Where AI Can and Cannot Be Used for Assessment

The question of AI in assessment is not simply a choice between a complete ban and full automation. Consider the specific action: prepare a criterion, help formulate feedback, match an excerpt with a requirement, or make a final decision. These actions have different consequences and require different evidence. A useful boundary lies where a support tool stops offering verifiable material and begins replacing instructor responsibility with an unexplained conclusion.

This article offers an instructional framework rather than a universal legal interpretation. Procedures for handling data, grades, and learning materials are determined by the particular institution's rules and applicable requirements. We will consider hypothetical situations: a feedback draft, a disputed written answer, and a practical demonstration. They do not describe completed experiments. The aim is to help course authors ask the right questions about AI use without attributing unvalidated capabilities to it.

Name the Decision First

'Using AI for assessment' is too broad a phrase. A model might merely shorten an already approved human comment or independently suggest a score. In the first case, the instructor has established the substantive basis. In the second, the tool influences the decision. Describe the process through specific steps and authority. Otherwise, students and the learning team may interpret one statement differently despite formally agreeing to use the same service.

Separate preparing the assignment, reading the work, formulating feedback, and the final decision. For each step, name what the model does and who checks the result. If the answer is unclear, the process is not ready. Do not hide complexity behind 'assistant.' An assistant can have different levels of access and influence. Its role must match actual use rather than merely a gentle wording in the course description.

Carnegie Mellon's Eberly Center suggests considering motivations, quality, data, and communication before using generative AI for assessment and feedback. This framework helps begin with the learning task. Saving time may be an intention, but is not evidence of suitability. If feedback requires extensive correction or introduces new unsupported claims, fast generation alone does not improve the process.

Appropriate Assistance During Preparation

AI can be used to draft criteria if the author checks them against the objective and task conditions. A model can help find an unclear word, propose observable features, or create a hypothetical borderline example. The final rubric remains the author's decision. Show it to students before they complete the task. Do not silently add a new criterion after reading work merely because the model called it important or professional.

Another appropriate step is editorial review of a test question against a verified source. Ask for alternative readings and the basis of the key. A person checks the conclusions independently. If the model cannot see a table or image, its review is limited to available text. Assignment preparation does not become automatically error-free, but gains another controlled perspective. This step does not require sending actual student work when illustrative fictional examples are used.

AI can also help simplify instructions. Check that shortening has not changed the requirement. 'Explain the choice' is not the same as 'name the option.' Language editing affects the learning action even when it seems technical. Authors must compare old and new versions and accept the one preserving the intended meaning. An elegant phrase is no advantage if students now perform a different task.

Feedback Drafts with Human Checking

A hypothetical student text contains a conclusion going beyond the data. The instructor has already identified the problem and wants to explain it clearly. A model can help turn a note into a short comment with a location, rationale, and next action. A person verifies every claim before sending it. If AI adds a quotation or psychological characterization, remove it. The tool's role is to formulate a controlled message rather than establish an invisible cause of the mistake.

The relevant excerpt, task conditions, and criterion are sufficient for this. Do not provide the student's entire history if it does not help explain the current step. Personalized feedback arises from specific work. A model must not judge motivation, attentiveness, or ability from one paragraph. Feedback should concern a decision the author can revisit. The supporting step then stays within the learning task without creating an unnecessary profile of a person.

When students receive AI-assisted feedback, they must know who is responsible for its content and how to ask a question. Do not send an unchecked result as a final decision. If the course author cannot explain the message's basis, the feedback is not ready. Responsibility is not transferred to a model by mentioning technology use. The learning team accepts text it can discuss against the actual work and announced criterion.

Supporting Comparison Is Not a Final Grade

A model can locate passages meeting a particular requirement, for example identifying a claim and its support. This is a useful starting point for rereading if the quotations are genuine. However, it may miss an indirect but sufficient explanation or attribute a stronger conclusion to the text. The instructor returns to the original. The final decision needs context: working conditions, available materials, level, and assessment procedure.

Do not use a model's response as an unconditional argument in a dispute. 'AI assigned this level' does not explain which feature supports it. Students may show a passage contradicting the conclusion. A person must consider it and change the decision if needed. A transparent procedure matters more than preserving the initial score. The tool helps with reading, but does not gain authority to interpret every permissible response finally without responsible review.

If the learning team considers wider use, verified examples, agreed criteria, and ongoing monitoring are needed. One successful comment does not prove reliability across all work. Different formats create different boundaries. A written note, diagram, and demonstration do not provide the same material. Name exactly what was checked and do not extend a local result to an entire discipline. AI should not receive a broader role merely because its first answer is confident.

Where Direct Observation Is Required

A practical demonstration may assess an action sequence, interaction, or professionally significant procedure. A textual description does not prove performance. If AI sees only a report, its conclusions are limited to that report. A final decision about an action requires appropriate observation and competent checking. Do not replace these with a convincing retelling. Preserving this boundary is particularly important where assessment determines access to subsequent practice involving responsibility.

A similar problem arises with group work. The final product does not automatically show each person's contribution. A model does not know who proposed a solution, listened to partners, or performed a particular stage. If those actions are assessed, arrange the necessary evidence beforehand. Do not ask AI to distribute individual scores from a shared file without grounds. Organizational convenience cannot compensate for missing data about the assessed action.

For an oral answer, a transcript can support content review but lose important context. If interaction is part of the objective, the lines and course of the conversation are needed. If only the argument is assessed, an approved text may suffice. The author defines the boundary. AI must not infer communication skill from an excerpt containing no response to other participants. Assess exactly what the material actually shows.

