← Back to blog
l'IA et les développeur June 29, 2026

AI and software development: what is actually changing about the job

Studies find different effects across tasks and teams. Here is what the evidence can reasonably tell us about developers’ work.

AI tools write code, explain existing systems, suggest tests, and sometimes execute a task from start to finish. The work is already changing. Claiming that developers will soon disappear, however, goes further than the available evidence allows.

The most useful studies do not answer one universal question. They examine different tasks, populations, and tools. Their findings need to stay attached to that context.

Real gains in specific settings

A controlled experiment published in 2023 asked participants to implement a JavaScript HTTP server. The group with access to GitHub Copilot completed that task faster. This was a bounded task in an experimental setting; it did not measure years of product maintenance or responsibility for a production service (Peng et al., 2023).

A 2025 field study combined three randomized trials at Microsoft, Accenture, and a third company, covering 4,867 developers. It found an average 26.08% increase in completed tasks, with substantial uncertainty and different effects across organizations (Cui et al., 2025). This shows measurable potential in those environments, not a universal rate.

Context can reverse the result

METR studied 16 experienced developers completing 246 tasks from open-source repositories they knew well. With early-2025 tools, tasks took 19% longer when AI was allowed (Becker et al., 2025). The authors explicitly restrict the conclusion to that setting.

In February 2026, METR released newer data but judged its central estimate unreliable. Developers who valued AI most were less willing to accept no-AI tasks, and participants selected tasks differently. The team changed its experimental design instead of promoting a fragile number (METR methodology update, 2026). The tools are moving quickly, and so is the challenge of measuring them.

The bottleneck moves

Generating more code does not guarantee a better product. Teams still need to understand the need, shape the system, protect data, review changes, test behavior, observe production, and make trade-offs. When generation becomes faster, review, integration, and user feedback may become the new constraints.

The 2025 DORA report describes AI as an amplifier of an organization’s strengths and weaknesses. It directs attention to the whole work system rather than adoption of one tool (DORA 2025, 2025). A team with small batches, useful tests, reliable delivery, and clear ownership can absorb generated code more safely. A team already blocked by dependencies and approvals may simply add more volume.

Skills that gain value

Developers remain responsible for decisions a model does not know: business context, operational constraints, acceptable debt, security, the cost of change, and consequences for users. Three abilities become especially valuable:

  • framing the problem with a testable goal and relevant constraints;
  • checking the result by reading the code, challenging assumptions, and asking for evidence;
  • connecting the layers across interface, API, data, infrastructure, and operations.

AI can accelerate exploration, documentation, and repetitive changes. It can also generate a plausible solution that misses an unwritten rule. System knowledge and fast feedback remain decisive.

Measure before generalizing

A team can select several categories of work and compare a baseline period with an assisted period. Examine lead time to production, review effort, rework, incidents, developer satisfaction, and delivered value. DORA similarly recommends keeping several dimensions of measurement instead of reducing productivity to code volume or suggestion acceptance (Measurement frameworks, 2025).

Software development was never only typing code. AI makes that easier to see: the durable value lies in turning an uncertain need into a useful, verifiable, operable system.

References

Share this article

Read also