Abstract
Modified Condition/Decision Coverage (MC/DC) testing is often considered an essential criterion for unit testing safety-critical software. However, many developers find it challenging to create accurate MC/DC test suites. In safety-critical software, logical errors can cause life-impacting failure, so strict testing is essential. Modern software developers can use generative artificial intelligence (GenAI) to produce MC/DC test suites automatically. However, reliance on AI-generated tests can be problematic, as large language models struggle with unconventional code constructs, potentially resulting in incorrect MC/DC coverage. The issue of inadequate coverage may be further exacerbated by the human tendency to equate GenAI's confidence with correctness, leading to the false impression that generated test suites are satisfactory. To quantify these risks, this paper first analyzes naïvely generated MC/DC tests for common logical expressions and explains deficiencies with the resulting GenAI created tests. This research continues by examining GenAI-generated tests on more complex decision algorithms, highlighting the risks of auto-generated test suites. The paper finally presents a workflow using appropriate safeguards, traditional analysis tools, and GenAI to produce complete and useful test suites. The findings reveal that, with proper human supervision, GenAI can be used reliably for MC/DC testing.