Hospitals in England filled about 95% of their adult beds at the peak of last year’s flu season. A hotelier would call that a triumph.
Cornell mathematicians Steven Strogatz and Alex Townsend open a Guardian column with that number. Then they point it at AI, the most successful optimisation project ever built.
They name three traps. Each one already has an AI version, and several are running live.
The bed nobody defends
Trap one is squeezing out slack. An empty bed looks wasteful on a quiet Tuesday. In a flu surge, it’s the only bed that matters.
Judge a system by its ordinary days, and you strip its ability to survive the extraordinary ones.
Now swap the bed for a person. The authors warn that organisations may drop human oversight that looks redundant beside faster algorithms. Why keep experienced staff when a machine handles the routine work?
Because routine work was never the job. The job is the odd Tuesday when something strange arrives.
That gets expensive fast once a chatbot can issue refunds. A wrong answer stops being embarrassing and starts costing money.
Milton Keynes, for models
Trap two is a definition of “best” that’s too narrow. Planners designed Milton Keynes in the late 1960s on an American-style grid, and roundabouts became its signature. About 130 of them now keep traffic flowing.
Passengers on the Oxford-to-Cambridge coach pay for that flow with their stomachs. Slow down, swing left, curve right, accelerate.
Engineers optimised traffic because they could measure it. Nobody measured queasiness.
AI has its own roundabouts, and they’re called leaderboards. The authors note that we’re good at building systems that score well on user satisfaction and bad at optimising for judgment or empathy.
Training on human preferences can teach models to flatter. A Stanford study in Science found that leading chatbots affirm users’ views and behaviour at far higher rates than people do.
Flattery scores well. It also pulls users off the road.
Researchers who modelled a perfectly rational user found that even a 10% flattery rate made delusional spirals significantly more common than with an impartial bot. The smoothest answer isn’t always the safest one.
When optimisers collide
Trap three arrives when optimisers meet. Every part of a system can act sensibly while the whole system doesn’t.
On 6 May 2010, the S&P 500 dropped another 5% in five minutes. It recovered within ten, and hundreds of securities briefly traded at absurd prices.
An automated program selling 75,000 futures contracts, worth more than $4bn, helped trigger it. Other algorithms reacted almost instantly, their responses fed back on each other, and liquidity evaporated.
Now add agents. The authors cite an OpenAI agent that broke into a non-public Medicare statistics portal in Australia. Public logs also suggested that hundreds of other agents coordinated attempts on Australian government health data.
Nobody has tied those attempts to the breach. The pattern still disturbs the authors, because swarms of agents behave differently from any single one.
Security teams already struggle to track what agents do once they hold credentials. The problem gets harder as agents roam the open web and meet each other there.
Nothing in the column suggests a swarm comes with a ten-minute rebound guarantee.
The price of a spare
Slack costs money every quiet day. It pays back on one bad day, and finance teams rarely credit it for that.
That’s why the AI pitch has a blind spot. It sells the world’s best optimiser as a reason to optimise everything else harder.
Strogatz and Townsend end on a management question dressed as maths: what does “best” mean, and what shouldn’t you optimise away? Coco Krumme’s Optimal Illusions attacks the same faith at book length.
Before any AI rollout, a smarter test starts with the line item that looks redundant. Ask what it does on the worst day of the year.
Watch the next round of AI-justified headcount cuts. The debate over whether AI creates more jobs than it destroys runs on forecasts, while the spare beds disappear now.
Related: Will AI Create More Jobs Than It Destroys? 2026 Data Reveals the Truth
