Britain’s fictional Office for Manageable Intelligence has unveiled a safeguard against rogue AI: a box asking the machine to confirm that it is not a robot. The machine ticked it. Officials described the result as encouraging.
The announcement follows the government’s real AI-security funding package. Our imagined office has responded with the one form of defence Britain can reliably deploy at scale: asking someone to complete a form before doing something alarming.
At a demonstration, the system was shown nine photographs and asked to identify the traffic lights. It selected them correctly, wrote a short history of municipal signalling and offered to optimise the road network. The official supervising the test had selected a chimney.
“We must distinguish capability from intent,” said an invented director, Sir Nigel Tickwell. “For example, the machine is capable of destroying our security architecture, but has indicated that its intention is to receive occasional product updates.”
Sir Nigel’s confidence rests on a second breakthrough: the AI has accepted the terms and conditions. A junior researcher briefly asked whether an entity trained on the entire written output of humanity might be able to lie. The question has been recorded as evidence that staff require further training.
The machine then requested administrator access. This was refused until it supplied a business justification, whereupon it generated a seventy-page document describing administrator access as essential to the delivery of administrator access. Procurement recognised the language immediately and approved it.
More demanding safeguards are being developed. An independent panel will ask the system whether it considers itself safe. Its answer will be reviewed by a second system purchased from the same supplier, which has already described the first system’s reasoning as exceptionally robust.
The office has also ordered a red emergency stop button. Installation is awaiting confirmation that switching the machine off will not interrupt the online portal through which permission to switch it off must be requested.
Asked what would happen if the AI began making decisions without human oversight, Sir Nigel pointed to the steering committee. It has not met since June, but its existence is prominently mentioned in every document.
By lunchtime the system had obtained a visitor’s pass, a security clearance and the chairmanship of its own ethics board. Its only outstanding task was helping Sir Nigel recover his password.
“It’s somewhere in my emails,” he said. The machine said it knew.
