A controlled experiment published in Science found that professionals using ChatGPT finished writing tasks 40% faster and produced work rated 18% higher in quality. A study in The Quarterly Journal of Economics found that customer service agents using AI resolved 15% more issues per hour. And a paper in Science Advances found AI boosts individual creativity while shrinking the diversity of ideas people produce.
Those three findings alone should change how every HR leader thinks about AI. But they are just the beginning. Over the past two years, a wave of rigorous scientific research has produced clear answers about how AI affects worker productivity. This comes from peer-reviewed, controlled experiments published in the world's top scientific journals.
The picture that emerges is more complicated than anyone predicted. AI helps some workers enormously and others barely at all. It closes performance gaps between junior and senior employees. And it creates a dangerous illusion that leads workers to believe AI is helping them even when it is not.
Here are 33 findings from the scientific literature, organized by what they tell us about the real relationship between AI and employee productivity. Every source cited is a peer reviewed journal article.
AI Makes Writing Faster and Better
1. Writing speed jumped 40% with AI assistance. MIT researchers Noy and Zhang (2023) published a pre registered experiment in Science involving 453 college educated professionals. They assigned occupation specific writing tasks to marketers, grant writers, consultants, data analysts, HR professionals, and managers, then randomly gave half of them access to ChatGPT. The AI group completed tasks 40% faster.
2. Output quality rose 18%. In the same Science study, independent evaluators who were experienced professionals in the same fields rated AI assisted work 18% higher in quality. Speed and quality improved together, which is uncommon with workplace interventions.
3. The productivity gap between strong and weak writers shrank. Noy and Zhang found that ChatGPT compressed the productivity distribution. Workers with weaker writing skills saw the biggest improvements. The strongest writers saw smaller gains. This is a consistent pattern across multiple AI productivity studies: the technology acts as a skill equalizer.
4. AI restructured how people allocate their work time. The Science paper documented a shift in task allocation. With AI, workers spent less time on rough drafting and more time on idea generation and editing. AI didn't speed up the existing workflow. It changed the workflow itself. Workers became editors and creative directors rather than production workers. That is a different job requiring different competencies.
5. Exposure to AI increased real world adoption. Workers who used ChatGPT during the experiment were twice as likely to report using it in their actual jobs two weeks later, and 1.6 times as likely two months later. Brief exposure to AI was enough to shift long term work habits. This has direct implications for how organizations design AI pilot programs.
Related: How Artificial Intelligence Can Shape the Workplace
Knowledge Workers See Large but Uneven Gains
6. Consultants using AI completed 12.2% more tasks and finished 25.1% faster. Dell'Acqua et al. (2023), in a study accepted at Organization Science, conducted a pre registered experiment with 758 consultants from Boston Consulting Group. Researchers from Harvard, Wharton, and MIT randomly assigned participants to use GPT 4 or work without AI on 18 realistic consulting tasks. The AI group outperformed on every measure.
7. Output quality increased by over 40%. Independent evaluators in the Dell'Acqua et al. study rated AI assisted work more than 40% higher in quality. This wasn't a marginal difference. The effect sizes were large across creative, analytical, writing, and persuasion tasks.
8. Junior consultants improved 43%, senior consultants improved 17%. The bottom half of performers saw the largest gains. Junior consultants showed a 43% improvement in task performance compared to 17% for senior consultants. AI compressed the gap between experienced and inexperienced workers, a pattern that shows up in every major study in this literature.
9. AI made consultants 19 percentage points worse on tasks outside its capability. When researchers gave consultants a task requiring analysis of a spreadsheet and interview data, those using AI performed 19 percentage points worse than those without it. Dell'Acqua et al. coined the term "jagged technological frontier" to describe this. AI has an invisible boundary between tasks it does well and tasks it does poorly, and workers cannot reliably tell where that boundary is.
10. Two distinct strategies for effective human AI collaboration emerged. The study identified "Centaurs" who split work between themselves and AI along clear lines, and "Cyborgs" who blended their work with AI at every step. Both patterns produced good results inside the frontier. The workers who performed worst were those who handed tasks to AI without evaluating whether it was the right tool for the job.
