Research behind the courses

We used the studies below when planning the courses. Some found that AI helped people finish work faster. Others found slower work or more wrong answers. The differences matter when deciding what to teach and what to check.

We built our courses on this research, so we have an interest in how it reads. Every figure links to its source. Check them.

Demand

Demand for AI skills is rising while overall hiring falls. PwC's 2025 Global AI Jobs Barometer, built from close to a billion job adverts, found postings for AI-skilled roles rose 7.5% year on year while total postings fell 11.3%, and that the skills employers ask for are changing 66% faster in AI-exposed occupations. The 2026 edition reports the trend accelerating.

Even here, single numbers disagree. PwC's 2025 Barometer puts the wage premium for AI-skilled roles at 56%, while Lightcast (2025), analysing over 1.3 billion postings with a different method, finds 28%. Both are defensible measurements of different comparisons, so the premium is real and its size depends on the method.

In a spring 2026 NACE survey, 185 employers said more than a third of entry-level jobs required AI skills, nearly three times the share they reported in autumn 2025. That tells us what these employers perceived. Studies that count job adverts find much lower shares, so we would not use the survey figure as a count of available jobs.

The New York Fed's Liberty Street Economics analysis (2026) gives another reason to examine training. Roughly 39% of employed Americans said they used AI at work, while only around 16% said their employer offered AI training.

There is also a regulatory strand. Article 4 of the EU AI Act obliges providers and deployers of AI systems to take measures supporting staff AI literacy. It has applied since February 2025, was softened by the 2026 Digital Omnibus from ensuring a sufficient level of literacy to supporting its development, and has been under national market-surveillance supervision since August 2026. It is a best-effort duty for organisations in the Act's scope, with no dedicated fine schedule. So it's a genuine reason to train staff, but it is not a training mandate, and a provider that presents it as one is overstating it.

None of this shows that training works, only that demand is real. Whether training changes what people can do is a separate question, and the evidence there is less comfortable.

Measured results

The published research does not support a single universal productivity number for AI-assisted work. The shape of the task, the worker's familiarity with it and the cost of checking the output all change the result, sometimes past zero.

Defined tasks

A controlled experiment with 95 professional developers (Peng and colleagues, 2023) found the group using GitHub Copilot completed one defined JavaScript HTTP-server task 55.8% faster, with no significant effect on task success. The Harvard and BCG field experiment (Dell'Acqua and colleagues, Organization Science, 2026) gave 758 consultants, about 7% of BCG's individual contributors, 18 realistic consulting tasks selected to sit within AI capability. The AI group completed 12.2% more tasks, finished 25.1% faster and produced work rated around 30% higher in quality.

These are useful results for bounded work with a clear finish line. They are not forecasts for a whole delivery workflow.

Repository work

METR's randomised trial (early 2025) followed 16 experienced open-source developers completing 246 real issues in mature repositories they knew well. With AI tools allowed, they took 19% longer. Before the study they expected AI to make them 24% faster, and afterwards, having been measured slowing down, they still believed it had made them 20% faster.

That gap between belief and measurement is the most important finding on this page. Our tutor Lee Crossley calls the lost time review churn and expansion: plausible changes grow, regressions surface and the time saved generating code is spent establishing what is safe to keep.

METR's 2026 update says newer tools probably improve the result, but it could not produce a reliable current estimate. Developers increasingly declined to work without AI, avoided submitting tasks they didn't want to do manually, and couldn't time their work cleanly once agents let them do other things while waiting. METR published the caveat beside the number, and this page keeps them together for the same reason.

Model limits

The BCG researchers also set a task the model could not do reliably. It looked much like the other tasks in difficulty. Without AI, consultants got the answer right 84.5% of the time. With AI, 70.6% did. Among those given AI and a prompt-engineering overview, only 60.0% did. Dell'Acqua and colleagues called this the jagged frontier: looking at a task did not reliably tell workers whether the model could handle it. Those tasks were run in 2023, and the models have changed since then. Learning a fixed list of what AI could do in that experiment would not solve the problem for today's work.

