“AI substantially increases individual knowledge-worker productivity.”

• Randomized controlled trial published in Science (Noy & Zhang, 2023): 453 college-educated professionals doing realistic mid-level writing tasks finished ~40% faster and produced output rated ~18% higher in quality when given ChatGPT; inequality in output quality also fell as weaker writers gained most.
• Field experiment with GitHub Copilot (Peng et al., 2023): developers assigned an AI coding assistant completed a standardized programming task 55.8% faster than the control group.
• Harvard Business School / BCG field experiment (Dell'Acqua et al., 2023) with 758 consultants: for tasks inside the AI's capability frontier, GPT-4 users completed 12.2% more tasks, 25.1% faster, with results judged over 40% higher quality — one of the largest real-workforce experiments to date.
• Brynjolfsson, Li & Raymond (NBER WP 31161; later Quarterly Journal of Economics): 5,179 customer-support agents using a generative-AI assistant resolved ~14-15% more issues per hour, with gains of ~30%+ for the least experienced workers — evidence from a real firm over a year, not a lab task.
• Effects replicate across heterogeneous task types (writing, coding, consulting, customer support) and across lab and field settings, with effect sizes typically in the 10-55% range — a consistency uncommon in productivity research.
• Survey evidence confirms wide use with self-reported time savings: Bick, Blandin & Deming find a large and rapidly growing share of US workers using generative AI at work, with users reporting meaningful reductions in task time.

• METR's 2025 randomized trial of 16 experienced open-source developers working on their own large, mature repositories found they were ~19% SLOWER with early-2025 AI tools — while believing they had been ~20% faster. This shows task-level lab gains do not automatically transfer to complex real work, and that self-reported productivity gains are unreliable.
• The same BCG/HBS study found that for tasks OUTSIDE the AI's capability frontier, consultants using GPT-4 were about 19 percentage points LESS likely to produce correct solutions than the control group — the effect is conditional, not universal.
• Measured economy-wide time savings are modest: Bick, Blandin & Deming estimate generative AI saves users roughly 5% of their work hours, implying only ~1% of aggregate work hours saved — far below what 'substantial' implies for a typical worker's whole job rather than an isolated task.
• Acemoglu's macroeconomic assessment projects only ~0.5-0.7% total factor productivity gain over ten years (~0.07%/year), arguing that only a small share of tasks are both AI-exposed and profitably automatable in the near term.
• Aggregate US labor-productivity statistics show no clearly identifiable AI-driven break in trend; firm-level AI adoption (per Census BTOS) remains a minority of firms, concentrated in information and professional services.
• Much of the strongest evidence comes from short, well-defined, individually-scored tasks with clear quality rubrics — conditions that favor AI and may overstate gains in ambiguous, collaborative, or accountability-heavy knowledge work. Emerging work on low-quality AI output ('workslop') documents verification and rework costs shifted onto colleagues.
• The claim is unqualified; evidence supports 'substantially increases productivity on many knowledge-work tasks, especially for less-experienced workers,' which is a narrower statement.

78/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Science (AAAS) — Noy & Zhangjournal · 96 · arXiv — Peng, Kalliamvakou, Cihon, Demirer (GitHub/Microsoft)journal · 72 · Harvard Business School Working Paper 24-013 (Dell'Acqua et al.)edu · 86 · National Bureau of Economic Research — Brynjolfsson, Li & Raymondjournal · 92 · METRother · 74 · National Bureau of Economic Research — Bick, Blandin & Demingjournal · 90 · National Bureau of Economic Research — Acemoglu, 'The Simple Macroeconomics of AI'journal · 88 · Federal Reserve Bank of St. Louisgov · 88 · U.S. Census Bureau (Business Trends and Outlook Survey, AI supplement)gov · 93 · Harvard Business Reviewnews · 68

• Randomized experimental evidence shows large gains on common knowledge-work tasks: in a Science study of college-educated professionals doing writing tasks, access to ChatGPT cut completion time by about 40% and improved output quality, especially for lower-performing participants.
• A large real-world field study in a customer-support setting found generative-AI assistance increased worker productivity by about 14% on average, with much larger gains for less-experienced and lower-skilled workers.
• An experiment with consultants found AI assistance improved speed and quality on tasks within the model's capability frontier: consultants completed more tasks, worked faster, and produced higher-quality outputs.
• Controlled evidence on programming tasks found developers using GitHub Copilot completed a coding task substantially faster than a control group, suggesting material productivity gains for at least some software work.

• The evidence is not uniform across all knowledge work: studies consistently find heterogeneity by task, worker skill, and whether the task is inside or outside the AI system's competence frontier.
• In the consultant experiment, when tasks were outside the AI frontier, participants using AI performed worse, showing that AI can reduce productivity or quality when users over-rely on plausible but wrong outputs.
• Some newer field evidence on experienced open-source developers found AI tools slowed participants down on selected real-world coding tasks, illustrating that gains do not automatically generalize to expert work or complex codebases.
• Most strong evidence is task-level or firm-specific rather than a settled measurement of economy-wide productivity; adoption, workflow redesign, verification costs, and error correction all affect realized gains.

