Over the last two years the QA profession has changed more than in the previous ten. Neural networks are no longer a "promising technology" from conference talks — they are everyday tools: they draft test cases, triage failed runs and repair broken locators. Let's look at what actually works in 2026.
LLMs as a working QA tool
Generating test cases from requirements, drafting autotests, analysing flaky tests — all of this is routine now. The important shift: teams have stopped "just asking a chatbot" and are building pipelines with proper test design techniques and mandatory human review. An LLM speeds up the routine, but the engineer still owns the quality.
Self-healing tests
Self-healing is no longer a marketing word: tools have learned to survive layout changes by picking new locators from context. This cuts the maintenance cost of UI tests — historically the most expensive part of automation.
Testing AI systems
The flip side: now we have to test the neural networks themselves. Non-deterministic output, hallucinations, prompt injection — this calls for new practices: golden datasets, LLM-as-judge with manual sampling, regression runs on every model or prompt change. It is one of the most in-demand new QA competencies.
What stays the same
- Shift-left and DevOps: continuous testing in the pipeline is basic hygiene, not a trend.
- Security: SAST/DAST in CI and attention to the dependency supply chain.
- Engineering fundamentals: test types, test design and the ability to read code have not been cancelled by AI — quite the opposite, you cannot verify its work without them.
My conclusion is simple: AI does not replace a QA engineer, but a QA engineer with AI is noticeably faster than one working the old way.