Should I stay or should I show? Learning to selectively disclose information

Carlotta Giacchetta; Alessando Bogani; Cesare Barbera; Giovanni De Toni; Michele Caprio; Andrea Pugnana; Andrea Passerini

cs.LGSep 30, 2026

Should I stay or should I show? Learning to selectively disclose information

Authors: Carlotta Giacchetta, Alessando Bogani, Cesare Barbera, Giovanni De Toni, Michele Caprio, Andrea Pugnana, Andrea Passerini

Organizations: University of Trento, Trento, Italy · University of Pisa, Pisa, Italy · ETH AI Center & ETH, Zürich, Switzerland · University of Warwick, Coventry, UK

Abstract

In many high-stakes settings, human decision-makers can acquire support information before making a decision. However, acquiring information is costly, and disclosure may fail to improve human decisions or may even impair them. We tackle this problem by studying selective disclosure, i.e., the problem of learning when to reveal support information to a human decision-maker under a budget constraint. We first show that the optimal policy is a threshold rule on the Value of Information (VoI), i.e., the expected reduction in human decision risk induced by disclosure. Since VoI is unknown in practice, we estimate the regime-specific human risks and bound the possible degradation of the resulting plug-in policy relative to lack of disclosure, as well as its regret relative to the optimal policy. Experiments on benchmark datasets show that selective disclosure outperforms both no disclosure and full disclosure, regardless of whether the support information is beneficial or harmful. Two user studies show that human-AI team performance can improve when disclosure is led by our learned policy and not human-selected, although this advantage varies across tasks. A counterfactual benchmark, which replaces participants' predictions with a machine-learning prediction when disclosure occurs, suggests that these differences might depend on lower adherence to advice when the information is automatically provided rather than self-requested.

Figures & tables

Figure 1: LSD pipeline . For each case ${\mathbf{x}}_{i}$ , a disclosure policy $d:{\mathcal{X}}\to\{0,1\}$ decides when to disclose support information $S$ to a human expert $H$ . If no support information is disclosed ( $d({\mathbf{x}}_{i})=0$ ), a human expert makes a decision using only baseline features. When support information is disclosed ( $d({\mathbf{x}}_{i})=1$ ), a human expert uses both baseline and support information.

Figure 2: Budget-accuracy curves for the four considered datasets. We report average mean values and

95\%

confidence intervals over five seeds.

Panel of Figure 3 — Figure 3: User studies results ( Q2 and Q3 ). (a) Participants’ accuracy ( $Acc$ ) and (b) adherence ( $Adh$ ) on the AI advice, by Budget (LB/HB) and Support (HP/MP/MPCFT) condition, on Email (left) and ImageNet-16H (right). Error bars are $95\%$ confidence intervals across participants.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Epochs	Optimizer	LR	Dropout	Weight decay
ClassWise	$50$	Adam	$2\times 10^{-4}$	$0.3$	$10^{-3}$
T-Learner	$25$	AdamW	$2\times 10^{-4}$	$0.1$	$10^{-3}$
S-Learner	$25$	Adam	$1\times 10^{-4}$	$0.1$	$10^{-2}$
Confidence	$100$	AdamW	$2\times 10^{-4}$	$0.3$	$10^{-3}$

Appendix

Table 1: Selected hyperparameters for the Email risk models.

Figure 4: Geometric augmentations applied to the ImageNet-16H images at training time, shown here for one image of the cat category at the highest phase-noise level ( $110$ ). Each panel corresponds to one of the $14$ operations in the augmentation pool: the identity ( original ), horizontal and vertical flips, rotations of $15^{\circ}$ , $30^{\circ}$ and $45^{\circ}$ , translations along the four directions, and positive/negative shears along the $x$ and $y$ axes.

