AI Automation ·
Does a better AI-assisted assignment show better learning? What should Korean universities assess separately?
A Google News US search for AI education led to Dartmouth’s November 20 announcement of course-redesign support. It includes purposeful AI use and learning designs that do not depend on AI.[1] We read this as an opportunity to compare designs against learning goals, rather than choose one campus-wide position on adoption. Funding plans, however, are not evidence of learning gains.
A polished submission and a student’s retained capability are different objects
Imagine a Korean university writing or policy-analysis course in which an AI-assisted report improves. Its quality can be assessed, but one report cannot fully establish whether the student can judge evidence independently or transfer an argument to a new case. Nor does AI use establish absence of understanding.
Our question is how Korean course redesign should distinguish supported output quality from independent understanding and transfer. We followed the announcement to operating information, a primary experiment separating assisted practice from unassisted assessment, and Korean university guidance. We conducted no student experiment or faculty interview. The assessment arrangements below are proposals.
Following the support model reveals work beyond purchasing access
Dartmouth’s funding page describes instructor cohorts, design consultation and help collecting student feedback.[2] These are stated support arrangements, not comparative findings about effective instruction or proof of superiority over Korean universities.
Our operational inference is that distributing AI access and redesigning a course are separate investments. Turning a learning objective into questions, feedback and assessment requires time. Comparing assignment grades after access expands may not distinguish the student’s understanding from the tool’s corrections. This is analysis of the evaluation structure, not an observed effect of the programme.
The primary experiment actually separated availability from removal
A June 25, 2025 PNAS paper reports a randomised mathematics study with nearly 1,000 students in one Turkish high school. In its autumn 2023 GPT-4 setting, the general-tool group improved assisted practice but performed worse than controls on a subsequent unaided test. The learning-guardrail group did not demonstrate a clear unaided-test gain either.[3]
We do not transfer an effect size to Korean university writing. School, subject, contemporary tools and short-term assessment differ. We also opened the August 20 correction, which concerns an author’s affiliation.[4] We did not reanalyse raw data.
The finding relevant to our question is not universal AI harm. It is experimental reason to separate assisted performance from unaided performance because they can yield different judgments. Whether similar changes occur in Korean writing or policy analysis remains a question for local evaluation.
Permission rules do not themselves establish learning effects
Korea National Sport University’s August 13, 2024 guidance recommends stating AI-use policies in syllabuses and considering varied assessment and questions requiring thought.[5] We checked the public page on November 21. It is neither a unified rule for all Korean universities nor measured evidence of the guidance’s effectiveness.
The comparison separates telling students what use is permitted from assessing whether it develops intended capabilities. Clear permission contributes to fair expectations; it does not prove learning gains. Instructor discretion also cannot justify comparing unlike class sections’ grades without adjustment.
| Comparison | Dartmouth’s public material | Korean university’s public material |
|---|---|---|
| Design scope | Supports purposeful AI and AI-resilient learning[1] | Recommends stating use policy in the syllabus[5] |
| Support and assessment | Design consultation and student-feedback support[2] | Varied assessment and questions requiring thought[5] |
| Not established | New programme’s learning gains | Guidance effects or practice across all Korean universities |
These are not treatment and control groups. Our proposed Korean implementation adds time for instructional design and assessment alongside clear permission. We omit unverified funding amounts and adoption rates.
Option A starts with an independent attempt, then comparison with AI
In an introductory unit, students first record their argument or solution, then examine differences from AI suggestions. Initial reasoning and revision explanations can reveal where feedback is needed. Compare A where exposing basic understanding before assistance is central.
It takes time and may delay support. Leaving an underprepared student alone with a difficult task does not guarantee learning. Connect bounded attempts with explanation, examples and instructor feedback. Low initial polish should not automatically become a grading penalty.
Option B permits assistance and separately examines independent explanation
Allow AI-supported drafting and counterargument exploration, then separately ask why the student selected or rejected evidence and how the reasoning applies to a new case. Consider this where collaboration with a tool is an explicit objective and foundational understanding has developed. Score the output and independent explanation against different criteria.
Collecting longer chat histories is insufficient. Review burden and privacy exposure can grow, and a polished retrospective explanation is not automatic proof of understanding. Ask about a few consequential decisions, collect only necessary records and provide equitable tool access and alternatives.
Neither option substitutes AI-detection scores for learning assessment. A need not exclude AI from the whole course; B does not permit unlimited use in every assignment. Scope depends on concepts, readiness and available feedback resources.
Separate output scores from reasoning about a new problem in the pilot
Departments and instructors should narrow a goal to observable behaviour, such as explaining the connection between an evidence source and a claim. The teaching centre can help create comparable questions and scoring criteria. Explain permitted activities, assessment targets and record collection beforehand. Preserve accessibility support and reasonable accommodations under both alternatives.
We propose checking starting levels and using similar concepts with consistent criteria, recording assisted products separately from independent explanations. A later new case can help distinguish immediate recall from transfer. Differences in section composition or task difficulty prevent a simple causal interpretation.
Proposed measures are evidential accuracy of outputs, validity of independent explanation, application to new cases, feedback time and differences in access burden. Satisfaction and usage frequency are supporting observations, not learning outcomes. Sample size, period and acceptance criteria must be specified for the course beforehand; we measured none.
If products improve without stronger independent explanation, withhold expansion and revise when or how assistance is used and how questions are designed. If independent-first work only increases burden and delays feedback, narrow A as well. Neither alternative has won before the proposed comparison is conducted.
We established why assessment needs separation, not a Korean effect
Our answer is to obtain different evidence for completed assisted work and retained understanding. The overseas support model identifies implementation resources; a different-context experiment adds reason not to infer learning from output alone. Combining them does not establish local effects.
Long-term retention and transfer, accessibility differences and total instructor workload remain unverified. Repeated comparisons with the same learning goals and assessment criteria could change the order of A and B. Does the improvement report distinguish a better submitted report from a student’s ability to explain its evidence independently?