Research map / Large language models / AI and society

Trustworthy AI Must Account for Interactions

NeurIPS 2026 · Large language models › AI and society · approach: Human-centered trust and agency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

A position paper that surveys negative interactions among five trustworthy AI aspects (fairness, privacy, robustness, explainability, uncertainty quantification) across all pairs and argues TAI research must account for these interactions holistically.

Problem

Trustworthy AI research typically improves one or two aspects in isolation, and technical fixes for one aspect often unintentionally harm others, so the goal of achieving all aspects simultaneously is undermined.

Approach

Systematically considers every pair of five TAI aspects in both directions, collating documented negative interactions (e.g., DPSGD increasing accuracy disparity, adversarial training raising membership inference risk, equalized coverage in conformal prediction inflating set sizes) and argues for holistic, interaction-aware TAI practice with practitioner guidance.

Key results

• Reviews negative interactions across all 10 ordered-pair directions of five TAI aspects, citing prior literature for each (e.g., DPSGD increases accuracy disparity; adversarial training increases membership inference success, with six defenses all increasing MIA success).
• Shows equalized coverage in conformal prediction increases average set sizes and that smaller calibration sets per group increase coverage-breach variance.
• Proposes six practitioner guidelines (prioritize aspects, define metrics, analyze interactions, quantify risk, ablate all aspects, manage compromises) and a financial-industry vignette.

Surprising findings

• All six common adversarial defence methods tested increased membership inference attack success rates relative to undefended models.
• Fairness-aware algorithms tend to memorize underrepresented groups, improving accuracy but weakening privacy.

Limitations

• Reviews are explicitly selective rather than comprehensive, covering only five aspects and limited topics.
• Focuses on negative interactions by design, so positive interactions are largely excluded.
• Does not provide an integrated solution; acknowledges achieving multi-faceted trust is not a solved problem.

Experimental setup

Builds on: differential privacy DPSGD conformal prediction LIME adversarial training

Closest papers by meaning

• U-trustworthy Models. Reliability, Competence, and Confidence in Decision-Making (AAAI 2024 · similarity 0.89)
• OpenDeception: Learning Deception and Trust in Human–AI Interaction via Multi-Agent Simulation (ICML 2026 · similarity 0.86)
• To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.85)
• More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness (ICLR 2025 · similarity 0.85)
• Position: Intent-aligned AI Systems Must Optimize for Agency Preservation (ICML 2024 · similarity 0.85)
• Bridging the Knowledge Gap: Understanding User Expectations for Trustworthy LLM Standards (AAAI 2025 · similarity 0.85)