Combinatorial testing methods such as pairwise testing and MCDC (modified condition decision coverage) combine high testing effectiveness with a manageable number of test cases. Considering this attractive combination for testers and managers, it is surprising that these test design methods are used relatively rarely in practice.
This is probably because they are difficult or impossible to use without tool support when multiple parameters or conditions are involved.
Are test design procedures a use case for GenAI?
Until now, these test design techniques had been a seemingly difficult problem to solve for generative artificial intelligence (GenAI). This is because the test cases are not only derived from natural language requirements, but are then calculated. And the mathematical abilities of LLMs (large language models) have generated a lot of cheap laughs in blogs and social media over the past two years. As good as LLMs are at working with text, slightly more complex calculations lead these models into hallucination faster than I can add 2 + 2.
A completely normal training course to become a GenAI assisted test engineer
You read that right – I wrote the previous paragraphs in the past tense because I experienced a big surprise during my GenAI Assisted Test Engineer training in Hamburg in June 2025.
Surprises are not uncommon in this training, as the capabilities of LLMs are developing rapidly at the moment, but this surprise was a game changer.
In the experiments on test design methods, we included pairwise testing, knowing full well that it doesn't work well. First, I show that the results in normal chat are rather poor (mostly I get a table with all possible combinations) and then I show a custom GPT that acts as an interface for a small Python script that calculates the pair combinations correctly.
After a short warm-up (this pattern is also taught in the course), in which I had pairwise testing explained to me, I gave the instruction to apply this technique to my example and received an almost perfect result. Without any extensive prompt engineering or other tricks.
Surprise! - In addition to pairwise testing, MCDC also works with GenAI.
This made me overconfident, and I tried the same procedure for MCDC. And here too: instead of full multiple condition coverage, which is essentially the definition of test case explosion, I got the correct number of test cases (N+1). A quick check to see if they were the right ones – that too.
We were fascinated and continued experimenting, achieving mostly correct results for both techniques.
The LLMs have learned something new again!
Stay careful!
Even though our experiments in training were promising, we also discovered a few errors. This is due to the non-deterministic nature of the current models and cannot be completely avoided.
The safest strategy here is to be vigilant, check results and take corrective action in the event of errors.
In this way, generative artificial intelligence can become a powerful tool in the tester's toolbox.
Would you like to apply MCDC and pairwise testing with GenAI?
Then join our ISTQB course on testing with Generative AI!
FAQ: Questions about using GenAI for test case design
What is pairwise testing – and why is it rarely used in practice?
Pairwise testing is a combinatorial test design method that combines high test effectiveness with comparatively few test cases. In practice, however, it is rarely used because without tool support, it is difficult or impossible to generate the correct pair combinations cleanly when there are multiple parameters.
What is MCDC (Modified Condition/Decision Coverage) – and how does it work?
MCDC (Modified Condition Decision Coverage) is a test design technique that is very effective but limits the number of test cases compared to full multiple condition coverage. The text emphasises that instead of test case explosion, you get the correct, reduced number (in the example, N+1 test cases).
Why are pairwise testing and MCDC so difficult to implement without a tool?
Because the test cases here are not only ‘derived’ from requirements, but must also be calculated combinatorially for multiple parameters or conditions. Without tool support, this quickly becomes confusing and is practically impossible to carry out cleanly. According to the text, this is precisely one of the main reasons why these methods are rarely used despite their benefits.
Is Pairwise/MCDC a useful application for GenAI?
Yes, this is precisely the question posed by the text – and it provides clear evidence from the training: the combination of GenAI and tooling can be a real lever here. During the training (June 2025, Hamburg), a short ‘warm-up’ dialogue resulted in the Pairwise model being applied almost perfectly to an example without extensive prompt engineering.
How do I deal with errors and non-determinism in GenAI in test design?
The text explicitly warns: Even though the results are promising, isolated errors occurred – due to the non-deterministic nature of current models. The recommended strategy is therefore to remain vigilant, check results and take targeted corrective action in the event of errors. This will make GenAI a powerful tool without you having to accept incorrect test cases without checking them.