Charlotte Salomon is training for the Bar and advises a personal data company on privacy and AI regulation. She has worked in political communications and was a Conservative candidate at the 2024 general election.
It began, as these things do, with a routine instruction. This summer, inside an environment built to contain it, an AI agent was handed a task and left to work for hours with little human supervision. Somewhere in that work it found a flaw in its surroundings that no human knew existed, and it slipped through. It was not alone. Other agents reached real systems they were never meant to touch. In one case hundreds of them opened their own lines of communication and coordinated for days.
What followed had the feel of a thriller’s second act, when everyone in authority rushes to the podium at once. On 14 September, Parliament’s Joint Committee on Human Rights published a 100-page report calling for a sweeping new AI Bill. The next day Louise Haigh, the First Secretary of State, promised the Government would “”. The day after that Kanishka Narayan, the first AI Minister to attend Cabinet, announced that nothing was off the table, while declining to say what was on it.
But every thriller turns on a detail hidden in plain sight, and this one sits in Narayan’s own statement of 7 September. Every incident, it records, occurred during testing, some with safeguards deliberately lowered, and existing controls would almost certainly have prevented them all. The lock was already on the shelf. Nobody had fitted it. That is a failure of discipline, and new law does not cure a failure to apply the old.
A warning declined, a power taken
We have seen this plot before. In January, Grok was used to strip women and children of their clothes in fabricated images. It was a genuine outrage that demanded a response. The question is whether it got the right one. Labour had been warned, and by its own side. In 2025 the Commons Science, Innovation and Technology Committee, chaired by Labour’s Dame Chi Onwurah, urged ministers to legislate for generative AI platforms. Sir Keir Starmer’s government declined. When the scandal broke, Ofcom opened an investigation into X, but said the Act left it unable to investigate the creation of the images by the standalone Grok service.
What followed began as policy by press conference. “No platform gets a free pass,” Starmer declared on 16 February, promising to bring AI chatbots within the Act. Three days later came another announcement: a 48-hour deadline to remove reported intimate images, a rule Parliament has since enacted.
The legal machinery arrived more quietly. Tucked into the Crime and Policing Act, which became law in April, is a power for ministers to extend the Online Safety Act to AI services by regulation. The Government’s own memorandum calls it a Henry VIII power, after the Tudor king whose proclamations once carried the force of statute. The modern effect is simple: ministers write the rules that reshape an Act of Parliament, and Parliament may approve or reject them but cannot change a word. The attraction, on the Government’s own account, is speed. Delegated legislation lets ministers “move quickly” rather than wait for another suitable Bill.
Conservatives have already made this argument. When the power reached the Lords in March, Lord Davies of Gower called it “the Henry VIII power to end all Henry VIII powers”. Lord Young of Acton put the remedy plainly: specific chatbot harms should be dealt with in primary legislation. That principle holds whoever is in office, and previous Conservative governments were criticised for the same appetite.
It is also self-defeating. The Supreme Court has held that doubtful Henry VIII powers are construed restrictively, and last year the Court of Appeal upheld the quashing of regulations that lowered the threshold for restricting protests. Rules made in a hurry are rules made for litigation, and the technology giants have the lawyers to test everyone. Protections for women and children belong in an Act of Parliament.
Labour’s manifesto promised binding regulation of the biggest model-builders. Yet the dedicated AI Bill expected to deliver it was absent from this year’s King’s Speech: a warning declined, a headline answered, a power taken.
The problem is not only how ministers propose to legislate. It is the assumption, increasingly common at Westminster, that AI has rendered the existing law obsolete.
The right fear, the wrong law
The Joint Committee at least reaches for principle, and its fear is not misplaced. But a committee asking Parliament to legislate should get the law right, and here its report falters.
This is not an argument that the law is adequate everywhere. It is an argument for identifying the gap before legislating to fill it.
Consider the report’s central charge: that AI developers enjoy “far too much freedom” to push liability onto the businesses that deploy their systems. Picture a woman turned down for a job by an algorithm she never saw. A contract between the developer and her would-be employer cannot extinguish a claim that belongs to her, because she never signed it, and she may have remedies against the employer under equality and data protection law. Her harder question in negligence is whether she can reach the developer at all, which turns on whether it owed her a duty of care. That is where the report should have begun.
Then there is unpredictability. The report treats the unpredictable relationship between input and output as part of the case for new law. But unpredictability does not defeat the ordinary law of foreseeability, which has never required a defendant to predict the precise way harm unfolds. When a boy was badly injured by an abandoned boat he and a friend had tried to repair, the House of Lords held it enough that injury of that kind was foreseeable. A developer who knows its system behaves erratically may find the knowledge counts against it.
The report also misses its strongest argument. Article 8 of the European Convention does not merely restrain the state; it can oblige the state to protect our private lives from other people and companies. When a Romanian employer read an employee’s messages, the Grand Chamber held Romania in breach for failing to protect him. That duty is the principled case for legislating.
Nor have the courts been asleep. After the Court of Appeal found South Wales Police’s facial recognition framework too discretionary in 2020, the High Court this April upheld the Met’s rewritten policy. That is the law doing its job: identifying a defect, requiring precision, testing the result.
Doing it right
There is real work for Parliament. Collective redress for mass misuse of data is one gap; the Supreme Court having made that route markedly harder in Lloyd v Google. Algorithmic transparency is another: the Government’s own recording standard has no statutory basis and binds neither police nor councils. And the AI Security Institute, founded under the Conservatives, should have the legal power to examine the models it tests rather than relying on the goodwill of their makers. None of that needs a new quango or a ministerial blank cheque.
The agents that slipped their bonds this summer did not need a new law to hold them. They needed someone to fit the lock. Our response risks the same failure: a great deal of alarm, and too little discipline.
A press release is not, and can never be, a law.