X Hacking: The Threat of Misguided AutoML

Explainable AI can be manipulated to support predetermined conclusions: Just as researchers can selectively report statistical findings (p-hacking), analysts can now exploit "X-hacking" to cherry-pick machine learning models that produce desired explanations while maintaining acceptable accuracy levels. This undermines the trust that XAI methods are supposed to build.

Automated tools make manipulation easier and faster: The paper demonstrates that automated machine learning pipelines can systematically search through multiple "defensible" models to find explanations supporting pre-specified conclusions. Using Bayesian optimization, this manipulation process is approximately 3 times faster than random searching, making it dangerously accessible.

Many datasets are vulnerable to this manipulation: The vulnerability depends on how much redundancy exists among features in a dataset. When multiple features contain similar information, it becomes easier to construct different models with comparable accuracy but contradictory explanations—creating opportunities for X-hacking.

Current XAI methods lack built-in safeguards: The paper highlights that popular explainability metrics like SHAP values can be gamed without obvious detection, raising concerns about the credibility of published findings that rely on these methods. This threatens the reproducibility of AI research and real-world applications.

Detection and prevention strategies are needed: The authors call for new methods to identify and prevent X-hacking, suggesting this is an urgent ethical issue for responsible AI deployment, particularly in high-stakes domains where stakeholders rely on model explanations for decision-making.