What happened

On 27 August 2026, OpenAI described a randomised experiment conducted with researchers at Bocconi University. More than 1,000 first-year students worked on a marketing case with ChatGPT access, causal-reasoning training, both or neither.

ChatGPT access improved the assessed quality of their submissions. The reasoning exercise produced a wider range of distinctive ideas. Students receiving both showed complementary benefits.

This does not establish that AI cannot support originality, or that the same results will appear in every workplace. It is a study of a particular task and population. It does invite a useful question about how organisations recognise good thinking.

Why it matters

A polished answer is easy to reward because its strengths are immediately visible. Clear structure, fluent language and an extensive list of recommendations can make a proposal look finished. The harder question is whether the author understands why the recommendation should work.

In a workplace review, I would ask the presenter to explain one assumption, one alternative and one condition that would change the decision. This makes the reasoning available for discussion without turning the meeting into a test of who used AI.

Consider an illustrative marketing proposal for a new audience. Ask what evidence supports the audience choice, which customer need the message addresses and what result would suggest the idea was wrong. A compelling campaign concept should survive those questions.

The objective is not to make work slower for its own sake. It is to ensure that the people responsible can defend the decisions embedded in the work.

The bigger shift

Training and assessment should be designed together. If a course encourages careful checking but the workplace rewards only rapid output, employees receive conflicting instructions.

A useful exercise can begin with an independent diagnosis before opening an AI tool. Participants then compare suggestions, inspect evidence and revise their proposal. Ask them to document what they accepted, rejected and changed. The deliverable is a reasoned decision, not a long transcript of prompts.

Give reviewers criteria that reflect the actual job. Accuracy, suitability, originality and explainability may matter in different proportions. A routine product description and a new market strategy should not be judged as though they are the same assignment.

Managers also need time to provide feedback. A training budget that covers attendance but omits application and review may leave the most valuable work unfunded. The guide to corporate AI training costs explains what to consider when scoping that investment.

My take

I would treat AI fluency and independent judgement as capabilities to develop together.

In the next team exercise, ask people to bring a brief initial view before using a model. Let them use AI to challenge it, generate alternatives or improve the presentation. Then ask for a final recommendation with the evidence that changed their thinking.

Reward a well-supported disagreement and an honestly identified uncertainty. Otherwise, an organisation may teach employees that the safest answer is whichever sounds most complete.

The value of an AI-assisted proposal lies in the decision it enables. Someone still needs to understand the problem, choose between alternatives and recognise when the assumptions no longer hold. That is the capability a serious training programme should help people practise.