A recent position paper—*Position: Trustworthy AI Suffers from Invariance Conflicts and Causality is The Solution* (arXiv:2605.02640)—argues that the core goals of trustworthy AI are in fundamental structural conflict, and that these conflicts are not accidental engineering gaps but stem from mutually incompatible invariance requirements.
The Core Argument
The four pillars of trustworthy AI pull against each other:
- Fairness requires the model to be invariant to protected attributes (race, gender, age)—the output distribution must not change regardless of the value of attribute A.
- Accuracy requires sensitivity to all predictive signals—including those correlated with A.
- Privacy requires invariance to the addition or removal of individual data points.
- Interpretability requires predictably responsive behavior to semantically meaningful changes.
- Robustness demands stability under distribution shift, while training-time accuracy rewards overfitting to shortcuts specific to the training distribution.
- Causal graphs can be wrong.
- Concept superposition in foundation models makes fine-grained interventions difficult.
- Some fairness conflicts are rooted in deep value disagreements that no technical method can resolve—a causal pathway may exist, yet relying on it in decisions may still be judged unfair.
- Title: Position: Trustworthy AI Suffers from Invariance Conflicts and Causality is The Solution
- Authors: Ruta Binkyte*, Ivaxi Sheth*, Zhijing Jin, Mohammad Havaei, Bernhard Schölkopf, Mario Fritz
- Affiliations: CISPA Helmholtz Center for Information Security; Max Planck Institute for Intelligent Systems; ETH Zurich; Google
- arXiv ID: 2605.02640
- Submitted: May 4, 2026
- Category: cs.AI (Position Paper)
Each goal demands invariance to a different set of variations—and those sets conflict. As the paper puts it, this is like asking an object to turn left and right at the same time. No amount of hyperparameter tuning or compute can dissolve it.
The Case for Causality
The authors' proposed escape route is causality, centered on the idea of selective invariance. Traditional statistical methods offer only all-or-nothing invariance: either block all signals correlated with A (killing accuracy) or block none (violating fairness). Causal graphs, by contrast, can distinguish normatively unacceptable pathways (e.g., direct racial discrimination) from causally legitimate ones (e.g., gender-correlated signals transmitted through valid medical indicators).
A striking supporting result: Richens and Everitt (ICLR 2024) proved that any agent robust across all environments must implicitly learn a causal world model—robustness itself entails causality. That is a mathematical necessity, not a philosophical preference.
Acknowledged Limitations
The authors are candid that causality is not a panacea:
Why It Matters
The industry's current whack-a-mole pattern—tuning fairness metrics today, adding privacy noise tomorrow, adjusting robustness training the day after—risks each intervention degrading another goal. The Google Gemini historical-accuracy incident (2024), the disproportionate harm of privacy noise to minority groups (Pujol et al., 2020), and accuracy collapse from over-optimized robustness are described as faces of the same coin.
The paper's recommendations: redefine trustworthy AI as multi-objective optimization rather than a pile of competing constraints; build benchmarks that jointly evaluate multiple trustworthiness goals; integrate causal knowledge across the foundation model lifecycle; and acknowledge that some trade-offs have normative, not engineering, roots.