AI Coding Tutors Raise Grades. A New Survey Says They're Lowering Skill.

The pitch for AI coding assistants in the classroom has always been simple: students who use Copilot or ChatGPT to work through assignments finish more of them, and finish them correctly. What a new survey of computing educators makes clear is that “correctly” and “competently” have quietly stopped meaning the same thing.

Why Everyone Believes This

It’s an easy belief to hold because the surface evidence supports it: assignment completion goes up, homework scores go up, and struggling students visibly unblock themselves on problems that used to stall them for hours. When a tool solves the intro-level assignment a student is stuck on, the natural read is that the student learned something in the process — the AI acted as a tutor, not a replacement.

What the Evidence Shows

The ACM Task Force on Generative AI and Programming Assessment surveyed 763 educators across 49 countries between May and October 2025, drawing roughly 500 complete responses. The results describe a split between what grades show and what students retain: assignment and homework scores rise, but developers who used AI assistance scored 17% lower on follow-up knowledge tests measuring the same material without AI help. Sixty-nine percent of educators said AI has already changed what software development skills are required, and 87% named growing student dependency as their primary concern — not occasional misuse, but a default habit of reaching for the tool before reasoning through the problem. Seventy-two percent flagged cheating and plagiarism risk directly.

The Correction

Educators are responding by moving the test itself rather than banning the tool. Sixty-four percent have already shifted teaching emphasis away from writing code from scratch and toward comprehension, debugging, and problem-solving — the skills that don’t disappear when the AI is taken away. Sixty-eight percent have adjusted assessment methods outright: proctored in-person exams were the most common change (56 mentions in the survey), followed by reduced homework weighting (38), oral exams and code-defense sessions (36), paper-based tests (35), and project-based assessment (34). What’s notably absent is confidence that any of this is solved: 48% of educators say they still lack proven best practices for integrating AI into coursework, 28% report insufficient AI expertise themselves, and only 45% work at institutions with any formal AI guidelines at all. The correction is underway, but it’s running well ahead of any consensus on what replaces the old assessment model.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.