All studies
Studies are listed by first author. The label shows how each one was done. A study that has not been through peer review is marked as such. Peer review means other researchers checked the work before a journal published it.
- Students did better with ChatGPT in practice and worse without it on the exam During practice, students with an ordinary GPT-4 chat interface got 48% higher grades and those with a hints-only tutor 127% higher. On the exam without AI, the ordinary-chat group scored 17% below students who never had AI. The hints-only group scored about the same as them.
- Shadow learning: how surgical trainees lost their way to learn In traditional surgery, trainees learned by doing more of the operation over time. Robotic surgery greatly reduced the trainee’s part in the operation. Only a minority became competent, through unofficial practice.
- Experienced developers were slower with AI and believed they were faster Tasks took 19% longer when AI was allowed. Before the study, the developers expected AI to save them 24% of the time. Afterwards they still believed it had saved about 20%.
- US Census: 18% of firms use AI in a business function 18% of US firms used AI in a business function in November 2025 to January 2026, or 32% when firms are weighted by their number of employees. 66% of users used AI only to help staff with tasks, and 2% of firms cut jobs because of AI.
- Fewer young people hired into AI-exposed jobs: US payroll data Employment of 22–25-year-olds in AI-exposed occupations is 19% below where it would be had it kept pace with less exposed peers. Experienced workers show no such gap. The gap comes mainly from fewer young people being hired.
- An AI assistant for 5,172 customer-support agents Agents with an AI assistant resolved 15% more customer issues per hour on average. Less experienced agents gained the most. The most experienced gained little and their quality dipped slightly.
- Which tasks can machines learn? 18,156 tasks scored Most occupations include at least some tasks that machine learning could do. Few occupations could be fully automated. Getting the benefit usually means redesigning what the job contains.
- After three months with AI, endoscopists found fewer growths without it In colonoscopies done without AI, the share in which doctors found at least one pre-cancerous growth fell from 28.4% to 22.4% within three months of AI being introduced. The doctors lost part of a skill that the tool was doing for them.
- Does readiness for change predict success? A review of 47 studies Whether an organization’s readiness for change leads to a successful rollout, and under what conditions, remains unclear. Studies define and measure readiness in many different ways, and most measure it only once.
- AI exposure and career moves in Brazil and the UK Young graduates often start in jobs that AI is likely to replace and move up to jobs that AI is likely to assist. AI could widen that path. It could also remove the starting jobs.
- The same AI pressure can raise or lower willingness to use AI When employees saw the demands of AI as a challenge they could meet, they felt positive and were more willing to use AI. When they saw the demands as obstacles, they felt anxious and were less willing.
- Across 160,097 developers, experienced programmers gained from AI coding tools and early-career ones did not AI wrote an estimated 29% of Python functions in the United States. Quarterly output, measured in code contributions, rose an estimated 3.6%. Experienced developers produced more and moved into new kinds of software; early-career developers showed no clear benefit.
- The jagged frontier: AI helped consultants on some tasks and hurt on another On 18 tasks that AI handles well, consultants using AI completed 12.2% more tasks and worked 25.1% faster, with better quality. On one task chosen to be beyond the AI’s ability, they were 19 percentage points less likely to get the right answer.
- Freelance job posts for writing and coding fell 21% after ChatGPT Within eight months of ChatGPT's release, posts for writing and coding jobs fell 21% compared with jobs that need manual skills. Posts for image creation fell 17% after image-generating AI arrived.
- Office AI assistant saved email time but did not change the work The 80% of workers given the tool who used it spent about two hours less on email each week and worked less outside regular hours. Apart from those time savings, the study found no change in how much work people did or what kind.
- Randomised experiments on a retail platform: AI raised sales most where there had been no service Sales effects ranged from none to 16.3%. The largest gain came from a pre-sale chatbot where shoppers had previously got no live help. Gains came mostly from more shoppers buying, and returns and ratings did not get worse.
- 46% of US small employer firms use AI; 31% of users report higher sales 46% of small employer firms said the business or its employees used AI for work. Among users, 71% reported higher productivity and 31% higher sales. 77% saw no change in labour costs, and 18% spent less on outside services.
