AI coding tools have created a difficult tradeoff for computer science education. They can help students produce working code, but they also make it harder for educators to know whether students actually understand what they are submitting.
A survey conducted by the ACM Task Force on Generative AI and Programming Assessment shows how quickly instructors are adapting. The task force surveyed 763 educators from 49 countries between May and October 2025, and about 500 nearly complete questionnaires were analyzed.
AI is changing what educators think students need to learn
The central finding is not only that students are using tools such as ChatGPT and GitHub Copilot. It is that educators increasingly believe the target skill set for software development is changing.
According to the survey, 69 percent of respondents believe AI has changed the skills needed for software development. That shift helps explain why many instructors are reconsidering the traditional programming assignment, especially in introductory courses where AI coding tools can solve typical tasks and often perform at the level of an average student.
The concern is not limited to academic misconduct. The top concern among educators was growing student dependency on technology, cited by 87 percent of respondents. Cheating and plagiarism followed at 72 percent.
Those two worries are connected but not identical. A student may follow the rules and still become too reliant on a tool to reason through code independently. That is why the education response is moving beyond detection and punishment toward a deeper redesign of teaching and assessment.
Teaching is moving toward understanding code
Sixty-four percent of respondents said they have already changed how they teach. The direction of that change is clear: less emphasis on writing code from scratch, and more emphasis on code comprehension, debugging, and problem-solving.
This does not mean programming fundamentals are disappearing. Instead, educators appear to be asking students to demonstrate understanding in ways that are harder to outsource to an AI system. Reading code, explaining behavior, finding errors, and choosing an approach all require judgment that a finished submission may not reveal.
The survey also found that 39 responses mentioned teaching AI use and prompt engineering as explicit topics. Some educators demonstrate AI tools in class, using them to show both what they can do and where their limits are.
"We know that students will likely be developing with AI tools when they enter the workforce," said Steven Gordon, a professor at Ohio State University and lead author of the report. "The key is to produce graduates who are competent in programming and who understand both the capabilities and limitations of AI."
That framing matters. If AI tools are part of the software development environment students will encounter, banning discussion of them may leave graduates underprepared. But allowing unrestricted use without changing instruction can make it difficult to tell whether students are learning durable skills.
Assessment is changing even faster
The survey found even more movement in testing. Sixty-eight percent of respondents said they have adjusted how they assess students.
The most common changes include more proctored in-person exams, reduced weight for homework, oral exams, code defense sessions, paper-based tests, and project-based assessments. In the survey responses, proctored in-person exams were mentioned 56 times, reduced homework weight 38 times, oral exams and code defense sessions 36 times, paper-based tests 35 times, and project-based assessments 34 times.
These assessment formats try to measure process, explanation, and independent reasoning. They also reduce reliance on unproctored homework as the main signal of competence.
- Oral exams can require students to explain choices and answer follow-up questions.
- Code defense sessions can test whether students understand the code they submitted.
- Project-based work can evaluate broader problem-solving and design decisions.
- Paper-based tests can focus attention on reasoning without immediate tool access.
Some educators now require students to disclose AI use and submit logs of their interactions with AI tools. That approach treats AI use as something to be documented and evaluated, rather than simply hidden or assumed.
Policies remain uneven
Institutional rules have not settled into a common model. Forty-five percent of respondents said their institution has guidelines on AI use, while 39 percent said it does not.
Where policies exist, they vary widely. Some institutions use outright bans. Others explicitly permit or encourage AI use, as long as students document how they used it.
This patchwork leaves instructors with a practical problem. A single course policy may need to account for institutional expectations, student behavior, academic integrity rules, and the reality that AI tools can now complete many conventional assignments.
The biggest barrier reported in the survey was a lack of proven best practices, cited by 48 percent of respondents. Twenty-eight percent said they do not have enough expertise with AI, while 20 percent said they see no need to integrate it into their teaching at all.
Professional development needs are also clear. Seventy-four percent want training on effective teaching methods, and 66 percent want help redesigning assessments.
The evidence points to a learning gap
The survey’s concerns are reinforced by several recent studies mentioned in the source article. An Anthropic study found that software developers who learned a new Python library with AI assistance scored 17 percent lower on a follow-up knowledge test than a control group without AI access. A small subset that fully delegated coding to AI finished tasks the fastest but averaged just 39 percent on the knowledge test.
A Chinese longitudinal study tracking more than 26,000 students found that AI use boosted homework grades by 18 percent and cut completion time by 30 percent. But on closed-book exams after six months, scores dropped by 20 percent. On later entrance exams, the decline was 18 and 24 percent, depending on the test.
A UC Berkeley analysis covering more than 500,000 grades found a sharp rise in top marks after ChatGPT's release, especially in courses heavy on writing and programming assignments. The increase was strongest in courses where unproctored homework carried significant weight in the final grade.
That evidence helps explain why computer science educators are changing what they reward. If AI can raise homework performance while weakening later test performance, then grades alone may no longer show whether students are building lasting ability.
The ACM survey has limits. Participants were recruited mainly through ACM and SIGCSE mailing lists and selected national educator directories. Respondents were concentrated in North America and Europe, and 77 percent of those who listed institution type work at universities. K-12 schools and vocational programs were barely represented, and the findings are based on self-reported data.
Even with those caveats, the direction is hard to miss. AI coding tools are not just changing how students complete assignments. They are forcing educators to define what real programming competence looks like when code generation is always within reach.