AI would have changed the calculation on Covid lockdowns: David Richards
Just over six years ago, I wrote about the computer model that helped ministers shut Britain down. I have had a long time to think since: how one projection, run in a hurry, came to move an entire country.
This February, Britain’s statistics watchdog wrote to the Covid Inquiry to tell it off. The Inquiry’s executive summary had declared that, had lockdown come a week earlier, modelling “established” that 23,000 fewer people would have died in England’s first wave.
One word did the damage. Established. Not estimated. Not suggested. Not modelled. As if a what-if had been dug from the ground like a fossil.
The complaint went past one word. The 23,000 did not even mean quite what the summary implied: the scenario behind it moved the voluntary measures of March 16 a week earlier too.
The Inquiry’s secretary conceded the word was too definitive. The same mistake, written into the official autopsy.
Go back to March 2020. The plan was mitigation: shield the vulnerable, slow the spread, keep schools and shops open. Then came Report 9 from Imperial College London. In an unmitigated epidemic, it modelled 510,000 deaths in Great Britain.
Worse for the government, it suggested mitigation could still mean overwhelmed hospitals and roughly a quarter of a million deaths. Within days, the plan was dead and the country was indoors.
It was not the only model, but it broke the deadlock. The scandal was simpler: an academic tool became the lever that shut a nation before anyone outside the team could read a line of it.
The model was no polished public instrument. John Carmack, the programmer brought in to ready it for release, found a single file of about 15,000 lines, a decade old, with functions that looked, in his words, “machine translated from Fortran.” He also judged it held up better than its critics claimed. That is the point. The fault was never that academic code looked like academic code.
It was that a tool built for experts to interrogate became an instrument of government before the public could interrogate it at all.
Sweden did not prove Imperial wrong. It proved something more awkward: people are not billiard balls. They move, panic, cancel, adapt and protect themselves before any law commands them.
To citizens, lockdown was a single switch: on, for everyone, everywhere. As a father with children in school, I did not experience that certainty as an abstraction.
Now run 2020 again, but with the tools we have today. Faced with the same pathogen now, what would AI-assisted models actually tell a minister to do?
Not lock down the country by reflex on Monday. First, watch. Wastewater, hospital admissions, genomics, absenteeism and mobility, read together faster than 2020 managed.
But this is still modelling: data, assumptions, judgement, uncertainty. AI would not know the answer. It would only make it harder for a minister to pretend that anyone did.
Then aim at the network, not the nation: the settings doing most of the spreading. Close those first. Leave the rest running where the evidence allows. Keep schools open as long as the data allows. Price every path: lives against lost school years, risk against ruin. The minister still chooses, but with the trade-offs before them.
So would AI-assisted models have told the government to lock us down? Faced with a pathogen that spread fast and silently, and killed at scale, yes, they might. Sometimes the only way to break an epidemic is to break a contact network. But which one? For how long?
Schools, pubs, care homes, cities, or the whole country? That is where 2020 failed.
The Inquiry now says the first lockdown was right, only late. Perhaps. It also found that lockdown was made inevitable by political acts and omissions. The machines would not have spared ministers every hard choice.
They would not have turned a guess into a verdict. That was the mistake. They would have made it harder for a guess to be sold as one.
David Richards is managing partner of Yorkshire AI Labs