When Can LLMs Replace Humans in A/B Tests? | Spotify Engineering (opens in new tab)
LLMs can approximate human outcomes in A/B tests only when strong assumptions hold; unlike randomized user experiments, those assumptions are not guaranteed by design. In the Upworthy headline dataset, raw GPT-4o-mini predictions recovered just 39% of the human treatment effect, but appropriate calibration and repeated sampling substantially improved accuracy. However, the assumptions are hardest to justify for genuinely new products or interventions—the cases where replacing human tests would be most valuable. ## Raw LLM Predictions Underestimate Treatment Effects - Researchers used the Upworthy Research Archive, containing click-through rates from thousands of headline experiments. - GPT-4o-mini predicted click-through rates for treatment and control headlines. - Treating those predictions as human outcomes recovered only 39% of the observed human treatment effect. - The error was systematic rather than random: LLM predictions attenuated effects toward zero. - This could lead companies to underestimate product improvements and make poor shipping decisions. ## Conditions for Valid LLM Surrogates Two assumptions are required: - **Surrogacy:** LLM predictions must capture everything about a treatment that affects human behavior. Once predictions and relevant pre-treatment characteristics are accounted for, treatment assignment should provide no additional information about the human outcome. - **Comparability:** The relationship between LLM predictions and human behavior—the calibration function—must remain stable between historical experiments and the new experiment. - If either condition fails, more LLM samples will not solve the problem; the procedure estimates an effect on the model rather than the effect on users. ## Calibration Methods Matter - Linear calibration with ordinary least squares failed a falsification test, producing estimates 3.8 standard errors away from the human benchmark. - Random forests and gradient-boosted trees performed better because they could model nonlinear relationships between LLM predictions and human click behavior. - Repeatedly sampling the LLM and averaging its outputs reduces temperature-driven noise, lowering bias toward zero and reducing variance. ## Limits for New Interventions - Historical data can partially test surrogacy and comparability, but neither assumption can be verified for a treatment never previously tested. - Trust declines as a new treatment becomes more different from past experiments. - The Upworthy setting is unusually favorable: binary outcomes, text-only treatments, similar headline variants, and an LLM trained on extensive textual data. - These assumptions are much less plausible for changes to layouts, recommendation algorithms, pricing, or entirely new product concepts. Human A/B tests remain essential for genuine product innovation. LLM-based testing is most defensible for incremental changes that resemble well-understood historical treatments, with human experiments used to validate the approach and calibrate its predictions.