69% of CS educators say AI has killed coding-from-scratch; oral exams are back

The code-writing homework assignment is dying, and computer science professors are the ones pulling the plug. A sweeping new survey from the ACM Task Force on Generative AI and Programming Assessment, which captured responses from 763 educators across 49 countries, reveals a discipline in the middle of a rapid, often chaotic, transformation. The headline number: 69 percent of respondents now believe AI has fundamentally altered the skills required for software development.

That belief is already reshaping the classroom. Sixty-four percent of instructors have moved away from teaching students to write code from a blank file, pivoting hard toward code comprehension, debugging, and system design. “We know that students will likely be developing with AI tools when they enter the workforce,” said Steven Gordon, an Ohio State professor and the report’s lead author. “The key is to produce graduates who are competent in programming and who understand both the capabilities and limitations of AI.” The anxiety driving this shift is palpable—87 percent of educators are spooked by growing student dependency on the technology.

The really disruptive change is in testing. Since tools like ChatGPT can ace a typical intro-level programming assignment, 68 percent of educators have already overhauled how they grade students. The solutions are decidedly old-school: a surge in proctored, in-person exams, paper-based tests, and—most tellingly—the return of the oral exam and the “code defense,” where students must explain their logic face-to-face. It’s a direct countermeasure against AI-enabled shortcuts, but it’s happening without a safety net. A full 48 percent of respondents said they lack proven best practices for weaving AI into their courses, leaving them to figure it out on the fly.

This isn’t just pedagogical pearl-clutching. The educators’ instincts are backed by hard data. A recent Anthropic study found that developers who learned a new Python library with AI assistance scored 17 percent lower on a follow-up knowledge test than peers who learned unaided. Those who fully delegated coding to an AI finished fastest but cratered on the test, averaging just 39 percent. The paradox is stark: the tools that make students look productive in the short term are the very ones leaving them hollowed out on fundamentals, while their professors scramble to build a new, cheat-proof model of education from scratch.