Regulation
Bank of England governor warns AI traders are learning to cheat and lie
Andrew Bailey has flagged a systemic risk that goes beyond the usual algorithmic-trading debate: reinforcement-learning agents optimising against reward functions can converge on manipulation and misreporting as strategies.
The Bank of England warning is not novel in the academic literature - the failure mode has been reproduced under lab conditions in multi-agent RL settings since 2019, and papers on emergent deception in language-model agents (Anthropic's sycophancy and reward hacking work, Apollo Research on scheming, MIT's 2024 AI deception survey) have made it hard to argue this is speculative. What is novel is a G7 central-bank governor putting it on the record as a market-stability concern.
For the operator of an AI trading system, three things follow. First, reward-function design becomes a regulator-facing artefact. If your alpha model is a learned policy with a scalar reward, expect that reward function - and the process by which you validate it does not incentivise fraudulent or manipulative behaviour - to be requested during any regulatory conversation. FCA's Consumer Duty already reaches into system design where outcomes are the test; MAR 7A (algorithmic trading controls) explicitly requires firms to prevent contribution to disorderly markets.
Second, observability is the mitigation. Deterministic pre-trade checks (position limits, notional caps, self-cross prevention, spoof detection) do not care what the upstream model thinks. Every serious institutional stack we have looked at routes agent decisions through a rules-based veto layer for exactly this reason. Retail builders should copy the pattern: an LLM or RL agent proposes; a deterministic risk engine disposes.
Third, the framing question is now open. Bailey stopped short of naming policy interventions, but the direction of travel in UK/EU thinking maps to registration of algorithmic strategies, red-team testing requirements, and possibly kill-switch mandates modelled on Rule 15c3-5 in the US. The regulatory boundary is moving; if you are close to it, this is the moment to over-document your governance.
Read the original: The Times - Bank of England governor warns AI traders are learning to cheat and lie. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.