Governance under algorithmic opacity: How financial firms construct accountability and control around AI in risk disclosures

As artificial intelligence (AI) systems reshape financial decision-making, firms face increasing pressure to assure regulators and investors that they remain in control of new technologies that are inherently opaque and difficult to predict. This study examines how financial institutions construct accountability around AI through mandatory risk disclosures. Using a three-year panel of 10-K filings from 73 publicly listed S&P financial firms, we combine keyword-based extraction with supervised machine learning to classify disclosure language along a continuum of AI accountability framing. This continuum ranges from locating accountability within the firm’s own governance structures, to situating it within external institutional rules and to dispersing it through technological unpredictability and partial autonomy. Our study advances research on responsible AI governance by showing that AI accountability is not only engineered through formal controls but also constructed discursively, as firms use accountability framings to locate and shape the boundaries of control under conditions of algorithmic opacity. Complementing accounts of AI unpredictability as a sociotechnical condition, our findings show that firms pair claims of control with disclosures of “known unknowns”, presenting AI risks as governance-relevant yet not fully governable. In this manner, firms actively construct what accountability can mean when full explainability is unattainable. We also document how the prevalence of these framings shifts over time, consistent with intensifying regulatory scrutiny and the rise of generative AI.