Codex novices: code-author, then edit it without AI

Original: Majeed Kazemitabaar et al., Studying the effect of AI Code Generators on Supporting Novice Learners in Introductory Programming, CHI 2023. Ninety 10–17-year-olds new to text-based programming began; after a Scratch baseline, one member of each pretest-matched pair was randomly assigned access to OpenAI Codex for Python code-authoring during seven 90-minute practice sessions. Each of 45 authoring exercises was followed by a manual, unaided modification of that code. All had documentation, instructor feedback and increasing difficulty. Sixty-nine completed all ten sessions (33 AI, 36 control); 21/90 attrition is not broken down by assignment in the article. An immediate post-test one day later and a retention test one week later denied both AI and documentation.

Results: The Codex group made more progress on authoring (reported 1.15× completion) and scored 1.8× as high for correctness during AI-allowed authoring, with less time and fewer errors. In immediately following no-AI code-modification tasks, group mean scores were 66.2% versus 58.4% (p=.202): no statistically established deficit, but not proof of equivalence. Both evaluation sessions showed no significant overall gap; at one week the AI group’s means were slightly higher, not significant. A higher-prior-Scratch-knowledge subgroup performed better at retention with Codex, but treat subgroup cautiously. The paper’s mechanism—generate/use/modify offers repeated reading and tracing—remains interpretation, not separately randomized.

Why this matters: This older and very different cohort is a direct counterexample to universalizing the Trio immediate unaided quiz as AI-driven inevitable deskilling. Yet its paired author→modify curriculum, age, basic syntax and 2022-era Codex do not establish how to train adults to supervise production agents. It also uses no-AI follow-up tests, so still misses Dru’s AI-allowed workplace-value endpoint. Original includes instructor help but not a comparative labor-cost ledger.