Arthur Turrell on Why AI Can't Fix Broken

A statistician handed a microphone to an AI live on stage and the model called it a hair dryer. That demonstration, at the Royal Institution in London on 11 July 2026, is a cleaner summary of where AI and official statistics stand than most policy papers manage. The Royal Institution hosted Arthur Turrell, whose career runs from plasma physics through the Bank of England, the Office for National Statistics, and most recently Downing Street, for a talk titled "Statistics and Stupidly Smart AI."

AI is "extremely smart in the situations that it knows about and it finds it somewhat difficult otherwise." The microphone-a-tennis-racket moment proved the point in front of a paying audience: the model crushed familiar objects (wine glass, person, 95% confidence) and whiffed on anything outside its training distribution.

Why the Numbers Under Everything Are Wobbling

Turrell's argument starts somewhere unglamorous: funding formulas. UK council funding and school pupil funding both draw on the Index of Multiple Deprivation, an index built from hundreds of statistics spanning income, health, education, even broadband speeds. Get those inputs wrong and the money going to your area is wrong. Wrong migration numbers mean the wrong public debate. Wrong house valuations mean wrong tax bands.

His favourite illustration is a Bank of England chart of UK average hourly earnings going back nearly a thousand years, adjusted to 2024 prices. For most of that millennium, everyone was poor. Then the industrial revolution hit and the line went vertical. And there's the Black Death spike in the 1340s: when a third to half of England's population died, wages surged, because labour got scarce. Economic data can tell stories that are triumphant, macabre, or both.

The problem is that the machinery collecting today's data was built for a different economy. Turrell made two audience volunteers "economic statisticians" to demonstrate: one counted identical £10 widgets in a toy factory in seconds. The other was handed poems and asked to value them. Sophia, to her credit, tried. Poe sold "The Raven" for about £300 in today's money, but that anchors exactly one poem.

Manufacturing and agriculture together now account for roughly 10% of UK output. The rest is haircuts, holidays, legal advice, and poems, none of which arrive with a serial number. On top of that, survey response rates have collapsed. "Most people would rather watch Netflix than fill in a 30-page official statistics survey," Turrell told the audience, and honestly, fair.

There's a third gap he flagged that deserves more attention than it gets: unpaid work. Caring for an elderly relative at home is work, and Turrell has done some of it himself, but it vanishes from economic statistics until the same relative moves into a paid care home, at which point it suddenly counts. An ageing country is exporting billions of pounds of economic activity out of the official ledger simply by keeping care in the family.

Where AI is Actually Delivering

Turrell is blunt about what AI is: pattern recognition. Chatbots match patterns in text, vision models match patterns in images, forecasters match patterns in numbers. Stripped of mystique, that framing is also what makes his examples legible.

Job classification. Survey respondents write messy free-text descriptions of their jobs; statisticians used to manually map them to a fixed occupational taxonomy. Turrell fed a description of feeding zoo animals into an AI and got back "zookeeper, occupation code 6129." He says a version of this system is already being implemented at the ONS.
Housing starts from orbit. The US Census Bureau has replaced some surveying of construction firms with satellite imagery: repeatedly photographing areas with building permits and running segmentation models to count new housing starts. By Turrell's account it's faster, cheaper, and more accurate than the paper forms it replaces.
Pandemic mobility. During Covid-19, working at the ONS, Turrell's team co-opted 1,500 publicly owned CCTV feeds, sampling every ten minutes, with an algorithm blurring faces and number plates before counting people, cars, and cyclists. It gave government an almost real-time read on movement when official statistics operated on monthly or quarterly cycles.
Early regional forecasts. Regional growth figures for England take 14 to 15 months to publish. Turrell and ONS colleagues built a highly experimental model that produces an early indicator, roughly 14 months ahead of the official estimate for somewhere like the East Midlands. It is, by his own admission, not perfect.

Then there's the weird frontier stuff: AlphaFold predicting protein structures, AI predicting plasma disruptions in fusion reactors, DeepMind's GenCast forecasting typhoon paths, and the Vesuvius Challenge, where AI helped read carbonised Roman scrolls from Herculaneum for the first time in two millennia. That last one has paid out millions in prizes to hobbyists, not just credentialed researchers, which is a fun signal about who gets to do this work now.

The Demo that Cut Both Ways

The wearable-camera experiment is the most provocative idea in the talk and the least finished. The concept: a device worn on the chest photographs your day, and a small model on the device (images never leave it, per the design) classifies your activity into official time-use survey categories. The prototype ran on an old laptop in about six gigabytes of memory, an iPhone's worth. It nailed everything on stage: inflating a balloon, pouring tea, fixing a plug, even a volunteer pretending to run a chemistry experiment, which Turrell admitted surprised him.

But notice what the demo proves and what it doesn't. It worked because volunteers acted out recognisable scenes for a camera in a controlled room. Time-use measurement has a long history of people misremembering what they did two weeks ago. Whether a chest-mounted camera captures an honest picture of a day, or just the parts of a day that look like training data, is an open question. Turrell himself noted the model "hasn't got very far" beyond the working prototype, and I'd file this one under promising, unaudited.

The CCTV project has a similar asterisk. In one frame, Turrell pointed to a red circle and asked the audience: person or lamp post? If humans can't tell, he said, the AI can't either. The system was a pandemic-era success, but the ambiguity didn't disappear; it just got averaged over 1,500 cameras.

What the Talk Doesn't Settle

To his credit, Turrell never claims AI replaces official statistics. His framing is a portfolio: fast, cheap, slightly rough AI estimates alongside slow, expensive, accurate official ones. The regional forecasting work makes the timeliness-versus-accuracy trade-off explicit rather than pretending it can be abolished. That's the intellectually honest position, and it's also the one most likely to get flattened in translation, because "AI can produce statistics faster" is a much easier headline than "AI can produce early drafts of statistics that still need the slow machine to check them."

The open questions are the ones he couldn't fully resolve on stage. Privacy around wearable cameras is handled in the prototype by keeping images on-device, but the governance question of who audits those models, and what happens when a chest camera classifies someone's day in a category they dispute, is unresolved. The Fortune reporting on AI's invisible economic contributions cuts the other way too: if AI is boosting productivity that current instruments can't see, better measurement isn't optional.

And the deepest tension is the one Turrell kept returning to with the microphone. Every success story here lives inside a well-defined pattern: jobs with taxonomies, houses with permits, activities with official categories. The failure mode is anything the training data never saw. A statistics system built on models that are only as good as their patterns needs someone checking the edges, and that someone is still, for now, a human with a form and a stamp.

Fill in the survey when it lands on the doormat, Turrell urged the audience. It's the least glamorous civic act imaginable, and the entire funding apparatus for schools, buses, and hospitals sits on top of it. Whether the next decade of that data arrives by satellite, chest camera, or dog-eared paper form, the rule from his opening blindfold game holds: feed the navigator false numbers, and they walk into a chair every time.

Yuki Okonkwo is an AI & Machine Learning correspondent at Buzzrag.