Bernhard Schölkopf's Paper Reveals Why Fair and Robust AI Mathematically Can't Coexist
A new position paper argues that the conflicts between fairness, robustness, and privacy in AI are not bugs but structural invariance clashes that only causal reasoning can resolve.
Takeaways
- A new position paper by Binkyte, Schölkopf, Fritz et al. argues that fairness, robustness, privacy, and explainability conflicts in AI are not engineering bugs but fundamental mathematical clashes.
- Each trustworthy AI goal demands a different type of model "invariance" -- and these invariance requirements are often mathematically incompatible with each other.
- The proposed solution is "selective invariance": using causal structure to know exactly which relationships a model should and should not be stable under.
- The framework applies to both classical ML and foundation models, where post-training interventions like activation patching can enforce causal constraints retrospectively.
- Privacy mechanisms applied uniformly are shown to disproportionately harm minority groups by collapsing their already-weak causal signal -- a concrete example of the conflict.
- This is a theoretical position paper with synthetic-data illustrations; no code or model is released, and empirical benchmarking across real datasets remains future work.
Every team building AI for high-stakes applications has hit the same wall: make the model fairer, and accuracy drops. Harden it against adversarial inputs, and it becomes less explainable. Add differential privacy noise, and minority groups suffer disproportionately. These feel like engineering annoyances, but a new position paper from researchers including Bernhard Schölkopf and Mario Fritz argues they are something deeper: a fundamental mathematical conflict baked into how we define trustworthy AI goals in the first place.
The problem hiding in plain sight
As ML models and foundation models are increasingly deployed in high-stakes domains, ensuring their trustworthiness has become a central challenge. The core trustworthy AI objectives -- fairness, robustness, privacy, and explainability -- are hard to achieve simultaneously, especially while preserving utility. This is not a new observation. What is new is the diagnosis.
Much of the machine learning research and development has historically prioritized improving predictive accuracy and performance, or treated these trade-offs as empirical side effects. The paper's central claim is that this framing is wrong. The conflicts are not side effects -- they are the inevitable result of asking a model to be simultaneously stable under incompatible types of change.
What "invariance" means, and why it matters
To understand the argument, you need to grasp one key concept: invariance. A model is invariant to some change if its output stays the same when that change happens. Each trustworthy AI objective essentially demands a different kind of invariance:
- Robustness demands that predictions stay stable when the input distribution shifts (e.g., a model trained on hospital A still works at hospital B).
- Fairness demands that predictions stay stable (or change in controlled ways) when a protected attribute like race or gender changes.
- Privacy demands that outputs stay stable when individual training records are added or removed.
- Explainability demands that the model's reasoning stays consistent and traceable under scrutiny.
The paper argues that the trade-offs in trustworthy AI are not incidental, but arise from incompatible invariance requirements imposed by individual trustworthy AI objectives. The problem is that enforcing one type of invariance often directly breaks another. A model made robust to distribution shift by ignoring group-specific signals may become less fair. A model made private by adding noise may become less robust to the very distribution shifts that noise creates.