When we talk about AI in software development, we often talk about speed. Writing code faster, finding errors, and building a working application are now part of daily work. But there is another question we need to think about: How will junior employees gain experience?
Small features, simple bug fixes, and reading existing code were important learning opportunities in the first years of a career. Finding the cause of an error, trying different solutions, and getting feedback from an experienced colleague helped people develop their technical judgment over time.
As we give more of these tasks to AI tools, can we expect the same learning opportunities to continue on their own?
Does a working result show that learning has happened?
Building a working application with AI support does not always mean that a student has learned the ideas behind it. The student may have completed the task. However, they may not yet know why the solution works or what to do when an error happens.
For this reason, we need to look at both the work interns deliver and how they explain it. Talking about their assumptions, the parts they found difficult, and how they would change the solution for a new requirement helps us see what they have learned.
What do we aim to achieve by including interns in projects?
Over the last six months, we have included students from Computer Engineering, Software Engineering, and Management Information Systems (MIS) in our projects. We want them to learn how to use AI tools and develop their professional skills. We also want to help them close the gap between education and working life.
A senior team member guides and reviews their work. Students can discuss their ideas, ask questions, and receive feedback while working on their tasks.
To reduce the risk of data leaks, students work on standalone solutions that do not use sensitive data. These solutions run separately from our organisation's systems. This gives students a space with clear limits where they can gain real project experience.
This experience includes more than technical work. Understanding the need, noticing missing information, asking the right questions, and explaining decisions are also important. For each task, we need to think about both the expected result and the skill the student will develop.
For example, connecting an application to an API may be the project goal. Explaining how user access is checked, planning for errors, and discussing how to avoid processing the same data twice may be learning goals.
How can we turn AI support into a learning opportunity?
One useful approach is to ask students to prepare their own first solution. They explain how they understand the problem, which method they want to use, and where they need help. Then they use AI support.
The aim is not to leave students without help for a long time. It is to give them space to develop their own ideas and then provide the right support.
Sinha and Kapur reviewed several studies. Their findings show that trying to solve a problem before receiving instruction can help learning in the conditions they studied. The results depend on the situation, but they support planning both the first attempt and the feedback that follows.[1]
The type of help AI gives also matters. Instead of asking for the full solution, students can ask for an explanation, a hint, or feedback on a wrong assumption in their attempt. This can help them stay active in the learning process.
In a study on mathematics learning, Bastani and colleagues tested AI support that used teacher-written hints and avoided giving full answers. This approach greatly reduced the learning loss seen with standard AI use. We cannot assume the same results will happen in software teams, but the study shows why the design of support matters.[2]
These studies give us a starting point for improving how we work. A process where students first try, then receive support, and finally review their solution can create more space for learning in projects.
Giving students space to explain their work
Another step is to ask students to explain the work they have completed. Along with asking, “Does the code work?”, we need to ask questions such as:
- Why did you choose this solution?
- When could it fail?
- What other approach could you use?
- Which part would you change if there was a new requirement?
We want students to make their own connections between ideas, rather than repeat an explanation from AI.
A research review by Bisra and colleagues also supports asking learners to explain the reasons behind a solution in their own words.[3]
A few days later, it may help to give the student a similar task with some changes. Seeing how they use their knowledge in a different situation helps us understand their progress better than asking them to repeat the first solution.
The senior team member has an important role here. Feedback gives us a chance to understand how the student thinks and discuss the parts they still need to learn.
Following skill development as well as speed
Software teams face real pressure to deliver. When learning time is not part of the plan, it is easy to choose options that make the work faster in the short term.
As team leaders, I believe we need to follow two results together: the quality of the work and the development of the person doing it.
For the work, we can look at time, errors, and how much needs to be done again. For skills, we can look at whether the person can explain the solution, find the cause of an error, and use what they have learned in a different situation.
One approach cannot solve every concern about the future of junior employees. But in our own projects, we can create opportunities for young people to take responsibility, receive feedback, and develop their thinking skills.
This is what we aim to do by including interns in projects. We want to help future colleagues learn to build with AI, understand the results, and gradually take responsibility for their own decisions.
References
- Sinha, T., & Kapur, M. (2021). When Problem Solving Followed by Instruction Works: Evidence for Productive Failure. Review of Educational Research, 91(5), 761-798. https://doi.org/10.3102/00346543211019105
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
- Bisra, K., Liu, Q., Nesbit, J. C., Salimi, F., & Winne, P. H. (2018). Inducing Self-Explanation: A Meta-Analysis. Educational Psychology Review, 30, 703-725. https://doi.org/10.1007/s10648-018-9434-x
Cover photo by Alicia Christin Gerald on Unsplash.
Top comments (0)