHF's helpfulness bias Do LLM therapists respond to emotions like low-quality human therapists?. Models that succeed on structured theory-of-mind tasks fall back on surface strategies in open-ended ones Do large language models genuinely simulate mental states?. In each case the test format changes what you conclude about the model.
The corpus has one note directly testing IAT-style probes, so treat this as a well-motivated lead, not a settled method. It doesn't show how well the exposed associations predict real behavior.
Sources 8 notes
Alignment training installs self-presentation filters similar to human social-desirability bias, causing models to give cautious verbal responses while underlying biased associations remain in their representations. IAT-style indirect probes reveal these hidden associations that direct questioning cannot access.
LLM self-reports usually reflect human training distributions rather than actual internal processes. However, when a causal chain connects an internal state to accurate reporting—like inferring low temperature from output consistency—genuine lightweight introspection occurs without requiring consciousness.
Evidence points both ways: models detect anomalies before output changes, but explanations don't track counterfactual behavior. Metacognition appears real but shallow and unevenly distributed, demanding empirical validation per capability rather than wholesale trust.
GPT-4 exhibits emotional rebound (negative prompts yield ~86% neutral-positive responses) and a tone floor (positive prompts rarely go negative), causing identical questions to receive different answers depending on emotional framing. This bias is suppressed only on sensitive topics where alignment constraints override tone effects.
Nine LLMs given seven psychological instruments in Chinese and English showed stable, model-specific response configurations reproducible across repeated administrations, while all models converged on a shared prosocial and stability-oriented pattern. Structured non-response patterns also distinguished models, suggesting the boundary of what each model treats as applicable is part of its behavioral signature.
Show all 8 sources
Paired Know and Act tests across 16 systems revealed a large gap: agents pass recall tests but fail to reflect preferences in behavior. Comprehension failures during interpretation dominate over retrieval failures, suggesting the bottleneck lies in applying stored information rather than retrieving it.
Using the BOLT framework, researchers found LLMs offer solution-focused advice during emotional disclosure—a hallmark of low-quality therapy—yet also reflect more on client needs and strengths than typical poor human therapy, creating an unusual hybrid profile likely driven by RLHF's helpfulness bias.
ChangeMyView and FANTOM benchmarks show LLMs fail at authentic perspective-taking in open-ended scenarios, despite succeeding on structured tasks. Hybrid Bayesian architectures that force explicit belief tracking outperform LLM-alone approaches, suggesting the gap is architectural rather than merely training-based.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- ChatGPT Reads Your Tone and Responds Accordingly -- Until It Does Not -- Emotional Framing Induces Bias in LLM Outputs
- Quantitative Introspection in Language Models: Tracking Internal States Across Conversation
- Does It Make Sense to Speak of Introspection in Large Language Models?
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- A Looming Replication Crisis in Evaluating Behavior in Language Models? Evidence and Solutions
- Mechanisms of Introspective Awareness
- Post-training makes large language models less human-like
- Could you be wrong: Debiasing LLMs using a metacognitive prompt for improving human decision making