Assignments, Assessment, and Feedback

When AI is involved in assignments and assessment, how can we tell what students have truly learned?

5 claims95 direct citations95 original sourcesUpdated 2026-09-21

Written and maintained by Robin · Updated 2026-09-21

This answer synthesizes public sources. Check the evidence and limits before applying it. Review the evidence

Citation text

Robin. When AI is involved in assignments and assessment, how can we tell what students have truly learned?. 2026-09-21.
Assessment reform continues to combine process disclosure, oral verification, independent performance, and teacher judgment. A return to oral exams, a no-AI writing requirement, and a hidden-instruction incident show task design changing, while corrected sourcing for New South Wales strengthens the policy evidence. Validity, fairness, student appeals, and tool-free transfer remain the tests.
https://edu.hhhh.life/en/topic/assessment-feedback/#answer

LIVING ANSWER

Follow the question. Read the answer directly.

Start with the claims, then open the evidence when you need it.

QUESTION 1

Which evidence comes closer to genuine learning?

Assessment should examine the process, independent performance, and long-term outcomes together, without treating faster completion, more output, or higher satisfaction as direct learning gains.

A working paper covering about 27,000 Chinese secondary students over 30 months reported higher homework scores and shorter completion time after generative-AI use, alongside an approximately 20% decline in monthly closed-book examination scores within six months. The result is currently available through media coverage of a working paper, so assessment should track both tool-assisted tasks and tool-free transfer while awaiting primary review and replication. [01][02][03] plus 16

19 sources for this claim Supports 2 · Illustrates 4 · Tracks 13 View all ↓Collapse ↑

Assignment assessment should record where students used AI and preserve drafts, citations, revision reasons, and oral explanations. Instructors at Wyoming and Penn have returned to oral exams, Boston University has set a no-AI writing expectation, and Danish schools are adding oral defenses and process monitoring. Each approach needs allowed-use rules, workload review, appeals, and alternatives. [20][21][22] plus 18

21 sources for this claim Supports 1 · Illustrates 13 · Tracks 7 View all ↓Collapse ↑

QUESTION 2

To what extent can AI participate?

AI can accelerate low-risk, reviewable formative feedback. Open-ended assignments, qualifying exams, and decisions that affect student opportunities require clear rules and final human responsibility.

AI can generate practice, organize error types, recognize handwritten responses, conduct spoken-language assessment, and propose feedback drafts, while teachers should check accuracy, task goals, and student context. Products from Google, Zuoyebang, Mobile Learning Companion, Medly, Yuanfudao, and others are expanding access to instant feedback. Product-reported score gains, reading time, willingness to express ideas, accuracy, and personalization still need to be validated through independent tasks and long-term outcomes. [20][41][42] plus 32

35 sources for this claim Supports 1 · Illustrates 14 · Tracks 20 View all ↓Collapse ↑

High-stakes assessment needs human judgment, explanations, and appeals. The official source for New South Wales assessment rules has now been corrected, while Instruction Partners' tool reviews and school authentication practices add implementation examples. Each approach still requires evidence on validation samples, fairness, error handling, and effects on student opportunity. [75][20][76] plus 7

10 sources for this claim Supports 2 · Illustrates 6 · Tracks 2 View all ↓Collapse ↑

Assessment is expanding beyond a single answer to the task process: question formulation, source verification, revision records, collaboration, and reflection can all become evidence of learning. Process records should serve feedback and should not be reduced to another automated scoring metric. [84][20][85] plus 10

13 sources for this claim Supports 2 · Illustrates 6 · Tracks 5 View all ↓Collapse ↑

WHAT WE STILL DON'T KNOW

Limits of this answer

Public evidence is still insufficient to support the following judgments. Future updates will prioritize original material that can narrow these gaps.

01
There is no consistently reliable method for automatically detecting AI-written work, so detection cannot be the sole basis for disciplinary action.
02
There are many formative assessment tools, but public evidence on validity, bias, and comparisons across groups remains insufficient.
03
The survey of teenagers is self-reported. It can describe behavior and perceptions but cannot by itself show that AI has caused harm to or improved learning outcomes.