Unchecked work

BetterUp Labs and the Stanford Social Media Lab asked 1,150 US employees about receiving poor AI-generated work from colleagues (Harvard Business Review, September 2025). Forty per cent reported receiving it in the previous month, with around two hours spent dealing with each instance. The authors called it workslop: output that looks finished but does not help the task along. Someone else then has to work out what it means, correct it or start again.

Deloitte Australia's report for the Department of Employment and Workplace Relations provides a specific example. An external academic found fabricated citations and an invented court quotation after the report had passed review. In October 2025, Deloitte agreed to repay the final instalment of the AU$440,000 contract, worth about US$290,000 (Fortune, 2025). AI-assisted content had reached the client with those errors still in it.

Microsoft Research collected 936 examples of AI use from 319 knowledge workers (Lee, Sarkar and colleagues, CHI 2025). People who expressed more confidence in the AI reported less critical thinking. Those more confident in their own skills reported more. They described the remaining effort as checking information, integrating responses and taking responsibility for the task, with less checking on routine work under time pressure. These are reports collected at one point in time. They show associations, but cannot establish what caused them.

Different tasks

Any of the figures above can be quoted alone to support almost any position: 55.8% faster sells tools, 19% slower sells scepticism. Both are careful measurements of quite different work (a defined greenfield task with an automated check on one side, real issues in large familiar codebases on the other). So the honest summary is conditional: AI-assisted work goes well where the task is well defined and checking is cheap, and where the work is ambiguous, unfamiliar or beyond the tool's capability, it costs time or correctness instead. And the people doing the work turn out to be poor judges of which situation they are in, which is what the METR belief gap and the BCG frontier result both show.

Course design

Prompting lessons

In the BCG experiment, the group trained in prompting did worse on the outside-frontier task than the group given AI with no training at all. That's one task in one experiment, and it should be held lightly. But it is a direct warning that teaching people to operate the tool, without teaching them to judge when its output should not be trusted, can make outcomes worse.

Observed work

In the GLAT study, people's ratings of their own AI proficiency did not predict their results on AI-assisted tasks. A performance test did (Jin and colleagues, 2024). We should be careful about extending that result: the participants were healthcare students, and the predictor only just passed the conventional significance threshold at p = .04. A 2024 review by Lintner found 13 AI-literacy measures based on self-report and three based on performance, with none of the performance measures covering generative AI at that time. Together with the CHI survey, this gives us a reason to look at what learners do as well as what they say they can do.

Other courses

We reviewed Google AI Essentials in August 2026, when it had around 1.9 million enrolments including free auditors. It covers fundamentals, productivity, prompting and responsible use, including spotting hallucinations within its modules. Learners take auto-graded quizzes at their own pace. There is no separate verification module or assessment of observed work. OpenAI's AI Foundations certification, with pilots announced from December 2025, is aimed at a similar general audience. In interviews about workplace training needs, Zhang and Lee (CHI 2025) found that workers wanted guided practice in verifying AI output, while much existing training was optional and focused on generic tool use.

We chose to spend our course time on the checking, delegation and recovery that receive less attention in those formats. Learners do prepared tasks live, without recording the sessions, and Lee reviews their work. He can then discuss the checks they ran and the decisions they made, including what they did when their first approach failed.

Our limits

We have not run a randomised trial of these courses, and we know of none showing that an AI training course improves work outcomes. A 2026 review in Computers in Human Behavior Reports links over-reliance on AI with deskilling. It suggests training as a possible response but does not test a course. Lau and colleagues (2026) developed a 13-item scale for critical thinking in AI use across six studies with 1,341 participants. It predicted accuracy in judging AI output, though this is new work from one research group.

The research gives us reasons to teach verification, problem definition, bounded delegation and recovery. It also gives us reasons to assess the work learners produce. It does not let us promise a productivity gain from attending. Lee only sees a limited amount of work during a course, so his feedback is not a certification of someone's overall capability.

We checked the figures against these links in August 2026. Please tell us if you find an error so we can correct it.

Last reviewed: 8 August 2026

Back to the courses