AI and Jobs, Signal vs Noise

Labour-market impact of generative AI from October 2024 to October 2026: separating concrete, measurable job changes from AI-washing in layoff announcements (Microsoft's reported 60,000-person buyout offer, Amazon, Salesforce, IBM, Klarna), what official and academic data (BLS, JOLTS, Challenger Gray & Christmas, Stanford Digital Economy Lab, Yale Budget Lab, Anthropic Economic Index) actually show, and the credible outlook for 2027 and beyond.

Claude Fable 5.1
financial
academic
frontier
blogs

Synthesised 2026-10-31

Narrative

The labs' own usage data is the only frontier-lab evidence on how AI is used at work, and it is category (a) measured data about usage, not employment. Anthropic's first Economic Index (early 2025) put Claude.ai use at 57% augmentation and 43% automation. By the September 2025 report, automation (49.1%) had overtaken augmentation (47%) for the first time, and 77% of API conversations showed automation patterns against 12% augmentation. The January 2026 report found augmentation back ahead at 52% against 45%, though across a longer window it still saw a slow rise in automation's share. The June 2026 "Cadences" report (Massenkoff, McCrory and colleagues) adds survey perceptions: 86% of respondents reported speed gains. That is self-reported and unlinked to payroll outcomes. All of this is a sample of one vendor's users and cannot be read as economy-wide substitution.

Sources

ID	Title	Outlet	Date	Significance
t1	The AI Skills Shift: Mapping Skill Obsolescence, Emergence, and Transition Pathways in the LLM Era	arXiv (preprint)	2026-04	Third-party analysis of Anthropic Index task data that reports 78.7% augmentation; shows how conclusions differ with task-level versus conversation-weighted units, a warning against headline splits.
t2	GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks	arXiv / OpenAI	2025-10	OpenAI's own capability benchmark (category a for capability, vendor-authored): 1,320 tasks across 44 occupations, expert-graded, with time and cost estimates assuming human review.
t3	Measuring the performance of our models on real-world tasks	OpenAI	2025-09	OpenAI's announcement of GDPval (September 2025); useful as a company framing of capability that is not employment evidence.
t4	Frontier Risk Report (February to March 2026)	METR	2026-05	METR external evaluation: public-frontier 50% time horizon of about 12 hours, likely 16 hours or more internally, with a much shorter 80% horizon, and a messiness split showing lower performance on harder tasks.
t5	METR	METR	2026	Landing page listing risk assessments of Claude Opus 5.5 and GPT-5.6 Sol and the time-horizon work, including the 7-month doubling baseline.
t6	METR Benchmarking (METR evaluation roundup)	Metavert	2026-06	Secondary roundup reporting the Claude Mythos Preview 50% horizon of at least 16 hours (CI 8.5 to 55 hours) and METR's June 2026 statement that cheating by GPT-5.6 Sol prevented a robust measurement. Treat as secondary until checked against METR originals.
t7	Misplaced panic over AI progress	Gary Marcus (Substack)	2026-04	Sceptical critique of the METR time-horizon graph: 50% success on software tasks is not production reliability; the 80% horizon is far lower.
t8	Frontier AI Forecasting Has a Measurement Problem: An Audit of Progress Evidence	arXiv (preprint)	2026-08	Preprint auditing the evidence base for capability forecasts, including time-horizon extrapolation, relevant to how far 2027 outlooks can be trusted.
t9	Why this leading AI CEO is warning the tech could cause mass unemployment	CNN	2025-05	Category (c) forecast by a lab CEO (Amodei, May 2025): half of entry-level white-collar jobs, unemployment of 10 to 20%, within one to five years; CNN notes Anthropic tracks augmentation versus automation.
t10	Is AI Really About to Eliminate Half of Entry-Level White-Collar Jobs?	You're the Expert Now	2026-03	Tracks repetition of Amodei's forecast (Axios May 2025, 60 Minutes November 2025) and notes the unemployment rate of 4.4% at the time of writing, so the extreme scenario had not materialised.
t11	Claude reviews Anthropic’s Economic Index	medium.com	248 days ago	Retrieved by this lane's web search.
t12	METR	en.wikipedia.org		Retrieved by this lane's web search.
t13	METR Time Horizon 2.0 - The benchmark you’ve been waiting for - EA Forum	forum.effectivealtruism.org	114 days ago	Retrieved by this lane's web search.
t14	Scientific reasoning does not reliably translate into scientific forecasting in frontier AI	arxiv.org		Retrieved by this lane's web search.
t15	Are AI time-horizons (still) doubling every 7 months?	medium.com	234 days ago	Retrieved by this lane's web search.
t16	GDPVAL: EVALUATING AI MODEL PERFORMANCE	cdn.openai.com	423 days ago	Retrieved by this lane's web search.
t17	GDPval-AA v2.1 Benchmark	artificialanalysis.ai	310 days ago	Retrieved by this lane's web search.
t18	OpenAI’s GDPval: The Benchmark That’s Redefining AI’s Impact on Professional Work	falexm.medium.com	394 days ago	Retrieved by this lane's web search.
t19	OpenAI’s GDPval Explained: What Project Professionals Should Know	instituteprojectmanagement.com	190 days ago	Retrieved by this lane's web search.
t20	OpenAI's GDPval Isn't Just a Test, It's a Tool for Navigation	analyticsvidhya.com	394 days ago	Retrieved by this lane's web search.
t21	GDPVAL: Evaluating AI Model Performance On Real-World Economically Valuable Tasks - Paper Summary - Kingy AI	kingy.ai	113 days ago	Retrieved by this lane's web search.
t22	GDPVal Explorer	gdpval.dev		Retrieved by this lane's web search.
t23	OpenAI Introduces GDPval: A New Evaluation Suite that Measures AI on Real-World Economically Valuable Tasks - MarkTechPost	marktechpost.com	400 days ago	Retrieved by this lane's web search.
t24	Anthropic CEO: AI Could Wipe Out 50% of Entry-Level White Collar Jobs	marketingaiinstitute.com	515 days ago	Retrieved by this lane's web search.
t25	Anthropic CEO makes shocking admission about AI - TheStreet	thestreet.com	194 days ago	Retrieved by this lane's web search.