- A measure of how exposed each occupation is to AI The study built and tested a score of how far each occupation relies on abilities where AI has made progress. The same data can be turned into scores for industries, regions and single firms.
- Small firms try AI less but use it as heavily: euro-area firm survey 41% of micro firms used no AI, against 13% of large firms. But the share of firms making significant use of AI was about the same at every size, at around 7–8%. Size goes with whether a firm starts, not with how far it goes.
- AI exposure and employment in 23 countries There was no clear link between how exposed an occupation was to AI and how its employment changed. In occupations with heavy computer use, more exposure went with faster employment growth.
- Generative AI and jobs: a global estimate from the ILO The study predicts that generative AI will mostly assist people in their jobs, not replace them. Clerical work is the most exposed, so richer countries and women are affected most.
- Startups most exposed to AI cut staff, mostly juniors, and grew faster Startups whose work was fully exposed to AI cut employment by 8% within two quarters of ChatGPT's release, mostly in junior and implementation roles. Their code contributions rose 64%, and they became more likely to rank among the startups with the highest sales per employee.
- AI decided which scans got one reader or two: more cancers found, 44% less reading Screening with AI found 6.4 cancers per 1,000 women, against 5.0 with two radiologists reading every scan. False positives did not rise. Radiologists did 44.2% fewer readings.
- Firms that put generative AI into their work hired fewer juniors From early 2023, junior employment at adopting firms fell about 9% relative to other firms within six quarters. Senior employment was largely unchanged. The fall came from slower hiring, not from people leaving.
- AI chatbots, earnings and hours: evidence from Denmark Two years after ChatGPT appeared, AI chatbots were widely used and workers reported benefits. Yet the study found no effect on earnings or recorded hours, and rules out effects larger than 2%.
- Fear of losing a job to AI and the hiding of knowledge Employees who feared AI would cost them their job were more likely to hide what they know from colleagues. Part of the reason was lower psychological safety. Confidence in one’s ability to learn AI softened the effect.
- AI adoption, psychological safety and depression among employees Where employees reported more AI adoption at their company, they later reported lower psychological safety, and then more symptoms of depression. Ethical leadership weakened the first link.
- OECD surveys of employers and workers on AI at work Workers and employers were generally positive about AI’s effect on performance and working conditions, with real worries about job loss. Where employers trained workers and consulted them, outcomes for workers were better.
- Workers who trust AI more report thinking less critically about its output Workers who were more confident in generative AI reported less critical thinking about its output. Workers who were more confident in their own ability to do the task reported more. With AI, the thinking shifted towards checking information and fitting AI output into the work.
- A review of 16 questionnaires that measure AI knowledge Of 16 questionnaires that measure AI literacy, 13 ask people to rate their own knowledge. Only 3 test it. None had been checked for whether it works the same way in different cultures.
- AI scribes in a randomised trial: one tool cut note time by 9.5%, and doctors used them in about a third of visits Doctors given Nabla spent 9.5% less time writing notes than doctors without a scribe. Doctors given DAX Copilot showed no clear change. Doctors used the scribes in about a third of their visits.
- Self-rated digital skills versus test results in 629 students Students’ ratings of their own digital skills differed from their test results by more than chance. Some were overconfident, others underconfident. Problem solving and handling information were weak even among the best.
- McKinsey’s 2025 global survey on AI in organizations 88% of organizations use AI in at least one business function, but only about 6% are “high performers” that credit AI with 5% or more of their operating profit. High performers are nearly three times as likely to have rebuilt workflows from the ground up.
- AI capability and company performance: a survey of 143 executives Companies whose executives reported stronger AI capability also reported more organizational creativity and better performance. The study produced a 46-question measure of AI capability.
- ChatGPT and professional writing tasks: an experiment Professionals who used ChatGPT finished writing tasks 40% faster, and the quality of their work rose by 18%. The gap between stronger and weaker writers narrowed.
