Big question: If productivity rises, who gains time, income, skill, voice, and control?
Research lock: 2026-08-28
Why this week matters
Job-loss headlines compress several different questions: which tasks are exposed, which are technically feasible, which are economically adopted, how jobs are redesigned, who captures productivity gains, and which pathways into expertise survive. Exposure is not displacement. Productivity is not worker welfare. A system can augment output while intensifying surveillance or eroding autonomy.
Deeper teaching spine
1. Use tasks as the unit, jobs as bundles
List a job's tasks by frequency, importance, social context, tacit knowledge, error cost, and complementarity. Classify each as potentially automated, accelerated, quality-assisted, newly created, or unaffected. Then model how the bundle and headcount could change. This prevents a task benchmark from becoming a job forecast.
2. Separate five evidence claims
- Capability: can a system perform the task under test conditions?
- Exposure: could the task be affected given current capabilities?
- Adoption: will organizations integrate it under real cost, workflow, and regulation?
- Productivity: does output per unit input improve in the field?
- Employment/distribution: what happens to wages, hours, hiring, quality, and bargaining power?
Every source should be labelled with the claim it actually supports.
3. Read the ILO index carefully
The ILO's 2025 refined global index estimates that one in four workers globally are in occupations with some generative-AI exposure and emphasizes transformation over full automation as the more likely broad effect. Its highest exposure category covers a much smaller share, and exposure differs by income level and gender. This is a scenario about task overlap—not a count of jobs already lost.
4. Productivity can compress learning
Brynjolfsson, Li, and Raymond studied 5,179 customer-support agents and reported an average productivity increase of 14%, with substantially larger gains for novice and lower-skilled workers. That supports a knowledge-diffusion hypothesis in that setting. It does not establish the same effect across occupations, long-run skill formation, worker satisfaction, employment, or wages.
Use the finding to ask a harder question: if novices receive expert-like suggestions, do they learn faster, or become dependent on an interface whose rationale they cannot inspect?
5. Algorithmic management is a governance system
Scheduling, routing, productivity scoring, monitoring, ranking, and deactivation distribute information and power. Assess data accuracy, visibility of rules, explanation, contestation, pace, safety, collective voice, and whether the worker can override an unsafe instruction without penalty.
6. Creative labour shows bargaining as governance
The Writers Guild's negotiated AI provisions are a primary example of workers defining permitted use, credit, compensation, disclosure, and training-material positions through collective bargaining. Compare contract governance with legislation, technical provenance, and individual consent.
Case-study dossier
Customer support: augmentation with uneven gains
Give teams the NBER study abstract/method and ask what could explain larger novice gains: access to tacit knowledge, task standardization, model imitation of high performers, manager use, or selection. Design a follow-up measuring quality, retention, learning after tool removal, and worker autonomy.
Writers Guild: negotiated boundaries
Students read the official AI provisions, not a headline summary. Identify which questions are settled contractually, which remain legal, and which require technical verification. Draft an analogous agreement for educators, designers, or software developers.
Algorithmic scheduling
Use a hypothetical delivery platform that predicts demand and issues shifts. A 5% utilization improvement creates last-minute schedule changes. Ask students to value employer efficiency, worker income volatility, caregiving constraints, safety, and appeal. Require a transition design, not a yes/no automation vote.
Seminar activities
- Headline autopsy: replace “AI will eliminate X million jobs” with a chain of capability, exposure, adoption, task redesign, and employment assumptions.
- Job unbundling: map 20 tasks from a real role; identify entry-level tasks that train judgment and could disappear first.
- Productivity dividend negotiation: allocate gains among lower prices, profit, wage, shorter hours, staffing, training, and quality.
- Worker-impact assessment: design notice, data limits, explanation, appeal, monitoring, and bargaining provisions for an algorithmic-management tool.
Visual evidence plan
| Visual | Capture target | Teaching use | Guardrail |
|---|---|---|---|
| ILO exposure distribution | ILO 2025 report page | Capture the exposure gradient and gender/income comparison. | Label exposure, not displacement. |
| Productivity study | NBER Working Paper 31161 | Crop the 14% average and 34% novice/low-skill findings with sample context. | One firm and occupation do not establish economy-wide impact. |
| WGA AI provisions | official WGA page | Highlight negotiated rules on AI-generated material and writer choice. | Contract scope and bargaining unit matter. |
| Task-to-job map | Create from an actual occupation using the five evidence claims above. | Visually interrupt forecast determinism. | Make assumptions editable and dated. |
Reading and citation ledger
- ILO, Generative AI and Jobs: A refined global index of occupational exposure (2025).
- Brynjolfsson, Li & Raymond, “Generative AI at Work”.
- Writers Guild of America, Artificial Intelligence contract guidance.
- NIST AI RMF — use the socio-technical risk frame for workplace deployment.
- OECD AI Incidents and Hazards Monitor — search current labour/algorithmic-management incidents, then verify against primary records.
Watch list
- Forecasts age quickly; record model capability, occupation taxonomy, geography, and publication date.
- Measure quality and rework, not only speed.
- Track entry-level hiring and learning pathways separately from current-worker productivity.
- Ask workers what changed in discretion, pace, monitoring, and appeal—not only whether they “use AI.”