Making an AI transparent doesn't automatically win more data: transparency raises stated willingness to share only for users who already trust AI, while immediate sharing decisions remain high and unchanged across human, white-box, and black-box scenarios.

Main Finding

Transparency in AI (white-box explanations) does not by itself increase actual data-sharing behavior. Its positive effect on stated willingness to share appears only for users who already trust AI. Immediate sharing decisions are largely driven by intuitive (System 1) processes, producing a privacy paradox: high actual sharing across conditions despite lower deliberative willingness to share.

Key Points

• Experimental design: pre-registered online experiment (N = 240) with random assignment to three processing-entity conditions — human expert, transparent white-box AI, or opaque black-box AI.
• Two outcome types:
  • Immediate/actual data-sharing (behavioral choice).
  • Willingness to share (deliberative, stated preference).
• Main empirical results:
  • No significant increase in actual sharing from AI transparency alone.
  • Transparency (white-box) increased willingness to share only among participants with higher pre-existing trust in AI.
  • High and similar actual sharing rates across all three conditions, indicating a privacy paradox and dominance of System 1 (intuitive) processing in on-the-spot sharing decisions.
• Moderation and mechanisms:
  • Trust in AI moderates the effect of transparency on deliberative willingness; transparency is effective primarily when trust exists.
  • Privacy concerns did not reliably predict immediate sharing behavior, consistent with dual-process accounts where System 1 overrides deliberative privacy evaluation.
• Practical takeaway: transparency without concurrent trust-building is unlikely to move actual data acquisition outcomes.

Data & Methods

• Sample: N = 240 participants recruited online (pre-registered study).
• Context: interaction with a fictional sleep-optimization app; participants told data would be processed by either:
  • a human expert,
  • a transparent (white-box) AI with explainable design,
  • or an opaque (black-box) AI.
• Randomization ensured balanced assignment to the three conditions.
• Measures:
  • Behavioral measure: immediate decision to share requested personal sleep-related data (yes/no).
  • Stated measure: willingness to share (deliberative response).
  • Moderators: pre-existing trust in AI and privacy concern scales.
• Analysis focused on main effects of entity/transparency on both outcomes, and interactions between transparency and trust/privacy.
• Pre-registration and randomized design strengthen internal validity; single-domain, online setting limits external generalizability.

Implications for AI Economics

• Cost–benefit of transparency investments:
  • Firms investing in explainability/transparency should not expect automatic increases in obtained data. The ROI of transparency on data acquisition depends on user trust levels.
  • For many products, direct trust-building (reputation, third-party certification, data governance commitments) may be a higher-value investment than transparency tooling alone.
• Market strategy and segmentation:
  • Platforms can segment users by trust profiles: for high-trust segments, transparency features can increase stated willingness and potentially long-term engagement; for low-trust segments, different incentives or trust-building interventions are necessary.
  • Bundling transparency with trust signals (brand reputation, guarantees, certifications) could be more effective than transparency in isolation.
• Implications for data markets and business models:
  • High baseline rates of actual sharing despite stated concerns suggest firms can acquire data even when consumers express reservations; however, reliance on intuitive acceptance raises ethical and regulatory flags.
  • If regulation mandates transparency to protect consumers, policymakers should recognize transparency alone may not change behavior and should pair mandates with enforceable trust/audit mechanisms.
• Welfare and regulatory considerations:
  • The privacy paradox implies that revealed preferences (actual behavior) may not reflect true consumer welfare. Policy should account for System 1 dynamics and consider nudges, required disclosures, or default protections rather than assuming transparency suffices.
• Measurement and evaluation practices:
  • Researchers and firms should measure both actual behavior and stated willingness; relying solely on surveys can misestimate economic responses to design or policy changes.
  • Longitudinal studies are needed: transparency effects on repeated interactions, retention, and downstream monetization may differ from one-off choices.
• Directions for future research important to AI economics:
  • Test generalizability across domains (health vs. finance vs. social apps) and higher-stakes data.
  • Investigate which trust-building mechanisms (reputation, guarantees, audits, insurance) most effectively unlock transparency’s potential.
  • Explore dynamic effects: does transparency increase long-run engagement or only affect deliberative attitudes?
  • Model equilibrium effects in data markets when some firms invest in trust versus transparency, and implications for competition and welfare.

Limitations to keep in mind: single fictional sleep-app context, online convenience sample, and short-term interaction—so effects may differ in higher-stakes, repeated-use, or offline settings.