- An AI business adviser helped strong small firms and hurt weak ones On average the AI adviser made no detectable difference to revenue or profit. Owners whose businesses were doing poorly at the start did nearly 10% worse with it. Those doing well may have done more than 15% better.
- Knowing how AI works improves decisions about what to hand over People who were taught how AI works handed over the tasks that suited AI and kept those that suited them. Their results improved. They also said they were less likely to use AI in future.
- Colleagues judge people who use AI as less competent People expected to be judged as less competent and less motivated if they used AI at work, and evaluators did judge AI users that way. The penalty also affected how evaluators rated job candidates.
- Ready for change, resistant, or both? Readiness for change and resistance to change are usually treated as two ends of one scale. The authors argue they are separate. A person can be high on both, or low on both.
- How automation eroded skills at an accounting firm Relying on automated software wore down accountants’ skills over several years. The loss went unnoticed by staff and managers. It became a problem because staff were still answerable for work they no longer fully understood.
- Digital skills at a financial company: employees versus supervisors Employees and their supervisors gave different accounts of the employees’ digital skills. The gaps were in cybersecurity, in using digital tools to streamline work, and in work-life balance.
- Exposure is not adoption: what AI could do versus what it is used for Scores of how exposed a job is to AI measure what AI could do. A measure of what AI is actually used for predicted demand for workers better.
- Experts decided better when the AI showed where it had looked With the picture, factory workers were 7.7 percentage points more accurate and radiologists 4.7 points more accurate than colleagues who saw only the score. Most of the factory workers with the picture did better than the AI alone.
- Whose view does a digital maturity score reflect? Many digital maturity models mainly record what senior managers think about progress, not measurable results. Most are built for large firms. The authors propose a model that asks general employees about day-to-day behaviour instead.
- Freelance demand for writing and translation fell 20–50% after ChatGPT After ChatGPT launched, demand for freelance skills it can replace, such as writing and translation, fell 20–50% below the trend expected without it. Short-term jobs fell most. Demand for building AI chatbots nearly tripled, and demand for novice freelancers fell.
- Ten years of digital maturity models: do they predict results? It is widely assumed that a higher digital maturity level means better performance. The review found that researchers question this: the models lack objective scoring criteria, and data linking maturity to performance is missing.
- Small firms that adopted AI hired faster: payroll records from Gusto Four quarters after owners said they adopted AI, their firms' headcount was 6.8% higher than at comparable firms that had not adopted it. Firms with fewer than 10 employees grew about 10% more. Firms with 10–24 employees showed no measurable change.
- When a human plus AI beats either alone: 106 experiments On average, a human working with AI did worse than the better of the two working alone. The combination lost on decision tasks and gained on tasks that involve creating content.
- Which barriers hold back AI maturity? A survey of 108 specialists Barriers to do with skilled employees had the strongest effect on AI maturity. They held it back directly, and they made infrastructure and data problems look worse. Infrastructure and data barriers had no direct effect larger than chance.
- Bank records: 17.7% of small US firms had paid for AI by end-2025 By December 2025, 17.7% of small firms had paid for an AI service at least once, up from 1.7% in January 2019. The median paying firm spent about $28 a month on AI in 2025. Firms with employees paid for AI more often than sole traders.
- Leaders use AI daily while their organizations stay in pilots 83% of the leaders surveyed use AI every day, but roughly 70% of their organizations have not progressed beyond pilots. Leaders who knew more about AI judged their own organizations more harshly.
- 6,000 executives in four countries: 69% of firms use AI, nine in ten see no effect yet 69% of firms actively used AI. Nine in ten executives reported no effect of AI on their firm's employment or productivity over the past three years. For the next three years they expect productivity to rise 1.4%, output 0.8% and employment to fall 0.7%.
- German small and medium-sized firms: skills and innovation predict AI use better than size 20% of German small and medium-sized firms used AI in 2022–2024, five times the share in 2016–2018. Use rose from 19% in firms with fewer than 5 employees to 36% in firms with 50 or more. Once graduates, innovation, research and digital strategy were counted, size and sector mattered little.