Every wave of automation has raised fears of mass unemployment — from the Luddites facing mechanised looms to economists debating computers in the 1960s. Generative AI renews the question with new force, because for the first time machines perform many cognitive and creative tasks previously done only by educated professionals. This lecture examines what we know, what we do not, and how to prepare.
Tasks, not jobs#
Economists find it more useful to analyse tasks than whole jobs (Autor, Levy and Murnane's task framework). A job is a bundle of tasks; technology automates some, augments others and creates new ones.
- Automation: AI performs a task instead of a person (transcribing audio, routing tickets).
- Augmentation: AI helps a person perform a task better or faster (drafting, coding assistance, decision support).
- New tasks: work that did not exist before (prompt design, AI evaluation and oversight, data annotation, AI governance).
Studies estimating exposure to large language models (e.g. Eloundou et al., 2023) suggest that a large share of workers in advanced economies have at least some tasks that LLMs could affect, with higher exposure among higher-wage, higher-education occupations — the reverse of earlier automation waves that mainly affected routine manual and clerical work. Exposure does not mean replacement: it indicates where work may change.
Early evidence on productivity#
Controlled studies of generative AI in real work have found meaningful gains on specific tasks:
- Customer support: a study of thousands of agents (Brynjolfsson, Li & Raymond, 2023) found an AI assistant increased resolutions per hour on average by around 14%, with the largest gains for novice and lower-skilled workers — the AI spread best practices.
- Writing tasks: an experiment (Noy & Zhang, 2023) found professionals completed writing tasks faster with higher rated quality when using ChatGPT.
- Consulting tasks: a field experiment with management consultants (Dell'Acqua et al., 2023) found large gains on tasks within the AI's capability — but worse performance on a task outside its "jagged frontier", where consultants over-relied on confidently wrong AI output.
- Software development: controlled experiments found faster completion of some coding tasks with AI assistants; effects in complex real-world settings vary.
Lessons: gains are real but uneven; AI can narrow skill gaps within tasks; and knowing when not to trust AI is itself a crucial skill.
Distributional effects#
Aggregate productivity gains do not guarantee broadly shared benefits. Concerns include:
- Wage and employment pressure in highly exposed occupations (translation, some writing, customer service, entry-level coding).
- Entry-level jobs that traditionally trained newcomers may shrink, disrupting career ladders.
- Concentration of gains among firms with data, compute and capital.
- Global effects: outsourced digital work (call centres, content writing, data entry) in lower-income countries may be affected — while AI tools may also open new opportunities for workers there.
- Invisible labour: AI depends on data annotators and content moderators, often poorly paid (see the data-labelling lecture).
Daron Acemoglu and others argue that the direction of technology is a choice: we can prioritise AI that augments workers and creates new tasks rather than AI that merely replaces labour.
Implications for students#
What organisations and policymakers can do#
- Organisations: involve workers in redesigning workflows; invest in training; use AI to augment and upskill; be transparent about monitoring; share productivity gains.
- Policy: education and reskilling programmes, social safety nets, portable benefits, labour protections for platform and data workers, competition policy, and support for AI that serves public needs.