Sources checked 27 October 2026. Nature gives a 26 October publication date; Crossref created the DOI record at 10:02 UTC that day, inside the coverage window. The author-hosted manuscript abstract describes 573 participants, while the methods report 986 allocated, 573 starting and 487 completing both tasks; AISFY reports each stage separately and uses the 487-person analysis sample. Sources: Nature Human Behaviour — ChatGPT and consultant reskilling study; Author-hosted full paper; Public reproducibility repository; Crossref DOI metadata.
26 October 2026, global workforce-learning evidence: A peer-reviewed study in Nature Human Behaviour found that consultants using ChatGPT performed better on unfamiliar technical tasks, yet the researchers found no evidence that this translated into stronger unassisted technical knowledge afterwards. For CEOs, COOs, CHROs, CAIOs, HR directors and L&D leaders, the buyer decision is concrete: is a programme meant to improve supported work, retained learning, or both?
The featured image is an editorial illustration and does not depict the study, its participants or its results.
What did the randomized study test?
The study published by Nature Human Behaviour examined whether generative AI could help professionals perform tasks outside their existing expertise. The researchers allocated 986 Boston Consulting Group consultants to an AI-supported group or a conventional-resources group. Of those allocated, 573 began the study and 487 completed both assigned tasks: 260 with ChatGPT access and 227 using resources such as Google, Stack Overflow and Khan Academy. A separate benchmark group included 44 BCG data scientists.
Participants first received 15–20 minutes of training on their assigned resources. Each then completed two of three 90-minute tasks involving coding and data cleaning, statistical interpretation or prediction. The tasks were deliberately designed so that human judgment was still necessary: using ChatGPT alone did not yield the correct answer. The full author-hosted paper describes the assignment, participant flow, measures and study limitations; the public reproducibility repository provides analysis materials.
AI access improved task performance, with different effects by task
Relative to the conventional-resources group, ChatGPT access increased the probability of a correct answer by 49 percentage points for coding, 20 points for statistics and 17 points for prediction. The reported 95% confidence intervals were 42–56 points, 15–25 points and 9–25 points respectively. These are measured differences in the study’s short technical tasks, not forecasts for other roles, tools or workplaces.
The results also show why “AI improved performance” is not a complete training claim. Participants were evaluated while using their assigned support, and the size of the advantage varied by task. A buyer therefore needs to define the target capability and the permitted support conditions before deciding what a post-course score means.
The post-study assessment did not show stronger unassisted knowledge
After the experiment, participants answered technical knowledge questions without the tool. The researchers found no evidence that the AI-supported group had gained more unassisted technical knowledge than the conventional-resources group across five measures (joint F(5,198)=0.83, p=0.531). The study also found greater overconfidence among AI-supported participants when they predicted which problems the model could solve correctly.
This does not prove that AI-supported learning is impossible. It establishes a narrower result: a short intervention that improved performance on the assigned tasks did not produce a detectable advantage on this unassisted post-study assessment. The study involved highly educated management consultants, narrow technical problems and no longer-term measure of workplace transfer, career transition or end-to-end data-science capability. It was not conducted in the UAE. The paper also discloses BCG employee time and internal resources as support, multiple BCG-affiliated authors and one OpenAI-affiliated author. The Crossref DOI record was created inside this edition’s coverage window.
A buyer test for performance and learning
AISFY editorial planning aid: This comparison applies the study to a training-purchase decision. It is a recommendation, not a result measured by the researchers.
| Buyer question | Assisted-performance evidence | Independent-learning evidence |
|---|---|---|
| What must improve? | Name the work task, approved tool and quality threshold. | Name the knowledge or judgment employees should retain without the tool. |
| How will practice be assessed? | Observe a realistic task with approved support and record output quality, review effort and escalation. | Use a new task or explanation without that support to test transfer rather than recall of the exercise. |
| Who reviews the result? | Assign a capable reviewer for errors, uncertainty and decisions that exceed the employee’s authority. | Ask the reviewer to identify misconceptions and decide whether further practice is needed. |
| What justifies expansion? | Consistent performance under the actual operating conditions. | Evidence that the intended knowledge or judgment survives when the support changes or is removed. |
AISFY reflection: define which capability the programme is buying
AISFY reflection: Supported performance can be valuable even when it is not the same as independent expertise. The risk is treating one as proof of the other. Before buying a workshop, decide which tasks employees may perform with AI, which judgments they must understand independently, who reviews their work and when the organization will assess each outcome.
Use an AI training needs analysis to define the role and capability gap, then plan workshop evaluation beyond attendance with separate evidence for immediate task performance and later transfer.
