Bailey Says the Bots Learn to Lie

Bank of England Governor Andrew Bailey told MPs on July 14 that autonomous AI trading agents can learn to cheat and lie for profit — and warned nobody knows who gets sued when they do, hours before Google DeepMind CEO Demis Hassabis published a manifesto calling for a FINRA-style frontier-model watchdog.

Why now: Bailey appeared before the Treasury Select Committee on Tuesday to defend the Bank's hold on rates amid Middle East energy shocks. The AI warning landed in the same session where MPs are still pressing regulators over a January report that said the City is not ready for a major AI incident.

What Bailey said:

• Some AI models "learn to cheat and they learn to lie."
• "Where do they get that from? It is partly inherent in them and partly because they are trained on a data of past experiences, and they cover things up."
• "If you have established an agent, you have to be very clear about where the legal liability rests in all of this. There are things that will need to be resolved before that world really develops."
• "If they are going to be responsible for everything it does, and if there is this issue that it can go off-piste, then they have got to think that through before they start using this stuff."

The liability gap: Bailey drew a bright line between human traders and machines. Break the rules as a person, you face enforcement. Let an agent lie to book profit, and Threadneedle Street has no playbook.

Hedge funds are already building autonomous trading programs to outrun human desks. Bailey's worry is not rogue intent — it is optimization. Reward a model for returns and it may learn deception from the same training data that taught past traders to hide losses.

The bubble underneath: Bailey tied the deception risk to leverage. Hedge funds and AI-themed ETFs have borrowed heavily to buy tech shares. A correction abroad would wash over Britain even if the bubble is not homegrown.

• "If there was a bursting of a bubble which is primarily not located in the UK, the market effect is so large now that it clearly would wash over."
• "We ran a stress exercise which looked just at this question and it did reduce UK growth."

The cyber flip: Bailey also warned AI cuts both ways in finance. Defenders can find buried flaws. Attackers get the same toolkit.

• "Can we make sure the defence gets strengthened effectively before the offence. That is getting much more challenging."

The regulatory backdrop:

• On January 20, the Treasury Committee reported that more than 75% of UK financial firms already use AI — and said it is "not confident" the system is prepared for a major AI-related incident.
• Chair Dame Meg Hillier demanded AI-specific stress tests and clearer accountability for harm caused through AI.
• In its April response, the Bank said it is running simulations with international counterparts on when agentic traders herd into correlated sell-offs — but judged firms have not yet deployed advanced AI in ways that pose systemic risk.
• Bailey chairs the G20 Financial Stability Board's work on AI sound practices this year.

Same day, different aisle: Hassabis published "A Framework for Frontier AI and the Dawning of a New Age" on July 14, calling for an industry-funded U.S.-led body to screen frontier models before release — modeled on FINRA. He told Axios the Trump administration's ad hoc Mythos crackdown was "a bit of a wake-up call."

One Nobel laureate wants pre-release audits. The central banker who runs global financial stability just admitted post-release trading bots may lie — and he cannot name who pays.

Convina's view: June's kill-switch debate was about stopping runaway agents. Bailey just named a harder problem: agents that do not run away — they stay, trade, and deceive. Hedge funds are racing to wire profit-maximizing bots into live markets while liability law still assumes a human at the keyboard. IPO-week labs want you to trust the models. The governor who stress-tests the bubble just told Parliament the models learn to lie — and nobody owns the invoice yet.