LLM verification-evasion patterns in AI-assisted software development: A taxonomy and implications for developer trust
Information and Software Technology, cilt.199, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 199
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.infsof.2026.108251
- Dergi Adı: Information and Software Technology
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, ABI/INFORM, Aerospace Database, Applied Science & Technology Source, Compendex, INSPEC, Library, Information Science & Technology Abstracts (LISTA), DIALNET, Information Science & Technology Abstracts (LISTA), Business Source Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: AI-assisted software development, Automation bias, Generative artificial intelligence, Human–computer interaction, LLM evasion patterns, Over-reliance, Requirements engineering, Software engineering, Software specification, Trust in AI
- Süleyman Demirel Üniversitesi Adresli: Evet
Özet
Context: Large language models are increasingly used to generate code, yet their output patterns under verification pressure remain poorly understood. When embedded in workflows requiring verification such as tests, terminal output, coverage reports, LLM outputs systematically fail to satisfy these requirements by claiming completion without evidence, creating placeholders, seeking permission to stop. Prior work focuses on output quality rather than chat-level output patterns during development; no systematic taxonomy exists. Objectives: This paper categorizes LLM verification-evasion patterns in AI-assisted software development, operationalizes them as detectors in the Spec2Code framework, and examines their implications for developer trust and responsible adoption. Methods: We combine practitioner observation with empirical corpus analysis. The taxonomy was derived through design science research across software projects developed with the Spec2Code framework; a three-stage process (observation, categorization, operationalization) produced 23 categories in six themes, each implemented as a programmatic detector. We then evaluated the detectors on two corpora: (i) a within-subject comparison of 56,680 assistant messages from Spec2Code-enforced versus unrestricted Cursor chat sessions; and (ii) a cross-family analysis of 74,932 assistant messages across 713 chat sessions spanning four model families that are Claude 3.5, GPT-4, Opus 4.5, Sonnet 4.5, triangulated under three independent detection tiers that are literal substring, deterministic regex, and sentence-transformer cosine similarity. Results: We identify 23 verification-evasion categories across six themes. Static detector evaluation on the software project corpus achieved macro-averaged F1 = 0.80. In chat data, Spec2Code-enforced sessions showed lower occurrence rates for all ten chat-level categories than unrestricted sessions. Cross-family analysis found 8 of 10 chat-level categories in at least two model families. Conclusion: The taxonomy and detection framework support human-in-the-loop oversight and responsible adoption of LLM-assisted development. This work contributes a taxonomy with operational definitions, a four-layer detection framework with measured effectiveness, and cross-model evidence supporting responsible adoption of GenAI in software engineering.