COAD Initiative — coad.life
← Home

AI & Job Displacement
Evidence Tracker

A living record of research, policy statements, and industry data on the impact of artificial intelligence on employment in Australia and globally — curated for the Citizens Of Australia Dividend initiative.

The measurement problem: Every official Australian figure on AI job displacement — including the "4% of jobs" and "600,000 workers" numbers repeated across government, media, and industry — comes from a single 2025 study that used pre-agentic AI as its capability baseline. That study explicitly excludes agentic AI systems and warns its scores cannot be used to forecast job losses. Emerging 2026 research using a workflow-level methodology finds exposure rates of 93.2% for knowledge-work roles — an order-of-magnitude difference driven entirely by the shift from task-level to agent-level AI. This page tracks both the official record and the growing body of evidence that renders it structurally obsolete.
⚠ Official baseline — JSA 2025
~4%
of Australian occupations scored ≥0.7 automatability under current Gen AI
Frequently reported as "~600,000 workers" — a journalist's arithmetic (4% × 14.8M), not a figure stated by JSA or the Productivity Commission.
  • AI baseline: generic Gen AI, circa 2023–2025 — predates agentic systems by 2–3 capability generations
  • Measures task-level exposure scores, not job losses or displacement forecasts
  • JSA explicitly states scores "do not provide direct insights on labour market outcomes"
  • Population weights from ABS 2021 Census — four years old at time of publication
  • Agentic AI systems placed explicitly out of scope
  • 998 occupations mapped; no modelling of firm adoption rates, retraining, or new job creation
vs
↗ Emerging evidence — Gupta & Kumar 2026
93.2%
of information-intensive occupations at moderate-to-high agentic AI displacement risk by 2030
Preprint (arXiv:2604.00186, April 2026) — peer review in progress. Applies to knowledge-work roles, approximately 40–45% of the Australian workforce (~6M workers). At 93.2%, that is roughly 5.6 million Australians in at-risk roles.
  • Measures workflow-level displacement — agentic AI automates entire job functions, not single tasks
  • Agentic Task Exposure (ATE) score captures multi-step planning, tool use, memory, and reasoning
  • Workflow Coverage Factor (COV) penalises roles requiring human coordination — rewarding well-bounded digital jobs
  • Represents a fundamentally different and more realistic methodology for post-2025 AI systems
  • Not yet peer-reviewed; treat as important emerging evidence, not settled science
Why the gap is so large
The official 4% and the emerging 93.2% are not measuring the same thing. The JSA asked: "Can current AI perform this specific task within a job?" Gupta & Kumar asked: "Can agentic AI operate this entire job function end-to-end?" As AI transitions from tools that assist humans with tasks to autonomous agents that complete workflows, the task-level measurement framework does not just underestimate risk — it becomes structurally the wrong question. The 4% figure is not wrong for what it measured. It is wrong as a guide to where we are now.
— Evidence entries tracked
— Verified
— Amended / challenged
— Pending verification
Filter:

Showing of entries