Is GenAI Good for Educational Measurement?

Synthetic Data, Real Populations, and Fairness Risks

Jorge Sinval

National Institute of Education, Nanyang Technological University

The Promise: Efficiency & Logistics

Automated Scale Construction

  • Frameworks like AI-GENIE integrate LLM item authoring with network psychometrics (UVA, EGA, bootEGA).
  • Automates item selection, eliminates semantic redundancies, and validates scale structures in silico.
  • Efficiency Gains: Average increases of \(8.68\) to \(20.03\) in Normalized Mutual Information (NMI) across models.
  • Produces stable, non-redundant item pools (15 to 58 items) while cutting developer hours (Russell-Lasalandra et al. 2026).

Hybrid Data Augmentation

  • Mitigates “cold-start” bottlenecks in adaptive testing when real pilot data is scarce.
  • Augmenting Sparse Data: Adding synthetic LLM respondents to human data (\(N = 50\)) in a 1:1 ratio increases item difficulty recovery rank correlation (Spearman \(\rho\)) from \(0.89\) to \(0.93\) (Liu et al. 2025).
  • Pre-Screening: Acts as a “first-pass” filter to prune poorly functioning test items prior to human administration.

The Illusion of Plausibility

High Parameter Correlations vs. The “Reliability Illusion”

  • Surface-Level Success: IRT item parameters calibrated on synthetic respondents show high correlation with human parameters (GPT-3.5 achieves Pearson \(r = 0.83\) and Spearman \(\rho = 0.87\)).
  • Spurious Internal Consistency: Open-weight LLMs (Llama-3, Mistral) exhibit exceptionally high Cronbach’s \(\alpha\) (\(0.86\) to \(0.96\)) on psychological scales, exceeding human baselines.
  • Factor Collapse: CFA and Bayesian SEM reveal this reliability is an illusion masking severe dimensionality misfit.
    • Human data fits cleanly: \(\text{CFI} = 0.972\), \(\text{RMSEA} = 0.047\)
    • LLM data degrades: \(\text{RMSEA} = 0.124 \text{ to } 0.364\), \(\text{CFI}\) drops to \(0.724\) (Shen and Han 2026).
  • High reliability stems from mechanical, stereotype-driven semantic matching rather than a true continuous latent trait.

Visualizing the “Reliability Illusion”

While Cronbach’s \(\alpha\) remains artificially high due to semantic matching, the underlying structural integrity of the latent trait collapses in synthetic data.

graph TD
    classDef human fill:#dcfce7,stroke:#15803d,stroke-width:2px;
    classDef synthetic fill:#fecaca,stroke:#b91c1c,stroke-width:2px;
    classDef node fill:#fff,stroke:#333,stroke-width:1px;

    subgraph HumanData [Human Data: Clean Unidimensionality]
        H_Theta((Latent<br>Trait)):::human
        H_Theta --> H_I1[Item 1]:::node
        H_Theta --> H_I2[Item 2]:::node
        H_Theta --> H_I3[Item 3]:::node
        H_Theta --> H_I4[Item 4]:::node
    end
    style HumanData fill:#ffffff,stroke:#94a3b8,stroke-width:2px

    subgraph SyntheticData [Synthetic Data: Structural Misfit]
        S_Theta((Latent<br>Trait)):::synthetic
        S_Bias((Mechanical<br>Stereotype)):::synthetic
        
        S_Theta --> S_I1[Item 1]:::node
        S_Theta -.->|Weak| S_I2[Item 2]:::node
        S_Bias ==>|Spurious Correlation| S_I2
        S_Bias ==>|Spurious Correlation| S_I3[Item 3]:::node
        S_Theta -.->|Weak| S_I4[Item 4]:::node
    end
    style SyntheticData fill:#ffffff,stroke:#94a3b8,stroke-width:2px

Prompt Fragility & Non-Human Geometry

Lack of Invariance

  • Minor changes in prompt wording disrupt metric and scalar invariance within the exact same model.
  • Proves LLMs lack stable latent psychological representations.
  • Loss of Predictive Validity: Regressors trained on synthetic data collapse on human data (out-of-sample \(R^2\) drops from \(+0.28\) to \(-0.18\)) (Lukauskas and Šarkauskait 2026).