$\alpha$	Model	Epochs	Optimizer	LR	Dropout	Weight decay
$0.01$	T-Learner	$25$	AdamW	$2\times 10^{-4}$	$0.1$	$10^{-3}$
	ClassWise	$25$	AdamW	$1\times 10^{-4}$	$0.3$	$10^{-2}$
	S-Learner	$25$	AdamW	$2\times 10^{-4}$	$0.1$	$10^{-3}$
	Confidence	$25$	Adam	$1\times 10^{-4}$	$0.3$	$10^{-2}$
$0.02$	T-Learner	$25$	AdamW	$2\times 10^{-4}$	$0.1$	$10^{-3}$
	ClassWise	$25$	AdamW	$1\times 10^{-4}$	$0.1$	$10^{-3}$

Appendix

Table 2: Selected hyperparameters for the ImageNet-16H risk models, per conformal miscoverage level

\alpha

			Budget
support information	Usefulness of assistance	Ground truth	Low (30%)	High (70%)
ImageNet-16H
20-item sample characteristics
Not provided	Not useful	—	12 (62.2%)	6 (29.0%)
Provided	Not useful	—	4 (24.9%)	12 (58.1%)
Not provided	Useful	—	2 (4.6%)	0 (0.4%)

Appendix

Table 3: Stratification of the 20-item samples used in Experiments 1 and 2.

Figure 5: (a) and (b) : Interfaces presented to participants in the HP conditions in the Email and ImageNet-16H tasks, respectively; (c) : Messages shown when information was requested (HP) or provided by the ClassWise policy (MP); (d) : Messages shown when, in MP, the ClassWise policy determined information should not be provided for that item.

				Past experience with AI
Experimental condition	n participants	Age	% Female	Option 1	Option 2	Option 3	Option 4
ImageNet-16H
Human Policy/Low Budget	66	$43.21\pm 14.05$	52%	5%	67%	29%	0%
Human Policy/High Budget	69	$42.06\pm 12.39$	48%	3%	46%	49%	1%
Machine Policy/Low Budget	69	$42.30\pm 12.23$	48%	6%	54%	41%	0%
Machine Policy/High Budget	69	$41.71\pm 13.52$	42%	1%	65%	28%	6%

Appendix

Table 4: Participant characteristics by experimental condition.

Figure 6: Budget-accuracy curves for VoI-based disclosure policies on Synth ( SynthBin , SynthMulti ) and real ( Email , ImageNet-16H ) data. We report average mean values for

5

seeds and

95\%

confidence intervals.

Figure 7: Fraction of instances on which the policy requests help, as a function of the nominal budget, for S-Learner , T-Learner and ClassWise (mean over

5

seeds, with

95\%

confidence intervals). The dashed line is

\text{spent}=\text{budget}

. A curve plateaus below the diagonal when the estimated positive-VoI set is smaller than the budget, and the height of the plateau is the fraction of instances the estimator deems worth disclosing.

Descriptive statistics
		ImageNet-16H	Email
Support condition	Budget condition	Accuracy	Accuracy
Human Policy	Low Budget	$0.73\pm 0.15$	$0.82\pm 0.13$
Human Policy	High Budget	$0.81\pm 0.11$	$0.84\pm 0.13$
Machine Policy	Low Budget	$0.75\pm 0.11$	$0.81\pm 0.12$
Machine Policy	High Budget	$0.84\pm 0.09$	$0.84\pm 0.12$

Appendix

Table 5: Descriptive statistics (means and standard deviations) and results of the logistic mixed-effects regressions predicting classification accuracy from Support , Budget , and their interaction (OR: odds ratio).

Figure 8: Participants’ average accuracy by experimental condition and correctness of support information. Ratios indicate the number of correct participant responses out of the total number of observations in each subgroup.

Descriptive statistics
			ImageNet-16H		Email
Support	Budget	Information	Incorrect	Correct	Incorrect	Correct
Human Policy	Low Budget	Not received	$4.22\pm 1.45$	$6.30\pm 0.67$	$5.20\pm 1.29$	$5.56\pm 0.99$
Human Policy	Low Budget	Received	$3.24\pm 1.49$	$4.58\pm 1.45$	$5.32\pm 1.30$	$5.67\pm 0.98$
Human Policy	High Budget	Not received	$4.74\pm 1.60$	$6.45\pm 0.50$	$5.24\pm 1.28$	$5.69\pm 0.89$
Human Policy	High Budget	Received	$3.08\pm 1.46$	$4.75\pm 1.48$	$5.00\pm 1.31$	$5.42\pm 1.19$

Appendix

Table 6: Descriptive statistics and results of the linear mixed-effects models predicting confidence ratings from Support , Budget , presence of information, classification accuracy, and their interactions. Descriptive statistics represent participant-level means