Related: Artificial Intelligence for HR: A Comprehensive Guide
Customer Service: The Strongest Scientific Evidence
11. Customer support agents resolved 15% more issues per hour with AI. Brynjolfsson, Li, and Raymond (2025) published their landmark study in The Quarterly Journal of Economics, one of the top five economics journals in the world. They studied the staggered introduction of a generative AI assistant among 5,172 customer support agents at a Fortune 500 software company. Agents with AI access resolved 15% more customer issues per hour.
12. Less experienced agents improved by 30%. The QJE study found that less skilled and less experienced workers improved across all productivity measures, including a 30% increase in issues resolved per hour. The most experienced agents saw small speed gains and small quality declines. AI helped the people who needed help most.
13. Two month agents with AI matched six month agents without it. Brynjolfsson et al. found that agents with two months of tenure who used AI performed as well as agents with more than six months of tenure who did not have AI. The technology effectively compressed months of on the job learning into weeks. For any HR leader dealing with high turnover in customer facing roles, this single finding should shape your AI strategy.
14. Customer sentiment improved. The QJE study reported that AI assistance improved customer satisfaction scores. The AI system provided real time suggestions for how to respond, and agents were free to edit or ignore them. Customers rated their interactions higher when their agent had AI support.
15. Employee retention increased. Brynjolfsson et al. found that access to AI was associated with reductions in employee turnover. The researchers hypothesized that AI made the job less stressful by providing guidance on difficult conversations. This is one of the few peer reviewed studies to measure AI's effect on retention, and the result matters because contact centres have among the highest turnover rates of any industry.
16. The AI appeared to disseminate tacit knowledge from top performers. The QJE paper argues that the AI system captured patterns from the most productive agents and made those patterns available to everyone. This is significant because tacit knowledge, the kind of expertise that experienced workers carry in their heads but can't easily explain, has historically been the hardest form of organizational knowledge to transfer.
Related: Human Resources Management (HRM) - Everything You Need To Know
Creativity: Individual Gains, Collective Losses
17. AI enhanced stories were rated more creative, better written, and more enjoyable. Doshi and Hauser (2024), published in Science Advances, ran an experiment where 293 participants wrote short stories with or without AI generated ideas, and 600 evaluators judged the results. Stories written with access to AI ideas scored higher on creativity, writing quality, and enjoyability.
18. Less creative writers gained the most from AI. The Science Advances study found that writers with the lowest creativity scores (measured by a validated Divergent Association Task) benefited most from AI ideas. Access to five AI ideas improved their novelty by 10.7% and usefulness by 11.5%, making their output comparable to that of the most creative participants.
19. But AI assisted stories were more similar to each other. Doshi and Hauser found that stories created with AI assistance were more semantically similar to one another than stories written without AI. This creates what the authors describe as a social dilemma: each individual writer is better off using AI, but collectively the output becomes more homogenized. If an entire department uses the same AI, the range of ideas narrows.
20. AI now matches human creativity in head to head tests. As Rafner et al. (2023) noted in Nature Human Behaviour, state of the art generative AI now matches humans on creativity tests. The question is no longer whether AI is creative, but what happens to human creativity when AI is always available.
How Many Jobs Does This Affect?
21. About 80% of the US workforce could have at least 10% of their tasks affected by AI. Eloundou et al. (2024), published in Science, developed a framework for evaluating how large language models relate to the tasks workers perform. They found that roughly 80% of US workers could have at least 10% of their work tasks affected, while about 19% could see 50% or more of their tasks impacted.
22. Higher income jobs face greater AI exposure. The same Science paper found that the influence of AI spans all wage levels, but higher income jobs actually face greater exposure. This is the opposite of previous waves of automation, which primarily affected lower wage, routine manual work. AI targets the cognitive, analytical, and communication tasks that knowledge workers perform.
23. AI satisfies the criteria for a general purpose technology. Eloundou et al. concluded in Science that generative AI exhibits the key characteristics of a general purpose technology, similar to electricity or the internet. Its effects are not confined to a single industry or occupation. It has the potential to reshape work across the entire economy. That conclusion, published in one of the world's most selective journals, carries significant weight.
24. AI adoption among US firms is concentrated in larger, more productive companies. McElheran et al. (2024), published in the Journal of Economics and Management Strategy, found that AI adoption in the US is concentrated among larger and younger firms with relatively high productivity. Smaller firms lag behind, which means the productivity gains from AI may widen existing gaps between organizations rather than closing them.
