The Erosion of Traditional Assessment Models
“AI is actively undermining several evidence-based teaching practices (e.g., see How Learning Works”
The integration of generative AI into the student workflow has rendered traditional homework models obsolete. For educators like Christian Kästner, the realization that LLMs could consistently pass rubrics without genuine comprehension necessitated a shift in strategy. The core challenge lies in the fact that AI tools now effectively externalize the cognitive labor of learning, allowing students to submit high-quality work without engaging in the underlying critical thinking processes. This shift has forced a departure from evidence-based pedagogical practices that previously favored frequent, low-stakes assignments. By moving away from take-home work, educators are attempting to reclaim the integrity of the learning process, though this often comes at the cost of increased student stress and a departure from the ideal of frequent, formative feedback. The struggle is not merely about preventing cheating; it is about maintaining a learning environment that rewards genuine intellectual effort in an era where the cost of generating a plausible solution has plummeted to near zero.
From Written Reflection to Oral Verification
“After every assignment, each student needs to schedule a 15-minute meeting with a TA”
As written assignments become increasingly susceptible to automation, the focus of assessment has shifted toward oral defense and live interaction. The move to 15-minute, one-on-one check-ins with teaching assistants represents a significant investment in human capital to ensure that students actually internalize the material. These sessions serve as a mechanism to verify that the student understands the technical tradeoffs and conceptual frameworks presented in their work, rather than just the final output. While this approach demands significant time from instructional staff, it aligns with the professional requirements of the field, where communicating technical decisions is as critical as the code itself. By treating these interactions as a core component of the grading rubric, educators can maintain high standards while mitigating the risks posed by AI-generated submissions. This transition reflects a broader trend in education: when the written word can be synthesized by a machine, the human voice becomes the primary site of verification and pedagogical engagement.
Scaling Human Interaction Through Automated Triage
“TAs spend 50 to 80% less time grading (grading 80% less content)”
To facilitate the shift toward more intensive in-person interactions, educators are increasingly leveraging AI to handle the mundane aspects of grading. By employing an 'LLM-as-a-judge' approach, instructors can filter student submissions, identifying those that clearly meet specifications and isolating those that require human intervention. This strategy does not eliminate the human element but rather reallocates it; teaching assistants spend less time grading repetitive code and more time engaging in substantive, face-to-face discussions with students. This model of human-in-the-loop assessment allows for a more efficient use of resources, ensuring that the time spent by TAs is focused on high-value pedagogical interactions rather than administrative review. By automating the initial pass/fail judgment, the course structure maintains rigor while acknowledging that the sheer volume of AI-assisted output would otherwise overwhelm traditional manual grading workflows. This symbiosis between machine-assisted grading and human-led verification is likely to become the standard operating procedure for technical courses.
Redefining Learning Goals in a Post-Automation Landscape
“I barely touched the overall learning goals . I am fortunate that this is”
Despite the radical changes in assessment methods, the underlying learning objectives for advanced engineering courses remain remarkably stable. The focus remains on engineering tradeoffs, risk mitigation, and collaborative teamwork—skills that are harder to simulate than rote coding tasks. Educators are finding that while AI can assist in the execution of these tasks, the ability to architect systems and navigate complex professional environments remains a distinct human competency. The decision to allow AI usage in almost all settings reflects a pragmatic acceptance of the technology as a new tool in the engineer's toolkit. By encouraging responsible use rather than policing it, instructors are preparing students for a reality where AI is a constant companion in the development process. The goal is no longer to test whether a student can write a specific function, but whether they can manage the lifecycle of a production-ready system while effectively utilizing the tools at their disposal to solve real-world problems.