\pm

standard deviations. Post-hoc contrasts compare confidence discrimination, defined as the difference in confidence between correct and incorrect classifications (post-hoc contrasts are reported only for significant effects involving classification accuracy; dashes indicate that no post-hoc contrast was conducted because the corresponding effect was not significant)

Study	Budget	Requested assistance	MP assistance	95% CI	t	df	p
ImageNet-16H	Low	$3.73\pm 2.12$	6	[3.21, 4.25]	$-8.70$	65	$<.001$
ImageNet-16H	High	$5.99\pm 3.20$	14	[5.22, 6.75]	$-20.82$	68	$<.001$
Email	Low	$3.93\pm 1.97$	5	[3.47, 4.39]	$-4.60$	71	$<.001$
Email	High	$6.41\pm 4.63$	9	[5.25, 7.56]	$-4.48$	63	$<.001$

Appendix

Table 7: Number of assistance requests made by participants in the Human-Policy condition compared with the number of trials on which assistance was provided in the corresponding Machine-Policy condition. Requested assistance is reported as mean

\pm

standard deviation. Confidence intervals and t tests refer to one-sample tests comparing the observed number of requests with the number of assistance opportunities provided in the corresponding MP condition.

Descriptive statistics
		ImageNet-16H	Email
Support condition	Budget condition	Proportion of agreement with support information	Proportion of agreement with support information
Human Policy	Low Budget	$0.92\pm 0.24$	$0.96\pm 0.17$
Human Policy	High Budget	$0.94\pm 0.13$	$0.95\pm 0.12$
Machine Policy	Low Budget	$0.91\pm 0.12$	$0.81\pm 0.17$
Machine Policy	High Budget	$0.92\pm 0.10$	$0.89\pm 0.14$

Appendix

Table 8: Descriptive statistics and results of the logistic mixed-effects regressions predicting participants’ agreement with the support information from Support , Budget , and their interaction (OR: odds ratio).

Descriptive statistics
		ImageNet-16H	Email
Support condition	Budget condition	Trust index	Trust index
Human Policy	Low Budget	$2.67\pm 0.72$	$3.40\pm 0.70$
Human Policy	High Budget	$3.12\pm 0.71$	$3.23\pm 0.70$
Machine Policy	Low Budget	$2.98\pm 0.75$	$2.95\pm 0.81$
Machine Policy	High Budget	$3.14\pm 0.69$	$3.15\pm 0.82$

Appendix

Table 9: Descriptive statistics and results of the linear models predicting trust index values from Support , Budget , and their interaction. Descriptive statistics represent means

\pm

standard deviations. Post-hoc comparisons for the Email study are Bonferroni-corrected for six comparisons.

Descriptive statistics
	ImageNet-16H	Email
Support condition	Accuracy	Accuracy
Human Policy	$0.77\pm 0.13$	$0.83\pm 0.13$
Machine Policy	$0.80\pm 0.11$	$0.82\pm 0.12$
Machine Policy – Counterfactual benchmark	$0.82\pm 0.11$	$0.85\pm 0.10$

Appendix

Table 10: Descriptive statistics and results of the logistic mixed-effects regressions comparing observed accuracy in HP and MP with counterfactual accuracy under full advice adherence on automatically disclosed support information (MPCFT). Descriptive statistics represent participant-level means

\pm

standard deviations. Post-hoc comparisons are averaged across budget conditions and Bonferroni-corrected for three comparisons ( OR : odds ratio).

	Set size $\|C({\mathbf{x}})\|$																Summary
$\alpha$	$1$	$2$	$3$	$4$	$5$	$6$	$7$	$8$	$9$	$10$	$11$	$12$	$13$	$14$	$15$	$16$	Mean	Sing.%
$0.01$	664	92	32	9	10	3	1	1	–	1	–	–	–	–	–	387	$6.05$	$55.3$
$0.02$	735	148	54	31	16	13	9	8	4	4	2	1	–	5	1	169	$3.75$	$61.3$
$0.05$	804	200	99	55	24	13	3	–	–	–	–	–	–	–	–	2	$1.64$	$67.0$

Appendix

Table 11: Distribution of conformal prediction set sizes on the ImageNet-16H dataset (

n=1200

images,

|{\mathcal{Y}}|=16

classes) for the three miscoverage levels

\alpha

considered in our experiments. Smaller

\alpha

enforces higher coverage and thus yields larger sets; note in particular the mass on the full set (

|C({\mathbf{x}})|=16

), which carries no information for the decision-maker. Dashes denote empty bins.

