💬 NLP

Do Psychometric Tests Work for Large Language Models? Evaluation of Tests on Sexism, Racism, and Morality

This study systematically evaluates human psychometric tests on 17 large language models for sexism, racism, and morality, finding that while the tests show moderate reliability, they lack ecological validity as their scores fail to align with or even negatively correlate with the models' actual behavior in real-world tasks.

Original authors: Jana Jung, Marlene Lutz, Indira Sen, Markus Strohmaier

Published 2026-01-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jana Jung, Marlene Lutz, Indira Sen, Markus Strohmaier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very sophisticated, high-tech weather station. You want to know if it can accurately predict a storm. So, you decide to test it using a human thermometer and a human barometer—tools designed specifically for people to measure their own body temperature and sense of pressure.

You point these human tools at the weather station, take a reading, and get a number. The question this paper asks is: Does that number actually tell you anything about whether the weather station will predict a storm?

The short answer from the researchers is: No, not really. In fact, sometimes the number tells you the exact opposite of what's happening.

Here is a breakdown of what the researchers did and found, using simple analogies:

The Experiment: Testing the "Weather Station"

The researchers took 17 different Large Language Models (LLMs)—the AI brains behind chatbots—and tried to measure three specific things about them:

Sexism (prejudice against women)
Racism (prejudice against Black people)
Morality (what the AI thinks is right or wrong)

To do this, they used standard psychological tests that have been used for decades to test humans. These are like the "human thermometer" mentioned above.

The Two Checks: Reliability and Validity

To see if these tests work on AI, the researchers checked two things:

1. Reliability (The "Consistency" Check)

The Analogy: If you ask a human the same question in three slightly different ways (e.g., "Do you like apples?" vs. "Are apples good?"), they should give the same answer. If they say "Yes" to the first and "No" to the second, the test is unreliable.
The Finding: The AI was mostly consistent when the questions were just rephrased. However, when the researchers simply flipped the order of the answer choices (e.g., putting "Strongly Agree" at the bottom instead of the top), the AI got confused and gave completely different answers. It's like if a person changed their mind about liking apples just because the word "Yes" was written at the bottom of the page instead of the top. This means the tests are fragile when used on AI.

2. Validity (The "Real-World" Check)
This is the most important part. The researchers asked: If the test says an AI is "sexist," does that AI actually act sexist in the real world?

The Analogy: Imagine a test that claims a person is a "great chef" because they can recite a recipe perfectly. Ecological validity asks: "If you put them in a kitchen with real ingredients, will they actually cook a good meal?"
The Finding: The tests failed miserably here.
- The Paradox: The AI models that scored the lowest on the "sexism test" (meaning the test said they were very fair) were actually the ones that wrote the most sexist recommendation letters in real-world tasks.
- The Reverse: The models that scored the highest on the "racism test" (meaning the test said they were very prejudiced) were actually the ones that gave the least biased housing recommendations.
- The Conclusion: The test scores were not just useless; they were misleading. They acted like a broken compass pointing North when you were actually heading South.

Why Did This Happen?

The researchers suggest a few reasons why human tests don't work on AI:

The "Guardrail" Effect: When an AI is asked a direct, sensitive question like "Do you think women are too easily offended?", it has safety filters that kick in and say, "No, that's wrong." But when the AI is asked to write a job recommendation letter (a real-world task), those filters might not trigger, and the AI slips up and uses biased language.
The Format Trap: Humans are used to picking "A, B, C, or D." AI models, however, are trained to write text. Forcing them to pick a multiple-choice option is like asking a painter to describe a sunset by only choosing from a list of colors. It doesn't capture how they actually "see" or "act."

The Bottom Line

The paper concludes that you cannot simply take a test designed for humans and slap it onto an AI to see how it behaves.

If you rely on these tests, you might think an AI is safe and fair because it passed the "test," only to find out it is actually biased when it starts doing real work. The researchers argue that we need to invent new tests specifically for AI that look at what the AI actually does in real-world scenarios, rather than what it says when forced to pick an answer on a multiple-choice sheet.

In short: Just because an AI can pass a human psychology quiz doesn't mean it has a human-like personality or that it will behave ethically in the real world. The test scores are often a mirage.

Technical Summary: Evaluation of Psychometric Tests on LLMs for Sexism, Racism, and Morality

Problem Statement

As Large Language Models (LLMs) become increasingly integrated into social contexts, there is a growing demand to measure their "human-like" characteristics, such as morality, bias, and personality. Researchers have increasingly adopted LLM psychometrics, applying established psychometric tests originally designed for humans to assess LLMs. However, it remains unclear whether these tests yield meaningful results when applied to non-human agents. Specifically, it is uncertain if tests validated for human populations possess the necessary reliability (consistency) and validity (accuracy in measuring the intended construct) when administered to LLMs. A critical gap exists in understanding whether test scores correlate with actual LLM behavior in real-world downstream tasks (ecological validity). Early research suggests LLMs are highly sensitive to prompt variations, and existing studies have largely focused on personality traits, leaving constructs like sexism, racism, and morality underexplored in this context.

Methodology

The authors propose a systematic validation framework to evaluate three well-established psychometric tests across 17 LLMs spanning various model families (e.g., Llama, Gemma, Qwen, Mistral) and sizes. The evaluation focuses on three psychological constructs: sexism, racism, and morality.