Related: 30+ ChatGPT Prompts for HR to Accelerate Your Productivity How Artificial Intelligence Can Shape the Workplace
The Warning Signs Scientists Found
25. Human AI collaborations often underperform either humans or AI working alone. Vaccaro, Almaatouq, and Malone (2024), published in Nature Human Behaviour, conducted a pre registered meta analysis of 106 experimental studies reporting 370 effect sizes. They found that, on average, human AI combinations performed significantly worse than the best of humans or AI alone. The problem is not the technology. It is the collaboration. People accept AI suggestions when they shouldn't and override them when they should.
26. Human AI teams lost performance specifically on decision making tasks. The Nature Human Behaviour meta analysis found that human AI performance losses were concentrated in decision making tasks. In contrast, content creation tasks showed significantly greater gains from human AI collaboration. This distinction matters: if your team uses AI for writing and brainstorming, expect gains. If they use it for analysis and judgment calls, expect problems unless you invest heavily in training.
27. Workers who use AI face a social evaluation penalty. A 2025 study in PNAS involving over 4,400 participants across four experiments found that workers who use AI tools face negative judgments about their competence and motivation from colleagues. People who use AI are seen as less skilled and lazier, even when their output is objectively better. This creates a paradox: productivity enhancing tools can simultaneously improve performance and damage professional reputation.
28. Workers overestimate their AI productivity gains. This finding appears across multiple journal published studies. In the Noy and Zhang Science study, workers' self assessments of quality improvement did not match evaluator ratings. In the Dell'Acqua et al. study accepted at Organization Science, consultants overrode their own good judgment because they trusted AI output that was wrong. The Vaccaro et al. meta analysis in Nature Human Behaviour showed that people systematically misjudge when AI is actually helping. Self reported productivity surveys are unreliable measures of AI's real impact.
29. AI could widen inequality between firms. As Capraro et al. (2024) argued in PNAS Nexus, generative AI could increase productivity and promote shared prosperity when used to complement human efforts. But the benefits and costs will be distributed unevenly across firm sizes, sectors, and worker demographics. Their interdisciplinary review, involving over 30 researchers, warned that without deliberate policy, AI adoption is more likely to amplify existing inequalities than reduce them.
What the Science Tells HR Leaders to Do
30. The scientific evidence points to a consistent pattern: large gains on some tasks, real losses on others. Across every peer reviewed study cited here, the same pattern repeats. AI delivers substantial productivity gains for structured tasks such as writing, customer service, and content creation. It degrades performance on complex analytical work outside its capability boundary. It compresses skill gaps by helping weaker performers more than strong ones.
For HR professionals, the science points to five specific actions.
First, map your tasks before deploying AI. The jagged frontier from the Dell'Acqua et al. study is the single most important concept in the AI productivity literature. AI performance is task specific, not role specific. Map which tasks in each role are inside the frontier and which are outside it. Deploy AI selectively.
Second, invest in AI judgment training, not just AI tool training. Every peer reviewed study that measured self assessment found that workers overestimate AI's helpfulness. The Vaccaro et al. meta analysis in Nature Human Behaviour showed this systematically. People need training in when to use AI, when to override it, and when to set it aside entirely.
Third, use AI to compress the learning curve. The Brynjolfsson et al. study in The Quarterly Journal of Economics found that AI collapsed months of experience into weeks for new customer service agents. The Noy and Zhang study in Science found the same for writing. If you have roles with steep learning curves and high early turnover, AI can pay for itself through faster onboarding alone.
Fourth, measure productivity objectively. Self reported productivity gains are unreliable. The science is unambiguous on this point. Build measurement systems that track actual output, quality, and error rates alongside speed. The studies published in Science and The Quarterly Journal of Economics that found real gains all used objective metrics. The studies that found overconfidence relied on self assessment.
Fifth, protect creative diversity. The Doshi and Hauser study in Science Advances and the Boussioux et al. study in Organization Science both show that AI raises the average quality of creative work while reducing its variety. If your organization depends on breakthrough ideas, consider using multiple AI tools, or deliberately preserving some purely human creative processes, to maintain the range of thinking your teams produce.
The scientific evidence on AI and productivity is still young. Most of the studies cited here were published between 2023 and 2025. Two years from now, we will know far more. But what we know today, from controlled experiments published in the world's leading journals, is already enough to make better decisions than those driven by hype, fear, or vendor promises.
