Organizations: McCombs School of Business, The University of Texas at Austin, Austin, TX 78712 · Naveen Jindal School of Management, The University of Texas at Dallas, Richardson, TX 75080 · Warrington College of Business, University of Florida, Gainesville, FL 32611
Artificial-intelligence (AI) policies are increasingly used to make consequential decisions in medicine, finance, education, and public services, yet deploying them responsibly requires knowing not only whether a policy helps on average, but for whom. Existing tools for off-policy evaluation and safe policy learning provide population-level guarantees, but some individuals may experience adverse treatment effects. We recast targeted deployment as a multiple-testing problem: for each candidate unit, test whether the outcome under the AI policy would exceed a user-specified threshold, and control the false discovery rate (FDR) among the selected set. Our method, Conformal Policy Selection, combines weighted conformal p-values that account for the distributional shift from behavior to target policy with a two-step selection-and-pruning procedure. We prove finite-sample FDR control and establish a doubly robust asymptotic guarantee, where FDR remains controlled if either the outcome model or the propensity model is correctly specified. Synthetic experiments verify these guarantees across four misspecification regimes, and a semi-synthetic study built on real electronic-health-record covariates illustrates how the same FDR control can act as a deployment safety gate, bounding the harmful treatments that a globally trained policy would otherwise administer to a high-risk subpopulation where it misgeneralizes. The framework offers a principled method for safe AI deployment in high-stakes applications.
Figures & tables
Figure 1: FDR and power with correctly specified scores and weights on synthetic data ( n=500 , m=100 , 400 replications). The FDR remains well below the nominal level q (dashed line) across all target levels, while power increases steadily.
Figure 2: Doubly robust FDR control experiment. Each panel shows the empirical FDR across nominal levels q under different combinations of score and weight specification. Top left : both correct, FDR controlled. Top right : correct weights, misspecified (but consistently applied) scores, FDR controlled. Bottom left : wrong weights, correct scores, FDR controlled. Bottom right : both wrong, FDR not controlled. The dashed line indicates the nominal FDR level q . Results averaged over 400 replications with n=500 calibration and m=100 test observations.
Scenario
q=0.15
q=0.25
q=0.55
q=0.95
Both correct
0.005
0.013
0.048
0.149
Correct W, wrong S
0.055
0.083
0.225
0.502
Wrong W, correct S
0.005
0.013
0.048
0.150
Both wrong
0.287
0.300
0.295
0.297
Table 1: Mean empirical FDR at selected nominal levels q across the four doubly robust scenarios (WBH method, 400 replications).
Figure 3: Conformal selection as a deployment safety gate on the EHR diabetes cohort (age ≥60 , BMI ≥30 , n=5,993 , semi-synthetic causal-forest ground truth, 20 replications). Left : the empirical FDR (blue) stays below the nominal level q (dashed) across the whole range, reaching the select-all floor ( 0.50 , dotted) only as q→1 , with power shown in orange. Right : the percentage of the AI policy’s intended (harmful) treatments that are actually deployed through the gate, which is zero for all q≤0.85 even though the gate certifies 93% of the batch for deployment at q=0.85 .
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: FDR and power on the LaLonde job training dataset (614 observations, 100 random splits) with threshold c=4,000 . Empirical FDR stays at or below the nominal level q for q≤0.65 and plateaus near the baseline negative rate ( ≈0.23 ) as q approaches 1 . Power increases smoothly with q , reflecting the trade-off the user controls by choosing the FDR budget.
Figure 5: LaLonde job training dataset. Mean post-program earnings ( re78 ) of the selected versus non-selected groups as a function of the nominal FDR level q . At small q the selected group has markedly higher mean earnings, and as q→0.9 almost all units are selected and the two curves converge.
Estimator
mean τ
frac. τ>0
max τ
S-learner (random forest)
−0.269
0.0%
−0.052
T-learner (random forest)
−0.981
0.0%
+0.000
X-learner (random forest)
−0.994
0.0%
+0.000
Causal Forest (DML)
−0.570
0.0%
−0.304
Appendix
Table 2: Estimated treatment effect on three-year avoidance for four heterogeneous-effect estimators (age ≥60 , BMI ≥30 subgroup). Every estimator returns a negative effect for essentially all patients.
