CRA urges computing programs to rethink assessment as generative AI changes student work

A new CRA-E white paper calls on computer science departments to move beyond product-only assessment, connect AI policies to learning goals and protect the practice students need to build expertise

The Computing Research Association has issued new guidance for university computing programs adapting to generative AI

The Computing Research Association’s Education Committee is calling on universities to rethink how computer science students are taught and assessed as generative AI makes it increasingly easy to produce working code, polished reports and other finished work.

Its new white paper, AI in Computing Education: Principles and Practical Guidance for Program Leaders, is aimed at department chairs, curriculum committees and program directors responsible for adapting computing degrees to generative AI.

Rather than setting rules around particular models or tools, CRA-E sets out four principles for programs: treat learning as a process rather than a finished product, connect AI policies directly to learning goals, redesign assessment to reveal what students actually understand, and build collaboration beyond individual computing departments.

That deliberately avoids guidance tied to the current generation of AI products.

“The tools change faster than a curriculum committee can meet,” says Jonathan Bell, Associate Professor in the Khoury College of Computer Sciences at Northeastern University and one of the paper’s six authors.

“If we had written recommendations for the models that existed when we started drafting, some of them would already be wrong.”

The paper is also explicit about what it does not attempt to solve. Financial and environmental costs, FERPA compliance and equity implications are acknowledged but are not its primary focus, while questions including what should be removed from existing curricula as AI capabilities grow are left for further work.

A working program is not proof that a student learned to build it

At the heart of CRA-E’s guidance is a distinction that generative AI has made harder for educators to ignore: producing the right answer and developing the skills required to reach it are not the same thing.

The authors argue that students still need to acquire foundational skills, practice integrating them and receive targeted feedback, even when an AI system can already complete the task successfully.

That problem may be particularly acute in introductory computing courses, where small, clearly defined programming problems have traditionally been used to build mastery.

Borja Sotomayor, Senior Instructional Professor of Computer Science and Director of the Masters Program in Computer Science at the University of Chicago, gives the example of asking students to build a chat server.

“When I ask students to write a chat server in my Networks class, it is not because the world needs more chat servers: it is because the process of building that chat server will help them to better understand network sockets, how to read a protocol specification, and how to debug network traces,” he says.

A student can now potentially produce the required server without going through that process.

“What GenAI changes is that a student can now submit a working chat server without actually experiencing that struggle: without, in other words, experiencing the process of learning.”

CRA-E is not arguing that AI should therefore be removed from computing education. Later in a student’s course, the paper suggests activities such as evaluating, debugging, critiquing and extending AI-generated code can themselves become part of the learning process.

The question for programs is when AI supports the skill being developed and when it bypasses it.

Assessment needs to show more than the finished code

CRA-E argues that when generative AI can produce working code, correct proofs, functional prototypes and polished reports, assessing only the final submission becomes a weaker measure of student learning.

Its suggested alternatives are decidedly practical. Programs are encouraged to use approaches including oral examinations, code walkthroughs, live problem-solving, iterative projects with checkpoints, code review and critique, and proctored assessments.

The paper points to computer-based testing as one option. It cites a University of Illinois facility that has proctored more than 90,000 exams annually, using randomized, auto-graded assessments that students complete in a supervised environment.

It also suggests a lower-overhead version for institutions without dedicated testing infrastructure, with students completing shorter assessments on laptops in monitored rooms.

AI does not disappear from this model either. CRA-E suggests generative AI could help create varied problems, provide preliminary feedback and support instructors preparing targeted questions for oral assessments.

The emphasis is on making students’ reasoning visible. Chelsea Troy, Senior Staff Software Engineer, Data Platform and Machine Learning Operations at Mozilla and a co-author of the paper, argues that traditional submissions have always been an indirect measure of ability.

“Most submission artifacts proxy students’ skill rather than measuring it directly,” she says.

“Students have always had options for juicing the artifact at the expense of the skill, but AI wedges itself between metric and proxy much more conveniently than many of its predecessors.”

The white paper also warns that students may overestimate their own abilities after using generative AI to complete a task. Asking them to explain code, solve a problem live or defend their reasoning can expose the difference between producing a successful output and understanding how it works.

CRA-E wants AI rules tied to the purpose of an assignment

The same thinking applies to AI policies. Instead of a blanket instruction telling students not to use generative AI, CRA-E recommends that educators explain why AI is restricted for a particular task and connect that restriction to the skill students are meant to develop.

A debugging assignment, for example, might prohibit AI because independently finding and fixing errors is itself the learning objective, and a skill students will later need when evaluating AI-generated code.

Individual courses can still adopt different approaches. The recommendation is not for identical AI rules across an institution, but for consistency in how those decisions are explained.

Programs are also advised to use existing faculty governance structures rather than leaving individual instructors to construct AI policy in isolation.

That includes making room for objections from students or staff concerned about privacy, data collection, ethics or environmental impact, and providing alternatives or accommodations where appropriate.

At institutional level, the authors want computing departments to look beyond their own classrooms too. The paper calls for collaboration with other disciplines on questions such as AI literacy, appropriate student use and professional expectations, as well as greater engagement with industry and AI-in-education researchers.

The Computing Research Association represents more than 250 North American organizations involved in computing research, including academic computer science and computer engineering departments, industry and government laboratories, research centers and professional societies.

For computing programs considering where to begin, the recommendations are concrete: map which courses build foundational skills and which emphasize professional practice, decide where AI should be restricted, structured or encouraged, audit assessments to check whether they still measure the intended learning, establish a program-wide framework for AI policies and give faculty opportunities to build firsthand experience with generative AI tools.

Bell argues that institutions do not need to wait for the technology to settle before making those changes: “Be explicit about the fundamentals each course is meant to build, independent of any tool. AI made both rigorous assessment and explicit learning objectives urgent.”