84/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Sciencejournal · 95 · Quarterly Journal of Economicsjournal · 95 · SSRN / Harvard Business School authorsother · 82 · GitHubother · 70 · METRother · 72

• A randomized controlled trial published in Science (Noy & Zhang, 2023) found that using generative AI decreased the time professionals took to complete mid-level writing tasks by 40% while independently raising the quality of the output.
• An NBER study (Brynjolfsson et al., 2023) analyzing over 5,000 customer support agents found AI tooling increased issue resolution rates by 14% on average, with novice or low-skilled workers experiencing a 35% productivity boost.
• A joint field experiment by researchers at Harvard Business School and Boston Consulting Group (BCG) observed that management consultants using AI finished 12.2% more tasks, completed them 25.1% faster, and produced 40% higher quality results on creative product innovation tasks compared to a control group.

• Productivity gains depend heavily on the task's location relative to AI's 'jagged technological frontier'. The Harvard/BCG study found that when knowledge workers used AI for complex tasks outside its current capabilities, they were 19% less likely to produce correct solutions because they relied on hallucinated or subtly incorrect outputs.
• The NBER study noted that the most highly skilled and experienced workers saw zero productivity gains—and in some cases a slight drop—when using AI, indicating it currently serves more to upskill novices than to multiply the efficiency of established top performers.
• For highly specialized coding or domain-specific tasks, the time required to prompt, verify, review, and debug AI-generated errors can sometimes offset the initial speed of generation, resulting in negligible net productivity gains.

95/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Sciencejournal · 95 · National Bureau of Economic Researchedu · 90 · Microsoftother · 75

• A randomized field experiment involving 758 Boston Consulting Group consultants found that GPT-4 access increased task completion by more than 12%, speed by more than 25%, and human-rated performance by more than 40% on tasks within the model’s capability frontier.[1]
• A six-month, cross-industry randomized field experiment involving thousands of Microsoft 365 users found that AI access reduced time spent on email and moderately accelerated document completion, though effects differed by behavior and task.[2]
• Research summarized by MIT Sloan reports productivity gains of about 14% in a customer-service field study, with especially large benefits for less-experienced workers.[3]

• The strongest positive results are often task-specific rather than proof of a broad increase in total individual productivity across all knowledge work; the Microsoft field experiment found no significant reduction in meeting time.[2]
• The consulting experiment found that AI could improve performance on tasks within its capability frontier but could also cause users to perform worse when working outside that frontier, limiting the claim’s generality.[1]
• AI access does not automatically translate into equivalent gains for every worker: the Microsoft experiment reported modest intent-to-treat effects because adoption and usage varied, while the observed gains were larger among frequent users.[2]
• The available evidence primarily measures selected tasks, time, or rated output rather than durable, organization-wide productivity and long-term knowledge quality; therefore, “substantially increases individual knowledge-worker productivity” is broader than the evidence directly establishes.[1][2][3]

82/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Harvard Business Schooledu · 96 · Microsoft Researchother · 91 · MIT Sloan School of Managementedu · 90

• Multiple RCTs find large individual gains on knowledge-work tasks: writers using ChatGPT finished tasks ~40% faster with higher quality (Noy & Zhang, Science 2023).
• Customer-support agents using a generative-AI assistant handled ~14% more chats per hour, with larger gains for less-experienced workers (Brynjolfsson, Li & Raymond, NBER 2023).
• BCG consultants using GPT-4 completed 12% more tasks, 25% faster, and at ~40% higher quality when work was inside the model’s capability frontier (Dell’Acqua et al.).
• Controlled GitHub Copilot trials reported developers completing coding tasks ~55% faster.

• Effects are heterogeneous and task-dependent: high-skill workers often gain less, and performance can fall when tasks sit outside current AI capabilities (the ‘jagged frontier’ documented in the same consultant RCT).
• Experimental tasks are typically narrow (emails, tickets, well-specified coding); complex judgment, coordination, and verification work may see smaller or zero net gains.
• Time spent prompting, checking hallucinations, and correcting errors can offset raw speed-ups; some later developer studies report slowdowns even when users believe they are faster.
• Macro productivity data have not yet shown a corresponding surge, so lab/field individual gains may not fully translate to sustained personal output.

78/100 — the lab’s own confidence in the score above, not a second reading of the claim.

Science/AAASjournal · 96 · National Bureau of Economic Researchother · 90 · SSRN/Harvard Business Schoolother · 82 · arXiv (GitHub/Microsoft Research)other · 78