TL;DR

An independent review of AI in the electricity networks, led for DESNZ by clean energy AI champion Lucy Yu, has landed with 34 recommendations and a warning most coverage will skip. Recommendation 17 gives Ofgem and the Competition and Markets Authority 12 to 18 months to build capability against algorithmic collusion — not cartels run by people using AI, but independent learning systems arriving at anti-competitive behaviour with no communication and no intent. The review notes that competition law as it stands may struggle precisely because nobody meant to do it.

Why the regulators are being put on a clock

The mechanism the review works through is flexibility baselining. The Demand Flexibility Service and schemes like it pay households and firms to shift consumption, which requires estimating what they would otherwise have used. Two problems follow once optimisation becomes continuous. Baselines derived from days a device did not respond stop being representative when it is always responding. Worse, an optimiser paid on the gap between baseline and actual usage has every incentive to move the baseline rather than alter any behaviour.

The review’s blunt formulation is that markets already get gamed by humans; AI does it “quickly, continuously and at scale”, widening the distance between a loophole appearing and a regulator closing it. Its remedy would abandon historical baselines altogether.

The collusion point is separate and harder. Independent learning agents can settle into anti-competitive positions without ever coordinating, and the review cites the CMA’s own March paper on the subject alongside Australian regulatory work. Given Britain’s fragmented market structure, it wants monitoring able to spot gaming that spans local and national markets simultaneously.

Looking forward

The remainder is a decade-long restructuring plan: probabilistic, risk-based grid operation by 2035, with planning following a year later, on the argument that today’s deterministic rules insure against worst cases regardless of conditions and so “hardcode higher than necessary costs”. It also asks for safety-case frameworks covering autonomous grid AI, developed with the AI Security Institute by 2030.

For UK businesses the near-term signal is narrow but real: a government review has instructed two regulators to build technical capability against a category of harm nobody has yet demonstrated in a British market. That is unusually anticipatory. It also rhymes with Moody’s warning that banks depend on AI they cannot properly test — the recurring problem is not the model, it is the absence of anyone able to check it.