17.16 Economic effects and labour
The economic literature on AI's impact on labour is large and contested.
Daron Acemoglu, MIT economist and 2024 Nobel laureate, has been the most prominent academic sceptic. His 2024 paper "The Simple Macroeconomics of AI" estimates that AI will increase US GDP by 0.93% over a decade, a meaningful but modest figure compared with the trillion-dollar capex now flowing into the sector. Acemoglu's broader argument, developed in Power and Progress (with Simon Johnson, 2023), is that the distribution of technology gains depends on institutional choices, and that prior waves of automation (from the early industrial revolution through the IT revolution) have produced either widespread prosperity or widening inequality depending on those choices.
Erik Brynjolfsson at Stanford has produced more optimistic estimates. The 2023 NBER paper with Li and Raymond (mentioned in Section 17.13) showed substantial productivity gains from generative AI in contact-centre work. Brynjolfsson's "productivity J-curve" framing argues that AI productivity gains will lag adoption by years as organisations restructure around the new technology.
Dario Amodei's October 2024 essay "Machines of Loving Grace" argued for a strongly transformative timeline in which AI compresses 100 years of biomedical progress into 5–10 years. Sam Altman's 2025 essay "The Intelligence Age" took a similar position. These optimistic positions are not universally held; the 2024–2025 wave of "AI bubble" warnings from Goldman Sachs, MIT economist David Autor and others reflect genuine analytic disagreement.
Empirical productivity studies. The 2023 Microsoft and BCG studies on Copilot showed productivity gains in software engineering (55% in the Peng study), customer service (Brynjolfsson, 14%), consulting (BCG/Harvard, 25–40% on tasks within the AI's frontier), and writing (Noy and Zhang, 40% completion-time reduction in business writing). The pattern is consistent: meaningful per-task productivity gains, larger effects on lower-skill workers, but uncertain implications for total output and employment.
Displacement. Several occupations have begun to show measurable contraction. Translation and copywriting have seen freelance-rate declines and redundancies. Junior software engineering hiring fell in 2024–2025. Customer-service agent roles are being automated at the rate described above. Voice acting, illustration and stock photography have faced significant disruption. Whether these effects are temporary (workers shift into AI-augmented roles) or permanent (workers exit the labour market) remains to be seen.
Universal Basic Income (UBI) debates. Sam Altman's funded UBI study with Open Research, completed in 2024, found that giving low-income Americans $1,000 per month for three years produced modest improvements in financial wellbeing, food security and mental health, but did not transform recipients' employment trajectories. The results are open to interpretation, UBI advocates emphasise the wellbeing gains; sceptics note the absence of dramatic effects. The political debate over UBI is unlikely to be resolved by any single study.
We do not yet know the magnitude of AI's effect on labour markets. The macro effects so far are smaller than the most aggressive forecasts and larger than the most dismissive. The composition of effects across occupations is uneven and partly contingent on policy choices. The 5–10 year horizon will produce empirical evidence that the next decade of debate will draw upon.
Sectoral exposure analyses in 2023–2025, including the OpenAI–Penn paper "GPTs are GPTs" (Eloundou and colleagues, 2023) and McKinsey's Generative AI report, identified the occupations with highest LLM exposure as: writers and translators (90%+ of tasks), legal professionals (50%), business and financial operations (40%), software developers (30%), and management (20%). Lower-exposure sectors include personal care, food preparation, construction, and many trades. The exposure measures, however, are based on theoretical task-level overlap with LLM capability and do not directly translate to employment effects; the actual labour-market response depends on substitution versus complementarity, organisational adoption pace, and regulatory constraints.
Policy responses have begun to take shape. The UK Pissarides Review (2024) recommended significant investment in worker retraining and adjustment support for AI-displaced workers. The EU AI Act's transparency provisions for AI-generated content and its prohibitions on social-scoring and certain workplace-monitoring uses of AI implicitly shape the labour-market impact. The US AI Bill of Rights (2022) and the various state-level acts (Colorado AI Act 2024, California SB 1047 vetoed 2024) are operating in the same space. The next decade of AI policy will determine whether the productivity gains accrue broadly or narrowly.