Financial regulators need to build ethics into their AI systems

Processing Content

• What’s at Stake: Unchecked supervisory AI threatens market integrity, financial inclusion and public trust.
• Expert Quote: “Supervisory decisions must remain explainable and accountable,” requiring a “human in the loop” for all significant interventions. —Marlene Amstad, FINMA
• Supporting Data: Sixty-seven percent of agencies use AI; 37% report no formal governance or ethical framework.

Source: Bullets generated by AI with editorial review

More than two-thirds of the world’s financial authorities are betting on artificial intelligence to manage our economies, so how is it possible that more than half of them are doing so without ethical guidelines in place? Why is it that less than 9% of financial authorities currently see algorithmic bias as a challenge worth solving? Can a technology-intensive financial landscape truly remain inclusive and transparent if the tools used to oversee it are operating in a governance void?

From detecting complex money-laundering patterns to predicting systemic banking shocks, financial authorities — including central banks, securities commissions and market conduct regulators — are increasingly betting on AI. These institutions play a critical role in safeguarding financial stability and protecting citizens’ savings. When effective, they can promote inclusion and sustainability by monitoring gender gaps in access to finance or supervising climate-related financial risks. However, when governance lags behind technology, the consequences can include eroded trust, weakened market integrity and unintended harm.

Over the past years, we have watched financial authorities move from paper-based spreadsheets to high-frequency data lakes. To date, 67% of supervisory agencies are deploying, piloting or exploring AI for a diverse range of high-impact use cases. Examples are everywhere. The European Central Bank’s Athena platform utilizes large-scale textual analytics, while the U.K.’s Financial Conduct Authority employs agentic AI for market abuse detection. In Finland, visual language models interpret charts. AI predicts nonperforming loans in Namibia, and assesses board functioning in India. And in Egypt, sentiment analysis decodes consumer complaints.

The latest data from the State of SupTech Report 2025 sheds light on a stark disconnect between technological ambition and institutional oversight. Over half of the surveyed authorities lack clear governance structures for AI-enabled supervisory technology (suptech).

The statistics are revealing. Over a third of agencies (37%) report having no formal governance or ethical framework for AI in supervision, while only 3% have developed a dedicated internal framework specifically tailored to suptech applications. Furthermore, only 4% align explicitly with international standards like the OECD AI Principles or the EU AI Act, 6% conduct regular ethical audits, and a mere 5% publish transparency reports regarding how AI impacts their supervisory decisions.

Perhaps most striking is the limited recognition of ethical risk as a supervisory challenge. When asked about barriers to deployment, only 8.8% of authorities identify ethical concerns or unintended societal impacts as an issue. Even fewer (8.1%) cite the risk of algorithmic bias or discrimination.

These figures are alarmingly low given the risk of AI amplifying existing inequalities. Risks may be underreported precisely because governance is underdeveloped. Where there are no bias audits, risks remain invisible and easy to dismiss.

Furthermore, 18% of agencies cite “black box” concerns (the inability to explain AI-driven outputs) as a major barrier. As Marlene Amstad, chair of the Swiss Financial Market Supervisory Authority, or FINMA, emphasized during SupTech Week, “Supervisory decisions must remain explainable and accountable,” requiring a “human in the loop” for all significant interventions. Without this, supervisors may unintentionally transfer responsibility from institutions to algorithms.

Accountability begins far upstream, rooted in data foundations. Among financial authorities, 64% report fragmented or inconsistent data as a key challenge, and this weakness tends to travel downstream into AI-enabled supervisory decisions. Poor-quality, incomplete or unrepresentative data increases the risk of biased or misleading outputs, particularly in areas such as consumer protection, financial inclusion and conduct supervision. This means that ethical failures are often baked in long before a model is trained or deployed.

Strong data governance is therefore a core element of ethical infrastructure. This includes clear data ownership, documentation of data provenance, ongoing quality controls and explicit consideration of who or what may be underrepresented in supervisory datasets. As Bernard Nsengiyumva of the National Bank of Rwanda underscored, strong data governance is not optional, it is the foundation for ethical and effective AI. Without it, even the most well-intentioned AI systems can reinforce blind spots.

The need for strong foundations becomes even more urgent as financial authorities move toward agentic AI — systems that are increasingly autonomous, goal-driven and capable of acting with limited human intervention. Agentic systems promise efficiency and scale, but they also expand the risk surface. New vulnerabilities such as prompt injection or unintended task execution can undermine supervisory control if safeguards are not built in by design. This requires moving beyond basic tool usage toward algorithmic literacy, equipping supervisors to interrogate model behavior, understand limitations and intervene when outputs conflict with supervisory judgment or public interest objectives.

Several authorities are proving that embedding ethics is possible. The U.K.’s FCA established a data & AI risk hub and an ethics framework that requires every use case to undergo independent evaluations prior to deployment. This embeds underlying knowledge of the right behaviors, assumptions, limitations and risks across supervisors, and fosters an assurance mindset. Similarly, the Bank of Tanzania established a dedicated AI and data innovation hub tasked with developing explicit AI guidelines focused on transparency, fairness and accountability, ensuring that the agency’s move toward API-enabled agentic AI remains grounded in responsible and ethical standards.

To close the accountability gap, leaders must move beyond intent and prioritize concrete operational policies that include transparency about how AI is used in supervision and where human judgment remains decisive. This involves translating high-level principles such as security, accuracy, fairness and explainability into measurable code and protocols, including auditability and bias testing, while defining clear liability for when a model or autonomous agent makes an error. Additionally, authorities must undertake ethical impact assessments that examine real supervisory effects and complete training that equips supervisors not just to use AI tools, but to question them — a critical gap given that the State of SupTech Report 2025 found that only 12% of authorities currently mandate training on ethical AI principles for developers and users.

The tipping point for supervisory transformation is no longer the availability of tools, but the governance and trust behind them. We cannot afford to have over 60% of authorities racing toward an AI-driven future while the majority still lack basic accountability frameworks. Deploying these systems without scaffolding is a systemic risk that could lead to discriminatory outcomes, unaddressed market vulnerabilities and a catastrophic loss of public trust that destabilizes the entire economy. If financial authorities are to remain trusted guardians of stability, ethical governance must become core supervisory infrastructure.