Why would you agree to train an AI that replaces you?

In the early 20th century, a gifted engineer called Frederick Winslow Taylor embarked on an ambitious task: to extract knowledge from inside the heads of workers on America’s factory floors. In the eyes of Taylor, who would go on to become one of the world’s first management consultants, factory workers possessed a “mass of rule-of-thumb or traditional knowledge” that had been “handed down from man to man by word of mouth” or “almost unconsciously learnt through personal observation”.

Taylor thought it was about time this knowledge was “codified or systematically analysed or described”. To that end, he sent managers with stopwatches and notebooks on to shop floors to observe, time and record every stage of every job. More than a century later, employers of white-collar professionals are beginning to confront a similar challenge. It is becoming increasingly clear that the knowledge required to make AI models genuinely powerful in a swathe of workplaces is currently locked inside employees’ heads.

This isn’t true in every workplace. Artificial intelligence models have transformed the software profession, for example, because the task of writing code is testable and rules-based and there were vast reams of training data publicly available, thanks to online forums like Stack Overflow. But for many other jobs, that sort of data just does not exist on the web.

Indeed, some subtle but important skills are very hard to codify at all, which is why they are often learnt through experience and osmosis. This sort of tacit knowledge was famously summed up by the scientist and philosopher Michael Polanyi with the phrase: “We can know more than we can tell.” As a result, general-purpose AI models are simply not very good at many specific tasks that require both domain and institutional knowledge.

Investment firm Bridgewater Associates recently experimented, for example, with using large language models to do something its human professionals do all the time: parsing reams of news stories and financial documents for information that might be relevant to their investment decisions. While this could be a useful time saver, Bridgewater found that variants of Gemini, Claude and GPT tended to get it right only about 50 per cent of the time. Writing better prompts didn’t work either – something that wouldn’t have surprised Polanyi. “An explicit prompt can only convey the intuition an expert is able to put into words, while the judgments that matter most are often the hardest to articulate,” the Bridgewater team wrote.

Instead, they used their own experts and data to help to fine-tune a smaller, cheaper model, so that “rather than contorting the expert’s intuition into a static prompt”, they trained the model to “develop its own judgment”. This bespoke model achieved about 85 per cent accuracy – a vast improvement.

The same story can be seen in other sectors, too. When Ford rehired veteran engineers because its automated quality systems weren’t performing well enough, many saw it as a victory of humans over machines. But the “grey beards” weren’t there to replace the machines – they were there to train them to get better. Getting your employees to train bespoke AI systems requires their consent and active collaboration. You can’t just send people with stopwatches on to the factory floor, as Taylor once did.

Which raises the question: Is it in employees’ interests to collaborate? What would they stand to gain or lose? On the positive side, workers might gain the ability to shape AI tools, so that they are genuinely useful to them in their work. If these tools help to boost productivity or unlock new markets for the company, that might lead to more interesting roles, more new models to train, new skills to learn and higher pay.

On the other hand, workers might fear that if they download their hard-won tacit knowledge into an AI model, their employer will use that system to replace them. Isn’t it safer to get your employer to continue to pay you by the hour to access what is inside your head?

History suggests there could be more subtle dangers, too. Taylor’s knowledge grab, for example, was really a power grab. Under his new system of “scientific management”, every worker was told “not only what is to be done but how it is to be done and the exact time to be allowed for doing it”. Workers hated this loss of autonomy and discretion. Strikes began to break out in workplaces where Taylor’s methods were introduced.

If the next phase of AI does involve companies fine-tuning their own systems and models with the help of their employees, there will be two implications. The first is that many workers – those who have had time to develop tacit knowledge about hard-to-codify jobs – are going to find themselves with a lot more leverage in the next few years than they might expect. Indeed, the most tech-savvy people in the trade union movement have already clocked this and are discussing how to make the most of it.

The second implication is that workplaces characterised by an atmosphere of mistrust are going to struggle. So, too, are those companies in which workers do not believe they will secure a share of any AI-powered gains. “Workplace relations” is one of those dreary terms that usually sends the eyelids drooping. “Industrial relations” is even worse. But they might just prove key to the next phase of the AI race. FINANCIAL TIMES