When Modernization Breaks Working Code: A Study of LLMs on Legacy PHP
13 points by tambourine_man
Pasting legacy PHP code into a large language model (LLM) for modernization may not yield better results than keeping the original code, according to a study evaluating five LLMs on 100 WordPress 4.0 files. The modernization attempts were assessed using 630 hidden behavioral checks and 511 Rector-detected findings, revealing that the outputs often failed to preserve the original code’s functionality. The study suggests caution when using LLMs for code modernization without project context or execution feedback.