Sales G. Aribe, Louie Jay S. Labastida
14 pages, 4 figures, 5 tables, Published by International Journal on Advanced Science, Engineering and Information Technology (IJASEIT)
Vibe coding (fully conversational AI code generation) cuts task completion time by 27% versus traditional coding and 12% versus AI-assisted coding, but produces lower maintainability and more security vulnerabilities, so its responsible adoption requires human oversight.
Quasi-experimental within-subject comparison of three coding conditions (traditional, AI-assisted with Copilot/Codex, and vibe-coding simulation) across three standardized Python/JavaScript tasks (inventory system, data parsing module, form-based app), with 30 purposively sampled participants (15 professional developers, 15 advanced students) at a Philippine college lab; measures were task time, errors, debugging iterations, SonarQube/CodeQL quality and security metrics, SUS and NASA-TLX surveys, and semi-structured interviews analyzed thematically.
27% faster than traditional coding and 12% faster than AI-assisted coding; SUS 71.4 (good); NASA-TLX 55.5 (moderate workload, reduced syntactic effort but higher linguistic reasoning); lower maintainability index and higher security vulnerability counts than both baselines. Scope: small n (30), three tasks of modest complexity, one lab setting, one unspecified LLM configuration, single country context; nothing about large-scale or production systems.
The authors concede the paradigm is 'transitional,' note trust-calibration and loss-of-control risks, and warn that perceived loss of control was associated with reduced validation of AI outputs. Unstated: no blinding or randomization detail, tiny and purposive sample, three short tasks cannot represent real software projects, 'vibe-coding simulation' is a constructed condition, no learning effects or order-counterbalancing reported in the excerpt, and results may not generalize beyond well-scoped, small applications.
A fair skeptic would note this is a small n=30 study at a single institution with self-constructed tasks and a simulated condition, published in a low-visibility journal; the headline percentage effects could be sensitive to task choice, participant skill mix, and prompt familiarity; no pre-registered protocol, no confidence intervals or significance values shown in the excerpt; the qualitative constructs (trust, loss of control) come from the same small sample and could reflect the authors' framing; and the security-vulnerability claim may hinge on which linter settings were used.
Little or nothing yet for the AI-economy at scale: this is a small educational-lab study, not enterprise evidence. At most it directionally supports the plausible hypothesis that conversational coding speeds small-task delivery while raising downstream maintenance and security costs, which would shift value toward code review and QA rather than reduce total software spend; treat it as a weak signal, not ROI evidence.
Labor & Knowledge Work: Junior developers and educators should note the paper's suggestion that conversational coding shifts skill demand from syntax toward prompt construction, critical validation, and code review, the very competencies that may define entry-level value as AI code generation spreads.
Technical Capability: Tool builders and evaluators get a concrete data point that intent-driven generation trades speed for maintainability and security, so capability claims about coding agents should be scored on downstream quality, not just completion time.