Computer science teachers are quietly rewriting their exams because of AI

If you have wondered how schools are coping with students who can hand a homework problem to ChatGPT, here is a broad answer. An expert group set up by the ACM, the international computing society, surveyed 763 computer science educators in 49 countries and found that most have already changed both what they teach and how they grade it.

The numbers are clear. Sixty-nine percent of respondents believe AI has changed which skills software developers actually need. Sixty-four percent have already shifted their teaching away from writing code from scratch and toward reading, debugging, and explaining it. Assessment is moving even faster: 68 percent have changed how they test. The most common shifts were more supervised in-person exams, less weight on homework, oral exams and code defense sessions where a student has to explain their own submission out loud, and project-based work. Some teachers now ask students to disclose their AI use and hand in logs of their chats.

Here is what sits underneath all of that. AI coding assistants can now solve a typical introductory programming assignment at roughly the level of an average student, which quietly broke the deal that homework used to represent. If an exercise no longer proves that a student can do the thing, the exercise stops being an assessment and becomes busywork. So teachers are moving grading to settings where a person is in the room. The top worry among respondents was not cheating, though 72 percent named it, but dependency: 87 percent said they are concerned students lean on the tools too hard, too early.

That concern has evidence behind it. An Anthropic study found developers who learned a new Python library with AI help scored 17 percent lower on a follow-up knowledge test than a group without it. A longitudinal study of more than 26,000 Chinese students found AI use lifted homework grades by 18 percent, then lowered closed-book exam scores by around 20 percent six months later. A fair caveat on the ACM survey itself: respondents skewed heavily toward North America and Europe, 77 percent teach at universities, and all answers are self-reported. The biggest barrier named, by 48 percent, was simply the absence of proven examples for how to teach with AI well.

What this means for you: If you are learning to code or picking up any technical skill with AI help, the useful lesson is not “avoid the tools.” It is that speed and understanding are different things, and the tools give you the first for free. A practical habit: let the AI produce the solution, then close it and rebuild the tricky part yourself, or explain it aloud as if defending it. If you hire or train people, expect a graduate’s portfolio to say less than a fifteen-minute conversation about their own code.

Sources

• ACM: Task force report on generative AI and programming assessment
• ACM Task Force: full report (PDF)
• The Decoder: The AI coding tutor paradox grows