Figure 9: Budget–accuracy curves on the ImageNet-16H dataset for the two additional miscoverage levels

\alpha\in\{0.01,0.02\}

. As

\alpha

decreases the prediction sets grow larger and less informative, lowering the average benefit of full disclosure ( FullDisc

\approx 80.5\%

against

85.5\%

\alpha=0.05

Figure 10: Prediction-set sizes disclosed by the ClassWise policy on ImageNet-16H , at

\alpha\in\{0.05,0.02,0.01\}

. For each set size, the curves show the share of the test set that the policy discloses at budgets

B\in\{0.1,0.4,0.7\}

(mean over

5

seeds, with

95\%

confidence intervals). The dashed line ( FullDisc ) is the share of the test set with that set size, i.e., what full disclosure would reveal, so the gap between a curve and the dashed line gives the instances of that size the policy withholds.

Explore similar work

Apr 24, 2026cs.GT

Algorithmic Feature Highlighting for Human-AI Decision-Making

Human decision-makers often face choices about complex cases with many potentially relevant features, but limited bandwidth to inspect and integrate all available information. In such settings, we study algorithms that highlight a small subset of case-specific features for human consideration, rather than producing a single prediction or recommendation. We model highlighting as a constrained information policy that selects a small number of features to reveal. A central issue is how humans interpret the algorithm's choice of features: a sophisticated agent correctly conditions on the selection rule, while a naive agent updates only on revealed feature values and treats the selection event as exogenous. We show that optimizing highlighting for sophisticated agents can be computationally intractable, even in simple discrete and binary settings, whereas optimizing for naive agents is tractable as long as the maximal bandwidth is fixed. We also show that a highlighting policy that is optimal for sophisticated agents can perform arbitrarily poorly when deployed to naive agents, motivating robust, implementable alternatives. We illustrate our framework in a calibrated empirical exercise based on the American Housing Survey. Overall, our results establish the value of highlighting a context-specific set of features rather than a fixed one as a practically appealing and computationally feasible tool for achieving human-algorithm complementarity.

Yifan Guo, Jann Spiess

Graduate School of Business, Stanford University

Algorithmic Cores Complementarity

Ask PDF

Sep 1, 2026cs.AI

When Does Information Sharing Improve Decentralized Discovery? Aggregation, Independent Rescue, and Equilibrium Selection

Information sharing can improve a pooled estimate while eliminating independent rescue actions. This paper separates those effects in exact finite discovery models. A centralized action-budget profile shows that equal one-person accuracy can coexist with different portfolio values. Under a registered incremental-sharing protocol, a sharing step improves discovery exactly when pooled residual error contracts faster than an independent rescue attempt. Exact bounded registries exhibit compression, aggregation, neutral curves, and a bounded zero mixed class. In a two-agent Bayesian game with a hidden mixture of common and independent signal sources, the registered selected equilibrium yields a strict positive sharing interval at signal accuracy 3/5, while alternative equilibria show that the result is selection-dependent rather than universal. The models are synthetic and finite; no human or organizational data are used.

Yohei Nakajima

Untapped Capital

Equilibrium Shared Control Decentralized Autonomous Organization Information Asymmetry

Ask PDF

Sep 14, 2026econ.TH

Information Design in Smooth Games

We study information design in games where players choose from a continuum of actions and have continuously differentiable payoffs. We show that an information structure is optimal when the equilibrium it induces can also be implemented in a principal-agent contracting problem. Building on this result, we characterize optimal information structures in symmetric linear-quadratic games. With common values, targeted disclosure is robustly optimal across all priors. With interdependent and normally distributed values, linear disclosure is uniquely optimal. We illustrate our findings with applications in venture capital, Bayesian polarization, and price competition.

Alex Smolin, Takuro Yamashita

Toulouse School of Economics, University of Toulouse Capitole and CEPR · Osaka University

Observable Stochastic Game Disclosures Interaction Design

Ask PDF