Can Large Language Model-Generated Responses Support Assessment Development? A Human-Calibrated Rasch Benchmark
TL;DR
Large language models had high internal consistency but used the lowest category in 0.2% of responses versus 13.7-21.5% for humans, and no persona chose it on every item, indicating responses more predictable than the Rasch model expects.
Abstract
Large language models (LLMs) are proposed as synthetic respondents for pilot testing, but their usefulness depends on whether they supply the evidence assessment development requires. We calibrated rating scale models on 14 digital-use skill items from 6,245 adults and used the human item parameters to evaluate responses generated for 1,300 demographically matched personas. LLM responses had high internal consistency ($\alpha \approx .94$) but used the lowest category in 0.2-0.3% of responses versus 13.7-21.5% for humans, and no persona chose it on every item. These gaps changed the response-category judgment on both subscales and the PC targeting judgment; on the human-calibrated scales, 12 of 14 items had infit below 0.70, indicating responses more predictable than the Rasch model expects. Preregistered changes to the prompt, category order, and persona information did not restore the human lower range. High internal consistency is insufficient evidence that LLM responses can replace human pilot data.