Central banks must secure trust as AI risks reshape finance: IMF tech chief

What it claims

IMF Chief Information Officer Shirin Hamid gave the 76th Anniversary Oration of the Central Bank of Sri Lanka in Colombo on 23 October 2026. She argued that a central bank's AI must be "explainable, governed, secure, ethical, and human-supervised." She framed the goal as trustworthy AI, not the most AI. She named three threat classes for financial institutions:

• impersonation of humans with cloned audio and video;
• autonomous AI swarms attacking at machine speed;
• indirect breaches that manipulate internal AI assistants already inside bank firewalls.

The third class is the one the existing summary omits. It is the most relevant to engineers deploying internal assistants, because the attack runs through a trusted system rather than across the perimeter.

How it works

Hamid proposed four pillars for a central bank's AI transition: dependable data, trustworthy systems, operational resilience and adaptive governance. She said AI must not be treated as an IT department project, and that human judgment is the scarcest resource. Her stated deployment rule was "augment judgment first, and grant autonomy only where it has been earned." She paired it with a named, answerable human for every AI-influenced decision.

On data, the IMF's response to commercial models returning wrong GDP figures was StatGPT. It runs queries directly against verified data instead of letting a model generate the numbers.

Limits

This is a news report of a speech, not a technical document. It gives no method for the IMF's GDP accuracy test: no model versions, prompts, countries, sample size or definition of "inaccurate." The "up to two-thirds" figure therefore has no baseline or setting. StatGPT is described in one sentence, with no architecture, data sources, access terms or evaluation. The three-second voice-cloning figure is stated without a source. The four pillars and the earned-autonomy rule are named but not turned into controls or criteria.