Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
The author is David Sarnoff professor of Management of Technology and professor at the MIT Sloan School of Management.
Recent research has fuelled optimism about AI’s potential. In a large call centre study, access to an AI assistant significantly improved agents’ ability to resolve problems, with the biggest gains recorded among newer workers. Other research on coding tools such as GitHub Copilot found that AI helped users, especially junior developers, complete tasks faster.
At first glance, these reports seem to be telling a straightforward story of technological progress. A tool helps people do their jobs better. Customers benefit. Companies benefit. Workers, particularly those early in their careers, gain access to expertise that once took years to accumulate.
But there’s a catch.
AI systems do not generate their capabilities from nowhere. In the call centre case, for example, the model was trained on transcripts of top agents’ conversations, in effect learning to replicate how they asked questions, managed stress, and solved problems. By digitising expertise, companies can scale an individual’s skills across time and location, allowing new hires — even after the expert has left — to perform more like veterans. For organisations and customers, the gains are clear: less experienced agents resolve issues more efficiently and customer satisfaction rises.
But for the workers whose performance trained these systems, this transformation carries real risks. Their contributions raise productivity and strengthen the organisation, but once their expertise is encoded in software it no longer belongs to them, and they are often not paid more for having supplied it. In extreme cases, higher skilled workers may be replaced by lower skilled workers, supported by the very AI models those higher skilled workers helped train.
This dynamic extends beyond call centres. Across occupations, daily work now produces rich digital traces that can be used as training data for systems designed to perform the same task. Consulting firms fine-tune models on past engagements; software teams train assistants on internal code; sales organisations mine call logs. Even creative industries draw on archives of past work to guide generative tools. As work becomes increasingly digitised, its byproducts double as inputs for AI model development.
This raises a deeper question: what happens when human expertise becomes training data?
Historically, economic security has rested on the scarcity of skill. People invest heavily in education and on-the-job learning precisely to acquire capabilities that are rare and difficult to replicate. That rarity underwrites higher wages and bargaining power. Generative AI alters this logic. When systems can absorb thousands of interactions, detect patterns, and generalise at scale, the judgment of a top performer can be codified and distributed far more quickly than human transmission ever allowed. The diffusion of knowledge can equalise opportunity, but it can also erode the scarcity that once made that knowledge valuable.
What should students entering this new labour market do to maximise the value of their expertise?
First, rethink productivity. Employees have traditionally expected to be paid for working, whether writing a report or dealing with a client. Yet, in a world where records of work are themselves valuable, workers need to think more expansively about their productivity: if their work helps train a model that benefits the company, they should seek to be recognised and, where possible, paid. More generally, workers should think carefully about how much of their job process they share with their employers. If AI training is likely to weaken their position, sharing less may be sensible; if it strengthens their role or pay, sharing more may be worthwhile.
Second, rethink competition. Many knowledge workers, those who create value by applying their expertise, assume they are more unique than they truly are. There may be only a handful of people with similar skills within a company or region. But data makes the market for expertise global. Anyone who can generate similar outputs becomes someone who can generate the data necessary to build an AI model capable of doing your work. AI provides the opportunity for people to scale their skills (and potentially be compensated for it), but it also exposes people to new competition from new places.
Finally, rethink co-operation. Many knowledge workers have long seen their jobs as unique and well paid, and so have been less inclined to support unions or other forms of collective action. But focusing only on individual career gains may not produce lasting benefits. Each time one person contributes data to improve an AI model, that model becomes better — not just at doing that individual’s job, but at doing the job of everyone in that role, anywhere.
In that sense, workers can end up competing against one another by supplying data too cheaply to companies or intermediaries that recruit people to train AI to do their jobs. When this happens, individuals may unintentionally undermine the bargaining power of others in their occupation.
If the goal is to ensure that workers share in the gains from AI, co-ordination may be necessary. Without it, individuals may strengthen the very systems that weaken their collective bargaining power. Instead, co-operation can benefit workers and companies alike. If fear about career risk leads people to hold back knowledge from AI systems, productivity may suffer. Smart companies will know that finding ways to recognise workers for their talent will ensure that they continue to supply it.
What, then, is to be done? As workers, people should think about how to use AI to expand their skills: whether by building complementary capabilities or by finding ways to scale their expertise through AI systems. As citizens, they should press for policies that give workers clearer rights over the data generated by their work and compensation for it.
