The next wave of artificial intelligence may transform a very different segment of the workforce than previously thought, according to a new study.
Since the release of ChatGPT in 2022, research has suggested that knowledge workers - such as writers, analysts, and software developers - would be among occupations with significant exposure to AI automation.
However, much of the improvement driving AI capabilities in future is expected to come from reinforcement learning, the impacts of which, researchers say, may be being overlooked by policymakers.
A new 'Reinforcement Learning Feasibility Index' developed by Dr Bouke Klein Teeselink (King's College London) and Philip Moreira Tomei (AI Objectives Institute) assesses which jobs AI systems can actually learn to perform by gauging how feasible it is to build a reinforcement learning environment around each role, rather than just measuring current software capabilities.
By analysing almost 18,000 occupational tasks across the US economy, the researchers found that roles operating in unpredictable environments, such as CEOs, creatives and microbiologists, are relatively resistant to being turned into reinforcement learning environments in which AI systems can be trained to do the job.
Instead, the researchers found that monitoring and control professions were subject to much greater exposure than suggested by previous AI exposure indices. Jobs such as power plant operators, railroad conductors, and aircraft cargo handling supervisors demonstrates features that reinforcement learning can exploit, such as verifiable outcomes, discrete actions, and environments that are easily simulated digitally.
Meanwhile, the research added weight to the assumption that fully digital clerical roles, including data entry and payroll clerks, faced the most significant exposure.
Professions like stonemasonry and floor laying remain protected from software-based automation because of the physical demands of the jobs and the need to interact with the material world, although physical AI and robotics may be rapidly encroaching into these domains as well.
The researchers said their findings echoed earlier waves of automation, which also fell hardest on the middle of the labour market. Reinforcement learning exposure shows an inverted U-shape, peaking among mid-career and upper-middle-wage workers, whereas the highest and lowest earners face significantly lower exposure. The labour market may already be shifting, as the researchers note a recent relative decline in job openings for highly exposed occupations.
