Can AI make clinical trials more powerful without bias?

Imagine testing a new depression app. You need many volunteers and months of work for solid results: these are Randomized Controlled Trials (RCTs), science's gold standard, but they cost effort and money. Now comes CALM, a method using Large Language Models (LLMs) as "super-assistants" to boost precision without cheating.

How does it work? Researchers feed participant data to the AI and ask it to predict individual outcomes. These predictions become extra info, like a personal weather forecast for each patient. The trick is "heterogeneous calibration": CALM knows AI can be wrong or biased, so it automatically corrects errors and reweights predictions to make them reliable. The result? The treatment effect estimate becomes much sharper, needing fewer people or detecting smaller effects.

The team tested CALM simulating that depression app trial. The method reduced variance—data "noise"—working well even when the AI had never seen similar cases (zero-shot) or had seen only a few (few-shot). It's like having a more powerful telescope to observe causality: science moves faster, saving resources while keeping mathematical rigor intact.

The CALM (Causal Analysis leveraging Language Models) framework formalizes the use of LLMs as prognostic score generators μ(X) to boost estimator efficiency in RCTs. The theoretical core lies in decomposing the LLM's prediction error: instead of assuming E[Y|X] = μ(X), CALM models the heterogeneity of the residual bias δ(X) = E[Y|X] - μ(X) via non-parametric regression on calibration sample residuals. The final estimator ψ_CALM combines the AIPW (Augmented Inverse Probability Weighting) estimator with the correction term δ(X), guaranteeing root-n consistency even if μ(X) is inconsistent, provided δ(X) is estimable at a sufficiently fast rate. The few-shot variant builds a U-statistic-like predictor by averaging over randomly sampled demonstration sets, restoring the i.i.d. structure required for standard asymptotic inference and reducing prompt-selection variability. Simulations calibrated to a real digital depression intervention RCT (N≈500) show relative efficiency gains up to 40% over standard AIPW and robustness across prompt specifications.