Propensity decile
ntreated
avoid
ncontrol
avoid
4
15
13.3%
584
72.1%
6
30
3.3%
570
68.4%
8
58
5.2%
541
65.4%
10
250
1.6%
350
57.7%
Appendix
Table 3: Three-year avoidance rate of treated vs. untreated patients within estimated-propensity deciles. Even in balanced (high-propensity) deciles the treated fare far worse, confirming the harm is not an overlap artifact.
q
FDR
Power
Frac. treated
Beneficial treated
Harmful deployed
≤0.35
0.00
0.00
0.00
0
0
0.45
0.22
0.85
0.49
76.5
24.4
0.55
0.29
1.00
0.49
91.4
30.5
≥0.65
0.29
1.00
0.49
91.4
30.5
Appendix
Table 4: Heterogeneous safety gate on real covariates, averaged over 20 repetitions. The policy treats about half of the cohort. The realized FDR stays below the nominal q .
Figure 6: Heterogeneous safety gate. Left, empirical FDR against the nominal level with the diagonal for reference. Middle, selection power. Right, the counts of beneficial treated units and harmful deployments as functions of q .
Policy learning aims to determine who should be treated based on individual characteristics. In high-stakes settings such as medicine and public policy where safety is a central concern, improving the average outcomes alone may not be sufficient: decision makers may also seek to protect individuals from harm, in line with the Hippocratic principle of ``do no harm.'' In this paper, we propose conformal policy learning (CPL), a policy learning procedure with a new distribution-free safety guarantee that controls the probability of assigning treatment to an individual who would be harmed relative to control. CPL views each treatment decision as testing a hypothesis of counterfactual harm and assigns treatment by thresholding conformal p-values. These p-values use observable proxies and selective calibration to address the challenge that the potential outcomes under comparison are never simultaneously observed. For randomized experiments, under standard exchangeability conditions, CPL provides finite-sample safety guarantee at a user-specified level, without imposing any outcome modeling assumptions. Moreover, when the outcome model is consistently estimated, CPL achieves asymptotically optimal welfare subject to the safety constraint. In observational studies, CPL with learn-then-balance weights achieves doubly robust safety guarantees. We evaluate CPL through extensive simulations and apply it to an empirical study of AI-powered interventions designed to reduce conspiracy beliefs.
Ying Jin, Naoki Egami
Department of Statistics and Data Science, University of Pennsylvania · Department of Political Science & Statistics and Data Science Center, Massachusetts Institute of Technology
In deploying artificial intelligence (AI) models, selective prediction offers the option to abstain from making a prediction when uncertain about model quality. To fulfill its promise, it is crucial to enforce strict and precise error control over cases where the model is trusted. We propose Selective Conformal Risk control with E-values (SCoRE), a new framework for deriving such decisions for any trained model and any user-defined, bounded and continuously-valued risk. SCoRE offers two types of guarantees on the risk among ``positive'' cases in which the system opts to trust the model. Built upon conformal inference and hypothesis testing ideas, SCoRE first constructs a class of (generalized) e-values, which are non-negative random variables whose product with the unknown risk has expectation no greater than one. Such a property is ensured by data exchangeability without requiring any modeling assumptions. Passing these e-values on to hypothesis testing procedures, we yield the binary trust decisions with finite-sample error control. SCoRE avoids the need of uniform concentration, and can be readily extended to settings with distribution shifts. We evaluate the proposed methods with simulations and demonstrate their efficacy through applications to error management in drug discovery, health risk prediction, and large language models.
Tian Bai, Ying Jin
Department of Statistics, Stanford University · Department of Statistics and Data Science, University of Pennsylvania
An agent must try new behaviors to explore and improve. In high-stakes environments, an agent that violates safety constraints may cause harm and must be taken offline, curtailing any future interaction. Imitating old behavior is safe, but excessive conservatism discourages exploration. How much behavior change is too much? We show how to use any safe reference policy as a probabilistic regulator for any optimized but untested policy. Conformal calibration on data from the safe policy determines how aggressively the new policy can act, while provably enforcing the user's declared risk tolerance. Unlike conservative optimization methods, we do not assume the user has identified the correct model class nor tuned any hyperparameters. Unlike previous conformal methods, our theory provides finite-sample guarantees even for non-monotonic bounded loss functions, and it introduces a new policy control setting. Our experiments on applications ranging from natural language question answering to biomolecular engineering show that safe exploration is not only possible from the first moment of deployment, but can also improve performance.
Drew Prinster, Clara Fannjiang, Ji Won Park +4
Prescient Design, Genentech, U.S.A. · Johns Hopkins University, Baltimore, U.S.A. · New York University, NYC, U.S.A. +1