Slow AI: knowing when to trust the machine, and what Cumbria taught us about expertise

A movement for critical AI literacy, and why the deficit model of public understanding is quietly repeating itself in the age of AI

By Sam Illingworth

Most of what we are told about artificial intelligence arrives as acceleration: adopt faster, prompt better, automate more, or be left behind. Therefore, Slow AI, a social movement started as a refusal since it asks a different question. When is this tool the right one, and when should we leave it alone?

Slow AI is an emerging idea that argues AI should sometimes be developed and used more deliberately, rather than always prioritising speed, automation and instant output. The basic principle is that faster is not always better, especially where AI affects knowledge, creativity, education, public communication or decision-making.

The Slow AI movement is about critical AI literacy. It treats friction as something to protect. It looks to systems and incentives when things go wrong, and to the people caught inside them with some sympathy. And it assumes that human judgement is the thing worth defending, because judgement is what these systems imitate without ever possessing.

For science communicators this should sound familiar. We have had this argument before. We had it in Cumbria.

The Cumbrian lesson

In 1986, after the Chernobyl reactor fire, rain carried radioactive caesium onto the hills of Cumbria in the north of England. Government scientists restricted the movement and sale of upland sheep and told farmers it would last a few weeks. Their models said the caesium would lock into the soil quickly, because those models were built on the lowland clay soils where most of the science had been done.

The Cumbrian fells are acidic peat, and in peat the caesium stayed mobile, moving through the grass and into the animals. A restriction announced for weeks lasted, on some farms, for more than two decades.

Brian Wynne, a sociologist at Lancaster University, studied what happened next. The hill farmers knew their land, their soils and their sheep in fine detail, and that situated knowledge was waved away as an anecdote. When the farmers grew sceptical of official reassurance, the experts read it as ignorance to be corrected. Wynne showed it was nothing of the kind. The farmers were making a rational judgement about institutions that had been confidently wrong and were now evasive, with the Sellafield nuclear plant sitting on their doorstep. Their distrust was earned.

What slowness offers science communication

Experts were interpreting farmers as part of the deficit model, the idea that public scepticism is an empty space to be filled with facts. Science communication spent thirty years learning why it fails. Trust is relational. It depends on who is speaking, whether they listen, and whether they have been honest when they were wrong. You cannot pour facts into that gap and expect it to close.

The lesson from Cumbria was not that scientific expertise had no value, or that local knowledge was always right. It was that expertise becomes unreliable when it mistakes a model for the world itself and dismisses knowledge that does not fit its assumptions. The farmers knew things about their land that the scientific models did not capture. Their scepticism was not a failure to understand science, but a rational response to experts who had been confidently wrong.

AI presents a similar challenge. When people distrust these systems, the standard response is to declare an AI literacy gap and propose to fill it with explainers about how the technology works and better prompt guides. That treats a question about trust as a shortage of information.

Misinformation is where the stakes show. Generative systems produce fluent, confident, plausible falsehoods at a scale we have not seen before. I built a small game called Dead Reference that asks people to tell real academic citations from AI-fabricated ones, and most score at around chance. At the NeurIPS Conference in 2025, more than one hundred hallucinated citations survived peer review across 51 papers. Fabricated references now carry real journal names, believable author combinations and plausible identifiers. The surface cues we were all taught to rely on no longer tell us anything.

The work of science communication is to help people decide when a tool earns their trust and when it does not, to keep friction in the places where verification happens, and to be honest about who built these systems, who profits, and who carries the cost when they are wrong.

Science communicators are well placed for this, because we have already lived the lesson before. Cumbrian hillside showed what can happen when institutions privilege abstract models over situated knowledge and respond to justified scepticism with more explanation rather than more listening.

Slowness is the discipline that lets us check what is true before we pass it on.