Trustworthy AI Suffers from Invariance Conflicts — and Causality May Be the Answer
In early 2024, Google's Gemini image generator caused a public uproar: asked to generate images of "America's Founding Fathers," it produced crowds of Black, Asian, and Native American faces. This was not a technical glitch but the direct consequence of fairness interventions pushed too far—engineers desperate to make AI "fairer" without asking a more fundamental question: what signals were these interventions suppressing?
A position paper jointly authored by researchers from CISPA, the Max Planck Institute for Intelligent Systems, ETH Zurich, and Google offers an answer that is genuinely unsettling: the four goals of trustworthy AI—fairness, robustness, privacy, and explainability—are in fundamental structural conflict. These are not accidental "engineering isn't good enough yet" problems; they stem from mutually incompatible invariance requirements.
Why the Goals Conflict
To make this concrete:
- Fairness requires the model to be invariant to changes in protected attributes (race, gender, age)—the output distribution must not shift regardless of the attribute's value.
- Accuracy requires sensitivity to *all* predictive signals—including those correlated with the protected attribute.
- Privacy requires invariance to the addition or removal of individual data points.
- Explainability requires predictable responses to semantically meaningful changes.
- Robustness requires stability under distribution shift, yet training accuracy rewards overfitting to shortcuts specific to the training distribution.
- Causal graphs may be wrong.
- Concept superposition in foundation models makes fine-grained interventions difficult.
- Some fairness conflicts are rooted in deep value disagreements that technical means simply cannot resolve.
- Even a causally legitimate pathway may still be judged unfair when used in decisions—an answer engineers cannot provide for you.
- 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
- Institutions: 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
- Type: Position Paper
Each goal demands invariance to some set of changes, but those sets conflict with each other. This cannot be fixed by tuning hyperparameters or scaling compute. In the paper's framing, it amounts to asking one object to turn left and turn right at the same time.
The Proposed Way Out: Causality
The authors stake everything on one word: causality.
The core insight of the causal framework is selective invariance. Traditional statistical methods can only do all-or-nothing invariance—either block every signal correlated with the protected attribute (killing accuracy) or block none (violating fairness). But causal graphs can distinguish pathways that are "normatively unacceptable" (direct racial discrimination) from those that are "causally legitimate" (e.g., gender-related signals transmitted through valid medical indicators).
In other words, causal reasoning refuses the binary choice between fairness and accuracy. It asks instead: what do you want your model to be invariant to, and what should it remain sensitive to?
This might sound like a theorist's self-consolation, but the paper offers a convincing argument: Richens and Everitt (2024) proved a striking result—an agent that is robust across all environments must implicitly learn a causal world model. Robustness itself implies causality. This is not a philosophical preference; it is a mathematical necessity.
Honest Limitations
The authors do not pretend causality is a panacea. They candidly list the constraints:
Paradoxically, it is in admitting these limitations that the paper's value emerges. It offers not a cure-all but a pair of glasses—put them on, and you can finally see which whack-a-mole you're actually playing.
The Industry Reality
The status quo: tweak fairness metrics today, add privacy noise tomorrow, adjust robustness training the day after. Each intervention may inadvertently trigger degradation in another goal. Gemini's historical inaccuracies, privacy noise disproportionately harming minority groups, accuracy collapse from over-pursuing robustness—these are different faces of the same coin.
The authors' call to action is direct:
1. Redefine trustworthy AI as multi-objective optimization, not a pile of competing technical constraints. 2. Build benchmarks that jointly evaluate multiple trustworthiness goals. 3. Integrate causal knowledge across the entire lifecycle of foundation models. 4. Above all, acknowledge that some trade-offs have normative roots, not engineering ones.
The uncomfortable implication: much of the engineering effort invested in trustworthy AI over the past five years may have been pointed in the wrong direction. If trade-offs are structural, better optimizers, bigger models, and more data will not make them disappear. You need a different framework, not more compute.
If you start designing a medical diagnosis AI tomorrow, you'd better draw a causal graph first—not to look academic, but to avoid discovering a year later that the noise you added for privacy disproportionately harmed health predictions for one ethnic group, the fairness intervention killed predictive signals critical for rare diseases, and your robustness training made the model oversensitive to hospital equipment brands.
These are not hypotheticals. They are happening now. The only question is whether you're willing to admit: this is not a bug—this is structure.
---
Paper Information
Key concepts: Invariance Conflicts, Selective Invariance, Interventional Accuracy, Explicit/Implicit Causal Integration.
Cited evidence: The Google Gemini historical inaccuracy incident (Vincent, 2024); robustness implies causality (Richens & Everitt, ICLR 2024); disproportionate impact of privacy noise on minorities (Pujol et al., 2020).