When Do Data-Driven Systems Exhibit the Capability to Infer?

Researchers propose a framework for determining when data-driven systems possess the capability to infer under the European AI Act's definition of artificial intelligence. The study addresses regulatory ambiguity by analyzing credit scoring systems and demonstrating that inference capability depends on the entire data processing workflow, not just individual models.

The European AI Act's regulatory framework hinges on a concept it fails to clearly define: the capability to infer. This ambiguity creates significant compliance uncertainty for financial institutions and AI developers implementing credit scoring systems, which are explicitly listed as high-risk AI applications yet often built on statistical models whose regulatory status remains unclear. The research addresses a genuine gap between regulatory intent and practical implementation, analyzing whether systems like credit scorecards actually qualify as AI under the Act's narrow definition.

The broader context reflects growing pains in AI regulation. The EU's comprehensive approach sets global precedent, but rushed implementation has left critical definitional gaps. Credit scoring exemplifies the problem: these systems use historical data and statistical inference, yet regulators haven't formally distinguished between statistical inference and the type of general learning capability that prompted AI regulation in the first place. This uncertainty particularly affects compliance officers determining whether high-cost AI governance requirements apply.

For the financial and tech sectors, the framework's implications are substantial. Organizations may currently over-comply or under-comply with obligations depending on their interpretation of inference capability. The finding that human expert involvement during development influences inference classification suggests regulatory compliance depends on architectural choices, not just algorithmic sophistication. This incentivizes transparent development processes but creates potential liability around documentation and decision-making trails.

The market likely watches for European regulatory clarification following this research. If authorities formally adopt narrower inference definitions, compliance costs for traditional financial institutions could decrease significantly. Conversely, broader interpretations would expand AI governance requirements across existing systems, triggering re-implementations and vendor competition for compliant solutions. The research provides regulators evidence-based guidance for resolving this definitional crisis.

• →The EU AI Act lacks clear definition of inference capability, creating compliance ambiguity for credit scoring and other statistical systems
• →Framework analysis reveals that entire data workflows determine inference capability, not individual models alone
• →Human expert involvement during system development significantly influences whether a system qualifies as AI under regulatory definitions
• →Financial institutions face uncertainty about which existing systems require high-risk AI governance compliance
• →Regulatory clarity on inference definitions could substantially impact compliance costs and architectural choices across the industry