ACM AI coding survey finds educators rewriting programming tests

An ACM AI coding survey of computer science educators found that generative AI is already changing how programming is taught and tested. The Association for Computing Machinery task force surveyed 763 educators across 49 countries between May and October 2025, with about 500 nearly complete responses analyzed, and reported that 69% of respondents believe AI has changed the skills required for software development.

The findings matter beyond universities because entry-level software hiring is already being reshaped by tools such as ChatGPT and GitHub Copilot. If introductory programming assignments can be completed by AI systems at roughly average-student performance, educators have less confidence that take-home code proves a student can reason through software problems.

How is AI changing computer science education?

According to the ACM Task Force on Generative AI and Programming Assessment, 64% of surveyed educators said they have changed their teaching. The shift described in open-ended answers was toward code comprehension, debugging and problem-solving, rather than asking students to produce code from scratch as the core signal of competence.

Some instructors are also making AI tools part of the curriculum. The task force said 39 responses specifically mentioned teaching AI use and prompt engineering, while others described classroom demonstrations intended to show both what the tools can do and where they fail.

Steven Gordon, a professor at Ohio State University and lead author of the report, said students are likely to use AI tools in the workforce. He said the goal is to graduate students who can program and who understand AI systems' strengths and limits.

Assessment is changing slightly faster than instruction. The ACM task force found that 68% of respondents have modified how they test students. Common changes included more proctored in-person exams, cited in 56 responses, less grade weight on homework, cited in 38, oral exams and code defense sessions, cited in 36, and paper-based tests, cited in 35. Project-based assessment was cited in 34 responses. Some educators now ask students to disclose AI use or submit logs of their interactions with AI tools.

Institutions are not aligned on AI rules

The survey found no single policy model across schools. Forty-five percent of respondents said their institution has guidelines for AI use, while 39% said it does not. The reported policies range from bans to explicit permission or encouragement, provided students document their use.

The largest blocker cited by educators was not access to tools. Forty-eight percent said they lacked proven best-practice examples for bringing AI into courses. Twenty-eight percent said they did not have enough AI expertise, and 20% said they saw no need to integrate AI into their teaching. For professional development, 74% wanted training on effective teaching methods and 66% wanted help redesigning assessments.

The report has limits. Participants were recruited mainly through ACM and SIGCSE mailing lists and selected national educator directories. The sample leaned toward North America, with 227 respondents, and Europe, with 106. Asia accounted for 57 respondents, South America 10, Oceania nine and Africa three. Among respondents who listed institution type, 77% worked at universities, with K-12 schools and vocational programs only lightly represented. The data is self-reported.

Other research cited alongside the ACM work points in the same direction: AI can raise short-term output while weakening measured learning. An Anthropic study found developers learning a Python library with AI assistance scored 17% lower on a follow-up knowledge test than a control group without AI access. A Chinese longitudinal study of more than 26,000 students found AI use raised homework grades by 18% and cut completion time by 30%, but later closed-book exam scores fell. A UC Berkeley analysis of more than 500,000 grades found top marks rose after ChatGPT's release, especially in courses where writing and programming homework carried heavy weight.

For software employers, the signal is clear enough: coursework and portfolios produced under loose AI rules may say less than they used to. Schools are responding by moving assessment closer to live explanation, code defense and supervised problem-solving.

This story draws on original reporting from The Decoder.