Artificial Intelligence in Software Testing and Beyond: A Review of Current Practices and Emerging Challenges
Codrina-Victoria Lisaru, Claudiu Vasile Kifor · Acta Informatica Pragensia · 2026
Background: Artificial intelligence (AI) is increasingly used both to test software (T1) and to assure AIbased systems (T2), with adjacent software-engineering work that shapes testing practice (T3).Prior reviews are mostly descriptive and rarely report comparable maturity or replicability signals.Objective: To provide a PRISMA-style systematic review (2015-2025, Web of Science) that maps T1-T2-T3 within a testing-centric frame, audits evidence maturity, threats reporting, and artefact openness per paper, and adds an explicit lens of large language models or generative AI (LLMs/GenAI).Methods: We queried the Web of Science Core Collection (2015-2025), screened via a predefined protocol, and extracted ten items (D1-D10) per study to normalize comparisons.Seventy-two papers met the criteria.Findings are organized into three themes: (T1) AI-based software testing, (T2) testing/validation of AI systems, and (T3) AI-related software engineering topics with implications for testing-T3 corresponding to the "beyond" in the paper's title. Results:The corpus is limited in practice-oriented evidence: 31 laboratory/simulation, 3 industrial, 10 hybrid, 6 conceptual/guideline and 22 secondary studies.Only 18/72 provide public artefacts; 33/72 report no empirical metrics.By theme, T1=32, T2=15, T3=25; the LLMs/GenAI subset totals 10 papers.Openness strongly co-occurs with measurable outcomes (88.9% of artefact-sharing papers report metrics vs 42.6% without), yet "all-three credible" studies (industrial/hybrid + open artefacts + metrics) are rare (4/72 overall; 1/10 for LLMs/GenAI).Conclusion: AI shows promise for testing, but evidence remains thin on industrial adoption and reproducibility.We recommend prioritizing hybrid/industrial validations, releasing artefacts by default, and using standardized task-metric bundles.The review presents T1 and T2 results, separates T3 for scope clarity, and provides actionable maturity and replicability signals to guide responsible, empirical adoption.