Governing artificial intelligence

The group uses learning systems in 26 approved ways and in no others. Each use is entered in a register before it runs, sorted into one of four classes, approved by the Ethics & Technology Council for a stated scope and a period of not more than twelve months, and withdrawn when it drifts outside either. The fourth class, autonomous action, is prohibited.

The register comes first

No learning or generative system is used for any group purpose until it is entered in the artificial intelligence register and approved. The entry names the class, the purpose, the accountable executive, the system owner, the datasets used for training, tuning and prompting with their origin, the population the system acts on, the inferences it is designed to draw and the inferences it is forbidden to draw.

The entry also states how those prohibitions will be tested and the tier the system carries under the Model Risk Management policy. A learning system is a model as well as a system, so it takes an owner, a tier and a validation cycle in addition to everything asked of it here. Nothing in the artificial intelligence rules relaxes a model requirement.

The register carried 26 approved uses at 30 June 2026: eleven of Class 1, nine of Class 2, six of Class 3 and none of Class 4. Six proposals were declined during the twelve months to that date. A use that is not re-approved within twelve months lapses and the system is switched off by the platform team without further notice.

Four classes of use

Every use is placed in a class according to what the system is permitted to do with its output, not according to how the system is built. Class 1 covers drafting, summarising, translating, searching and code assistance, where a named person reads, edits and issues everything produced. Class 2 covers pattern detection, anomaly flagging, screening prioritisation and research signal generation.

A Class 2 output is an input to a person and is never shown to a client as a conclusion. Class 3 covers systems that put a recommendation and the evidence for it in front of a named approver, who may accept, amend or reject it and whose decision is recorded with reasons. Class 3 is permitted only where the approver has the time, the information and the authority to decide differently.

Class 1, assistive: every output is read, edited and issued by a named person. Eleven uses.
Class 2, analytical: the output is an input to human analysis and never a client-facing conclusion. Nine uses.
Class 3, decision support: a recommendation with its evidence goes to a named approver who records the decision. Six uses.
Class 4, autonomous action: prohibited without an amendment to the Charter. No uses.

The prohibition that holds the rest together

Class 4 is any system that would take an action affecting a client, an employee, a counterparty, a position or a regulatory filing without a human decision in between. It is prohibited. The Council cannot grant an exception, nor can the Executive Committee, nor can the Board without first amending the Charter itself. That is a deliberately high bar and it has not been approached.

The effect is visible in the ordinary run of the business. A signal proposes orders and a named person authorises their transmission. A model produces a valuation and the Risk & Valuation Committee adopts the figure. A screening tool surfaces a match and a compliance officer reads it. No client, in seven years of operation, has had a decision taken about them by a machine.

Public generative services are blocked at the group boundary. Approved systems run inside the group environment or inside a contracted vendor environment bound by Charter terms: no retention of group prompts or outputs, no training on group data, no onward disclosure and audit rights for the group. A vendor that will not accept those terms is not approved, whatever its system can do.

How a use is stopped

An incident is any output that breaches a prohibition, any use outside its approved scope, any client complaint about a system-assisted decision and any failure of a Class 3 approver to record a decision with reasons. Incidents are reported to the Head of Technology and Ethics within one business day. Waiting to establish the full facts is not permitted, and a first report may be incomplete.

Two incidents were recorded in the twelve months to 30 June 2026. Both were scope breaches by Class 1 tools, in which staff used an approved drafting system for a purpose the approval did not cover. Both uses were suspended until the teams concerned had been retrained and the scope restated in the user documentation.

A system found in use to produce a prohibited inference is withdrawn on the day of discovery by the system owner or by the Head of Technology and Ethics, without waiting for a committee. The withdrawal is reported to the Council and to the Board. Restoring the use requires a fresh approval, not a resumption of the old one.

What the Board is told

The Ethics & Technology Council reviews the register, the incident log, the approval rates of every Class 3 system and every request for a human decision at each quarterly sitting in Singapore. A Class 3 system whose recommendations are almost never rejected is re-examined, because an approver who always agrees is not exercising a judgement.

The Council reports to the Board once a year with the full register, the proposals it declined and its own assessment of whether the seven principles were kept. Internal Audit examines the register and a sample of approvals on a two-year cycle. The Head of Technology and Ethics presents the report, and the independent director with a data-governance background reads it before the Board does.