← CHI 2026 map

CHI '26 · Honorable mention · full-paper review · confidence medium-high

The Data-Dollars Tradeoff: Privacy Harms vs. Economic Risk in Personalized AI Adoption

Alexander Erlei , Tahir Abbas , Kilian Bizer , Ujwal Gadiraju

DOI PDF Program page Page JSON

This is a clean, well-scoped causal paper with a genuinely interesting reversal: quantified leak risk did not move adoption, but ambiguous leak ranges did. The contribution is strongest as an empirical finding about information environments and privacy behavior, not as a broad theory of all AI privacy decisions.

Axes Lens

Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.

Review Summary

This paper reads as a solid CHI empirical contribution because it asks a crisp question, uses a reasonably large preregistered experiment, and reports a result that is easy to understand yet not obvious from common intuition. The central finding is that users did not reduce adoption when leak probability was stated as a concrete 30% risk, but they did reduce adoption when the same threat was framed as an ambiguous 10–50% range. That makes the paper interesting beyond a standard privacy-preference story: it suggests that ambiguity itself can be behaviorally consequential, even when expected outcomes are held fixed. The validation is appropriately matched to the claim type: a controlled between-subjects experiment with N=610, choice behavior, willingness to pay for a privacy label, and a bargaining task. At the same time, the paper is careful about its own limits. The leak is not a literal breach but an abstraction implemented through algorithmic price-setting, so the result should be read as evidence about framed privacy threats in experimental settings rather than direct evidence about real-world data leaks. The novelty is therefore best understood as an empirical disentangling of risk versus ambiguity and of data-type sensitivity, not as a new system or artifact. Overall, the evidence supports a medium-to-strong claim that ambiguity over data leaks can drive avoidance of personalized AI, but the deployment scope remains bounded by the experimental task and the artificial leak mechanism.

What Changed

Canon before

Prior CHI work on privacy and AI adoption typically treats privacy concern or disclosure cost as the main driver; this paper instead centers ambiguity about leak probabilities and tests whether quantifiable risk behaves differently from ambiguous risk.

Departure from common sense

The paper’s core result is counterintuitive because a quantified leak risk did not change adoption, while an ambiguous leak range did. That reverses the common expectation that clearer, more concrete risk should deter users more strongly than uncertainty.

Actual novelty

The paper’s novelty is the empirical disentangling of information environment and threatened data type in a controlled experiment, with expected monetary outcomes held fixed. It claims to be the first study to separate risk from ambiguity while also comparing sensitive demographic data against anonymized preference data.

Evidence

The evidence supports a controlled behavioral study with a 2×3 between-subjects design, N=610, measuring AI personalization choice, willingness to pay for a privacy label, and bargaining behavior. The main finding is that ambiguity, not quantified risk, suppresses adoption. The paper also explicitly acknowledges that its leak mechanism is an abstraction rather than a literal real-world breach.

Limits

Method limits

The study is experimentally strong for causal inference within its framing, but it remains a framed online experiment with an artificial leak mechanism and incentivized tasks rather than a field deployment or real breach context.

Deployment limits

Results may not transfer directly to real-world AI personalization settings where leak consequences, trust, stakes, and institutional context differ from the experimental seller-and-label setup.

Boundary conditions

Findings are bounded by the specific manipulation of risk versus ambiguity, the product-basket personalization task, and the operationalization of leaks through algorithmic price-setting rather than actual disclosure to third parties.

Manual check: none

Position in field

This paper positions itself as a causal behavioral contribution to privacy and AI adoption research, shifting attention from privacy preference alone toward ambiguity in information environments as a driver of avoidance.

Abstract

Privacy concerns significantly impact AI adoption, yet little is known about how information environments shape user responses to data leak threats. We conducted a 2 x 3 between-subjects experiment (N=610) examining how risk versus ambiguity about privacy leaks affects the adoption of AI personalization. Participants chose between standard and AI-personalized product baskets, with personalization requiring data sharing that could leak to pricing algorithms. Under risk (30% leak probability), we found no difference in AI adoption between privacy-threatening and neutral conditions (ca. 50% adoption). Under ambiguity (10-50% range), privacy threats significantly reduced adoption compared to neutral conditions. This effect holds for sensitive demographic data as well as anonymized preference data. Users systematically over-bid for privacy disclosure labels, suggesting strong demand for transparency institutions. Notably, privacy leak threats did not affect subsequent bargaining behavior with algorithms. Our findings indicate that ambiguity over data leaks, rather than only privacy preferences per se, drives avoidance behavior among users towards personalized AI.