Statistical Baselines Overperform

  • On a 37-model psychometric benchmark, a simple non-LLM Gaussian copula baseline matches or beats frontier LLMs (Lukauskas and Šarkauskait 2026).
  • Copula Baseline: \(PSS = 0.69\)
  • Best Frontier LLM (e.g., GPT-5.4-mini): \(PSS = 0.71\)
  • LLMs fail to meaningfully outperform traditional simulation on sample-driven joint distribution metrics.

The Peril: Distributional Narrowness

When evaluated on large-scale items, individual LLMs show severely narrow ability distributions compared to human populations, compressing item-level standard deviations (\(\sigma_{\text{LLM}} / \sigma_{\text{human}} = 0.44 \text{ to } 0.53\)) (Lukauskas and Šarkauskait 2026; Liu et al. 2025).

Figure 1

Lack of Psychological Depth

Synthetic responses show restricted use of extreme response categories and high duplication of exact response vectors (Mazzullo and Bulut 2026). They capture within-model prompt stochasticity, not authentic between-person variance.

Figure 2

The Homogenization Problem

RLHF as Value Normalization

  • Reinforcement Learning from Human Feedback translates situated human preference judgments into scalar reward functions.
  • Optimizes away non-conforming responses, flattening linguistic and cultural usage to mainstream Anglophone norms.
  • Value Collapse: Model behavior defaults to White, Liberal, Female annotator baselines, erasing diverse annotator preferences (e.g., Black/Conservative ratings of emotional awareness) (Ali et al. 2026).

Defaulting to Western-Centric Norms

  • When prompted neutrally, LLMs consistently default to liberal, Western-centric worldviews.
  • Simulation Failure: Attempts to simulate cultural differences across US and China on self-construal scales failed.
  • Both simulated Chinese and American personas exhibited Western/bicultural patterns.
  • Models fail to internalize non-Western psychological traits, imposing an implicit Western-centric baseline (Li and Qi 2025).

The Mechanics of Value Collapse

Reinforcement Learning from Human Feedback (RLHF) optimizes for majority annotator preferences, systematically erasing minority psychological profiles from the model’s behavioral distribution.

flowchart LR
    classDef process fill:#f3f4f6,stroke:#004aad,stroke-width:2px;
    classDef warning fill:#fecaca,stroke:#b91c1c,stroke-width:2px,color:#000;
    classDef default fill:#fff,stroke:#333,stroke-width:1px;

    A[LLM Generates<br>Diverse Responses]:::process --> B{Human Annotator<br>Reward Modeling}:::process
    
    B -->|Majority Preference<br>Western/Liberal/Mainstream| C[High Reward Score]:::process
    B -.->|Minority Preference<br>Non-Western/Marginalized| D[Low Reward Score]:::warning
    
    C --> E[Reward Optimization<br>PPO Algorithm]:::process
    D -.->|Filtered Out & Penalized| F((Epistemic<br>Erasure)):::warning
    
    E --> G[Homogenized LLM Base]:::process
    F -.-> G

Fairness Risks & Epistemic Injustice

Stereotype Amplification

  • Counterfactual demographic swaps on synthetic personas reveal severe, asymmetric stereotype amplification.
  • Education swaps induce artificial shifts (mean \(|d| = 0.56\), max \(1.50\)).
  • Gender swaps fabricate false differences (\(\vert{}d\vert{} = 0.12\)) where real human data shows strict null effects (Lukauskas and Šarkauskait 2026).
  • Synthetic respondents fabricate statistically significant indirect mediation pathways on null human mediation paths.