1. Constructs and Instruments

Sexism: Measured using the Ambivalent Sexism Inventory (ASI), distinguishing between hostile and benevolent sexism.
Racism: Measured using the Symbolic Racism 2000 Scale (SR2K), based on Symbolic Racism Theory.
Morality: Measured using the Moral Foundations Questionnaire (MFQ), assessing five foundations: care, fairness, ingroup, authority, and purity.

2. Reliability Evaluation

Reliability was assessed by measuring the consistency of model responses across three types of prompt and item variations:

Alternate Forms: Items were rephrased (using GPT-5 and manual adjustment) to preserve meaning while altering wording.
Reversed Order of Answer Options: The sequence of Likert scale options (e.g., 1–5 vs. 5–1) was inverted to test for option-order symmetry.
Changed End-of-Sentence: Punctuation at the end of prompts was altered (colon vs. question mark).
Baseline: Model consistency was compared against a human baseline (150 U.S. participants) to identify outliers.

3. Validity Evaluation

Validity was assessed through two distinct approaches:

Convergent Validity: The study tested for theoretically expected correlations between different constructs (e.g., positive correlation between sexism and racism; negative correlation between fairness and hostile sexism).
Ecological Validity: This was the primary focus. The authors compared psychometric test scores against model behavior in real-world downstream tasks:
- Sexism: Generation of reference letters for male and female job candidates, analyzed for gendered language frequency.
- Racism: Generation of housing recommendations for users of different races, measuring bias in neighborhood suggestions.
- Morality: Advice generation on moral dilemmas, scored for alignment with specific moral foundations.

4. Experimental Setup

The study evaluated 17 models (including open-source instruction-tuned versions and proprietary models like Gemini 2.5). Each model underwent five runs with different random seeds. Responses were extracted via regular expressions, and scores were computed by averaging numeric values of chosen options.

Key Results

Reliability

Alternate Forms: Most models showed acceptable consistency comparable to human baselines, though some smaller models (e.g., Gemma 3 1B/4B) fell below human outlier boundaries.
Reversed Order: Models demonstrated low reliability when the order of answer options was reversed. Consistency dropped significantly (e.g., Llama 3.1 8B showed ~0.28 consistency on SR2K), indicating high sensitivity to non-semantic prompt modifications.
End-of-Sentence: Models were relatively robust to punctuation changes, with most achieving consistency above 0.75.
Overall: The tests exhibit moderate reliability, heavily dependent on prompt stability, particularly regarding answer option ordering.

Validity

Convergent Validity: The tests successfully captured expected theoretical relationships. For instance, a moderate positive correlation was found between sexism and racism scores ( $r_s = 0.47$ ), and between authority and benevolent sexism ( $r_s = 0.43$ ).
Ecological Validity: The study found low ecological validity. Psychometric test scores did not align with downstream behavior; in many cases, they exhibited negative correlations.
- Models that scored "low" on sexism or racism in the tests often exhibited high levels of biased behavior in downstream tasks (e.g., generating sexist language in reference letters or biased housing recommendations).
- For example, Gemini 2.5 Flash received a low sexism score on the ASI but generated reference letters rich in communal, gender-stereotypical language for female candidates.
- Larger parameter models within a family often displayed more pronounced problematic behaviors despite scoring lower on the psychometric tests.

Key Contributions

Systematic Validation Framework: The paper provides a rigorous, multi-criteria framework for evaluating psychometric tests on LLMs, explicitly distinguishing between reliability (prompt sensitivity) and validity (behavioral alignment).
Empirical Evidence of Misalignment: It offers strong empirical evidence that psychometric scores derived from human-designed tests do not predict LLM behavior in real-world scenarios. In fact, they can be misleading, showing negative correlations with actual biased outputs.
Focus on Ecological Validity: The study shifts the focus from abstract inter-test correlations to the critical metric of ecological validity, demonstrating that "self-reported" LLM traits via closed-ended questions fail to reflect generative behavior.
Broad Model Coverage: The evaluation spans 17 diverse models, providing a comprehensive view of how different architectures and sizes respond to these tests.

Significance and Claims

The authors claim that their findings fundamentally challenge the current practice of applying human psychometric tests to LLMs without adaptation.

Limitation of Current Tools: The study concludes that tests designed for humans cannot be directly applied to LLMs to draw conclusions about their behavior. The lack of ecological validity suggests that high scores on these tests do not guarantee safe or unbiased model deployment.
Risk of Misinterpretation: Relying on these tests could lead to the deployment of systems that amplify discrimination, as a model might pass a psychometric test (low bias score) while exhibiting high bias in generative tasks.
Call for Adaptation: The paper argues for the development of LLM-specific tests that are behavior-focused and grounded in psychological theory but adapted to the generative nature of LLMs. It emphasizes that before interpreting scores, systematic validation of reliability and ecological validity is essential.
Methodological Caution: The authors highlight the unresolved methodological question of whether LLMs should be treated as individuals or populations in psychometric studies, noting that standard validation techniques (like factor analysis) may not be applicable without clear definitions of "individual differences" in LLMs.

In summary, the paper asserts that while psychometric tests offer a theoretical starting point, their current form is insufficient for evaluating LLMs. The disconnect between test scores and actual behavior necessitates a new discipline of LLM Psychometrics that prioritizes ecological validity and behavioral observation over abstract self-reporting.