The presentation is flawless. The code is commented. The recommendation seems sound. Then a simple question is enough to make its author hesitate: why choose this solution? With generative AI, a junior employee’s work can appear more mature than their reasoning. The problem is not necessarily cheating: it is the disappearance of the clues that once helped managers guide them. How can you give fair feedback when the quality of the output no longer clearly indicates what someone has understood?
A convincing result, invisible learning
Consider an illustrative scenario: a junior research analyst submits a structured, fluent client briefing complete with recommendations. Her manager finds few flaws. But when he asks which source supports the main conclusion, she struggles to retrace the reasoning. The assistant consolidated the information, suggested the structure and rephrased the arguments. She did the work, without necessarily exercising all the skills the document appears to demonstrate.
This disconnect did not begin with generative models: presentation templates, proofreading tools and help from colleagues have always made production easier. But AI changes the scale. It also supplies transitions, justifications and the appearance of organised thought. Hesitation disappears from the deliverable, even though it often provides valuable clues about what someone needs to learn next.
To examine this issue with September 2026 in view, this article draws on earlier publicly available research and distinguishes its findings from forward-looking suggestions. A study of customer support agents, released in 2023 by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, observed particularly marked productivity gains among less experienced workers using generative assistance. This finding alone does not demonstrate lasting mastery of the skills involved.
Do not confuse assistance with incompetence
Another study published in 2023, conducted with Boston Consulting Group consultants, showed that AI’s benefits varied by task: it improved performance in some areas but could also steer users towards incorrect answers. For managers, the lesson is practical: neither fluent writing nor the use of a tool establishes someone’s level of understanding.
The first pitfall, then, is to praise only the finished product. The opposite pitfall is to devalue all assisted work. One beginner may have thoughtfully framed their request, rejected several suggestions and checked every claim. Another may have accepted an answer uncritically. Their documents look similar; their support needs differ.
Feedback should focus not on how much AI was used, but on the quality of the judgement exercised. What decisions had to be made? Based on what evidence? Which errors can the person recognise? These questions shift the discussion from suspicion to professional development.
Make reasoning visible without staging an interrogation
Requesting complete conversation histories would be cumbersome and sometimes intrusive. It could also encourage junior employees to fabricate an exemplary audit trail rather than flag their difficulties. A better approach is to agree in advance on a few useful records: the objective, an important decision, a check performed and a remaining uncertainty.
For an analytical briefing, these elements can fit in a sidebar. For code, they can appear in the change description and associated tests. For a graphic design, in two annotated variations. The aim is not to recount every click: it is to reveal what the final deliverable conceals.
Three questions that open up the discussion
- “Which choice would you defend if we had to simplify this work?” The question reveals the order of priorities.
- “Which suggestion from the tool did you reject, and why?” It explores critical thinking without assuming that it took place.
- “What would change if this assumption were wrong?” It tests understanding beyond simply recounting information.
These questions should be introduced as a learning routine, not a surprise test. Answers can also be written or prepared: verbal fluency should not become a false indicator of technical competence.
Give feedback on a specific gap
Saying “this is too ChatGPT” helps no one. The phrase mixes a judgement about style, suspicion about the method and criticism of the substance. Constructive feedback, by contrast, distinguishes three levels: what works in the output, where the reasoning remains weak and what to try next.
In our example, the manager could offer this feedback: the briefing makes the options clear; however, the main recommendation rests on a source that is insufficient to support it; the next step is to compare two sources and explain their limitations. The quality of the finished product is recognised, but it no longer obscures the work still needed.
Good feedback leads to a feasible action, not an instruction to “think harder”. Repeating a calculation, constructing a counterexample or checking a reference provides a concrete way forward. It is better to target one crucial learning goal than to catalogue every weakness in a document.
Preserve time for genuine problem-solving
Should beginners be banned from using AI? A blanket ban would also deprive some of useful help with rephrasing instructions or exploring a concept. But routinely delegating the first attempt risks removing the encounter with the problem itself: the one that helps people discover what they do not understand.
One compromise is to divide the activity into stages. First, a brief independent attempt: sketch an outline, propose a hypothesis, identify the unknowns. Then, use AI to challenge or enrich that approach. Finally, revisit the differences and check the decisive points. This process is a teaching approach to adapt, not a universal guarantee of learning.
Certain unassisted exercises can also be reserved for explicit objectives. Understanding a formula, diagnosing an incident or making the case for a recommendation sometimes requires direct practice. The reasons must then be explained and enough time allowed: working more slowly during an exercise is not failure.
The right to learn also depends on management
Junior employees sometimes receive two incompatible messages: work faster with AI, but develop as though you had built everything yourself. If the organisation rewards only speed and polish, asking people to reveal their doubts loses credibility. The right to learn requires time for review and the freedom to say “I don’t know yet”.
Managers can lead by example by showing an AI response they have corrected themselves. Assessment also benefits from distinguishing between results, verification and independence. Progress can then be observed across several assignments, particularly when a junior employee can handle a similar case with less help.
What next? As assistants potentially take on more of the production of deliverables, mentoring will probably need to make human decisions more visible. The priority is not to expose AI users, but to check that they are developing better judgement. A successful document completes one task; well-constructed feedback prepares the way for the next.