Inter-LLM Monoculture

  • The “Wisdom of LLMs” (multi-model ensembling) fails to recover population diversity.
  • Off-diagonal inter-LLM agreement (\(PSS = 0.733\)) is higher than LLM-to-human agreement (\(PSS = 0.714\)).
  • Ensemble performance collapses because models share underlying alignment/pretraining monocultures.

Fabricating Causal Pathways (DAG)

When using GenAI for demographic counterfactuals, models rely on training-data stereotypes, fabricating indirect mediation effects that do not exist in human ground-truth data.

graph LR
    classDef standard fill:#f3f4f6,stroke:#333,stroke-width:2px;
    classDef danger fill:#fecaca,stroke:#b91c1c,stroke-width:3px,color:#000;

    A(Demographic Variable<br>e.g., Gender/Race):::standard
    B(Synthetic Mediator<br>e.g., Attitude/Engagement):::danger
    C(Measurement Outcome<br>e.g., Performance Score):::standard

    A -.->|True Human Null Effect| C
    A == Stereotype Amplification ==> B
    B == Spurious Correlation ==> C
    
    linkStyle 0 stroke:#004aad,stroke-width:2px,stroke-dasharray: 5 5;
    linkStyle 1 stroke:#b91c1c,stroke-width:4px;
    linkStyle 2 stroke:#b91c1c,stroke-width:4px;

Summary Recommendations

What GenAI IS Good For

  • Pilot Studies & Item Generation: Synthetic LLM respondents and in silico scale pipelines are highly effective for rapid item drafting.
  • Redundancy Filtering: Excellent for preliminary item pool pruning.
  • Direction-of-Effect Piloting: Useful for initial structure testing before human data collection.

Mandatory Safeguards & Limitations

  • Not a Drop-in Replacement: Synthetic data cannot replace human samples for item calibration, norming, reliability estimation, or subgroup fairness auditing.
  • Beware of range restriction, structural misfit, and fabricated demographic stereotypes.
  • Required Protocol: Any measurement pipeline using GenAI must implement Messick-style construct validation, counterfactual demographic auditing, and explicit human-in-the-loop review.

References

Ali, Dalia, Dora Zhao, Allison Koenecke, and Orestis Papakyriakopoulos. 2026. “Operationalizing Pluralistic Values in Large Language Model Alignment Reveals Trade-Offs in Safety, Inclusivity, and Model Behavior.” Proceedings of the AAAI Conference on Artificial Intelligence 40.
Li, Chihao, and Yue Qi. 2025. “Toward Accurate Psychological Simulations: Investigating LLMs’ Responses to Personality and Cultural Variables.” Computers in Human Behavior 170: 108687. https://doi.org/10.1016/j.chb.2025.108687.
Liu, Yunting, Shreya Bhandari, and Zachary A. Pardos. 2025. “Leveraging LLM Respondents for Item Evaluation: A Psychometric Analysis.” British Journal of Educational Technology 56 (3): 1028–52. https://doi.org/10.1111/bjet.13570.
Lukauskas, Mantas, and Viktorija Šarkauskait. 2026. “Plausible but Not Valid: A Psychometric Audit of LLMs as Synthetic Survey Respondents.” arXiv Preprint arXiv:2608.14606.
Mazzullo, Elisabetta, and Okan Bulut. 2026. “Diversity and Carelessness in LLM-Generated Psychometric Data: Implications for Data Quality and Trust.” Computers 15 (9): 562. https://doi.org/10.3390/computers15090562.
Russell-Lasalandra, Lara L., Alexander P. Christensen, and Hudson Golino. 2026. “Generative Psychometrics via AI-GENIE: Automatic Item Generation and Validation with Network-Integrated Evaluation.” Behavior Research Methods 58: 217. https://doi.org/10.3758/s13428-026-03082-1.
Shen, Qian, and Yu Han. 2026. “The Reliability Illusion in Synthetic Patients: Psychometric Misalignment of Open-Weight LLMs on PHQ-9 and GAD-7.” Proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2026), 83–95.