Learning with AI while preserving judgment · Software Engineering in the AI Era

13.1Production and learning have different goals

Receiving a solution ≠ understanding it

In a 2026 experiment, 52 mostly junior engineers learned the Trio library. On the post-test, the AI group averaged 50% and the group without AI 67% — a 17-point gap. The small time difference was not statistically significant. [F15]

50%

With AI assistance

67%

Without AI

Post-test score

−17 pp

13.2A learning cycle that uses AI

Seven steps from gap to asset

A proposal of this course: AI is your laboratory, not your answer key. Step through the cycle.

Gap → understanding → asset

Step 1 / 7

Choose a concrete gap

Something that affected a recent decision: isolation, cancellation, caching, authorization, accessibility, profiling.

13.3A sustainable agenda

Three levels of study

The first two levels decide which tools deserve attention. If your biggest limit is understanding concurrency, switching editors won’t close the gap.

Current tools

Features that let you carry out the practices better.

Operational practices

Specification, context, delegation, verification and flow.

Durable fundamentals

Reasoning about systems, data, algorithms, security and product.

13.4Training people who are starting out

Make judgment visible — on purpose

Include AI from the start, with tasks that expose the learner’s reasoning. If mechanical tasks are automated, training must create experiences of judgment deliberately.

Progression happens through demonstrated responsibility — not through the volume of code produced.

Key takeaways

1
Design AI use by goal: producing familiar work and learning a gap need different interactions.
2
Use the seven-step cycle: predict, experiment, rebuild, vary, capture.
3
Fundamentals and practices decide which tools deserve your attention.
4
For juniors, create experiences of judgment deliberately; evaluate what they understand and control.

Read the full section of the original reportThe complete, unabridged text — with every source and nuance.

13.1. Production and learning have different goals

When producing familiar work, heavy delegation may be the right choice. When a gap compromises your ability to evaluate, the activity needs to include deliberate learning.

A 2026 experiment with 52 engineers, mostly junior, examined learning of the Trio library. The average score on the post-test was 50% in the AI group and 67% in the non-AI group: a difference of 17 percentage points. The small difference in time was not statistically significant. The study measures immediate learning on a limited task. The analysis of usage modes suggests value in interactions aimed at understanding, but those modes were not separately randomized. F15

The practical guidance is to design AI use according to the goal. Receiving a solution and understanding a solution are different outcomes that need their own criteria.

13.2. A learning cycle that makes use of AI

The method below is a proposal of this report:

Choose a concrete gap. Use something that affected a recent decision: isolation, cancellation, caching, authorization, accessibility, or profiling.
Write down a prediction. Before consulting the answer, describe the expected behavior and the reasons.
Use AI as a laboratory. Ask for small examples, variations, counterexamples, and explanations connected to your system.
Run and observe. Compare the prediction with concrete results.
Rebuild the explanation. Use your own words and make the conditions of validity explicit.
Solve a variant. Change inputs, scale, or type of failure to test the transfer of knowledge.
Produce a useful asset. Turn the discovery into a test, decision note, or procedure that helps with the work.

You can ask for gradual hints instead of a complete fix. You can also ask the model to critique an explanation you wrote. What matters is that there be some demonstration of understanding that does not depend on repeating the answer you just received.

13.3. A sustainable study agenda

Organize study into three levels:

Durable fundamentals: reasoning about systems, data, algorithms, security, and product.
Operational practices: specification, context, delegation, verification, and flow.
Current tools: features that let you carry out those practices better.

The first two levels guide which tools deserve attention. If your biggest limitation is understanding concurrency, switching editors probably won't close the gap. If you understand the domain but manually repeat every configuration, improving the environment may be the most useful investment.

One or two short sessions of deliberate learning per week can be a starting point, adjusted to availability. This cadence is a suggestion, not a scientifically optimal frequency. The indicator of progress is being able to explain, predict, detect, and correct something that previously depended on trusting the model's answer.

13.4. How to train people who are starting out

Early career development should include AI from the start, with tasks that make the learner's reasoning visible. A small, complete project makes it possible to observe requirement, implementation, test, deployment, and effect, preserving the relationship between the parts.

The mentor can ask the person to formulate the contract before generation, anticipate a failure case, explain a decision, and diagnose a variant. AI can help in all of these steps, but the evaluation should verify what the engineer has come to understand and control.

If mechanical tasks are automated, training needs to create experiences of judgment on purpose. This includes analyzing incidents, testing hypotheses, talking with users, and taking part in decisions of appropriate scope. Progression happens through demonstrated responsibility, not through the volume of code produced.