Intelligence at the Edge: Design Requirements for Accountable AI in Retail and Algorithmic Trading Under the UK and EU Regimes

Abstract

Accountability arrangements for artificial intelligence assume there is time to review a decision, room to reverse it, and a single organisation carrying its consequence. Financial-market AI has none of the three. This paper states what a system operating without them must provide, as design requirements derived against the UK and EU instruments governing retail and algorithmic trading, against four literatures not previously read together, and against the author's disclosed practice. Two requirementscarry the argument. The first concerns timing. Where an action cannot be undone, the record supporting later review has to be produced and cryptographically committed before the action, because a reconstruction assembled afterwards comes from the party whose conduct is in question, after it has learned what it will be asked. The literature specifies what such a record must contain and never says when it must be written. The second concerns a position studied so far only where the analysed person cannot leave. Once a system analyses the conduct of the person using it, that person is neither deployer, operator nor user but the subject of an inference, and five arrangements follow: separate consent, access to one's own inferred record, contestability, baseline revision, and withdrawal. The same model that tells a trader they are about to trade on tilt tells the platform whenthat trader is most likely to trade, and nothing in the model or the data separates the two uses, so the obligation is enforceable only by inspecting the record. Every requirement carries its cost and the requirements it works against. No artefact is evaluated.