Where Style Alone Cannot Support a Conclusion

An unusual phrase or polished structure does not establish how work was created. If the course has AI-use rules, authorship questions are considered under an agreed procedure using available grounds. Do not turn a model's guess about 'machine text' into evidence. Product assessment and checking compliance with rules are different tasks. The latter needs a separate procedure extending beyond an impression of style.

Students may explain their argument, show revisions, or discuss a decision if the conditions provide for this. That work can offer evidence of understanding, but is not a universal way to prove a file's entire history. Do not promise absolute certainty. Course authors must name the learning action being checked and the conclusions the material permits. AI does not automatically reconstruct an invisible process, even when it proposes a plausible account.

When a tool makes a suspicious characterization of a person, stop that conclusion. 'The student is prone to cheating' is not subject-specific feedback on a paragraph. Likewise, 'low motivation' does not follow from an omitted answer. Consider the specific fact and appropriate clarification procedure. A model can help formulate a neutral question about the work, but must not create a personal label that subsequently affects all the instructor's decisions.

Data and Authority

Before use, determine which material can be sent to the chosen tool under your organization's rules. Anonymization helps but does not always remove every risk: text may contain recognizable experiences or other people's data. Use the minimum necessary excerpt or perform the review without sending it. This article does not replace checking specific requirements. Consider instructional benefit alongside the acceptability of the process, rather than after bulk uploading work.

A provider's security statement is not automatic permission for every task. Conditions depend on the service, account environment, and organizational procedures. Do not transfer one institution's rules to another merely because the product name is the same. Authors need to know where the supporting step occurs and what material goes there. If this is unclear, the process needs clarification before use, rather than a confident AI conclusion about general compliance.

A Transparent Course Policy

A short description can name permitted support tasks, human review, and how to ask questions. For example, AI drafts comments on practice versions, while the instructor approves every message. This hypothetical example must be adapted to the actual procedure. Do not promise automatic feedback if the tool has not been validated for that mode. Stated roles must match what the team actually does.

Explain the limits on student use too. If structural suggestions are allowed but independent justification is required, say so before completion. Instructor and student policies may serve different tasks, but their reasoning should be clear. Do not create a situation where students are forbidden formulation assistance without explanation while instructors send unchecked automatic decisions. Transparency supports discussion of the learning objective, rather than only the fact of technology use.

Define how concerns about feedback are considered. Students identify the excerpt and criterion; the instructor explains the basis and corrects the decision if needed. Not every question needs to become a major formal dispute. However, the channel must be real and accessible. AI must not become a final authority that people cannot question on substance. Responsible assessment leaves room to clarify evidence and consider an acceptable alternative.

How to Decide on a Specific Step

Ask: what result is needed, which data is available, can every conclusion be checked, who will approve the message, and what are the consequences of an error? If there is no evidence for a required feature, AI must not assess it. If the basis is verifiable and human oversight is organized, a supporting step can be considered. This is a framework for the author's decision, not automatic permission. Assess the specific process rather than technology as a single unchanging object.

Start with a small task without real personal data, such as editorial review of hypothetical feedback. Check whether the tool preserves the criterion and avoids adding conjecture. Then decide whether it benefits your course. Do not scale up merely because the first answer sounds successful. Quality lies in verified content rather than speed or persuasive wording. If the process does not help, it can remain outside assessment.

A Practical Boundary Map for One Assignment

It is useful to assign one responsible person to each stage of this map. Do not rely on a general 'the team checks.' If the instruction author and work reviewer differ, pass the approved criteria and material together. The boundary of AI's role then survives both in the prompt and in the actual workflow between people making the learning decision.

Consider a separate hypothetical assignment: a student prepares a short explanation of a source choice. They must show the material's origin, name the supporting basis, and acknowledge missing data. Before completion, the instructor can ask AI to check the instruction for ambiguity. This concerns the author's text, not a student's grade. The result is a list of possible misunderstandings. A person reads it, accepts relevant corrections, and saves the final task conditions. Risk is lower at this step, but still not zero: the tool may add a new requirement.

After submission, AI can help find where the written text names the source. However, failure to find an excerpt does not prove it is absent. The author may have used another form or placed the link in a separate field. The instructor must check the full material. A useful working note is: 'The machine result is a candidate for reading, not a final decision.' This short but important boundary prevents convenient search from becoming an automatic accusation of failing a requirement.

At the feedback stage, AI can offer several short formulations based on an already established human conclusion. For example, the instructor has determined that the source is named but its connection to the claim is unexplained. The draft should then request that specific connection. There is no need to give the tool authority to reassess all the work without an explicit task. If a new shortcoming appears in the draft, it needs separate reading. Good language does not make an unchecked claim correct or justify adding it to the assessment.

The final conclusion must rest on the task conditions, criteria, and actual work. A hypothetical record can contain three fields: what is confirmed, what needs correction, and what cannot be determined. The third is particularly useful. It reminds us that not every question can be resolved from the supplied text. If we need to know how a student selected material, a written product without a process description does not show this. A conclusion about the working method needs other evidence or appropriate clarification, rather than a model's assumption.

Five Stop Conditions

The first is an undefined source of criteria. If AI receives only the work and a request to 'assess quality,' it will apply its own ideas. Return to the announced requirements before using the result. The second is partially read material. A recognized PDF excerpt is not the whole assignment. A final conclusion cannot be made if important tables or attachments are inaccessible. Readable material or another permitted review method is needed, rather than a confident generalization from a fragment.

The third is feedback about behavior or intent. 'Unwillingness to work' does not automatically follow from a short answer. Remove it and return to the observable action. The fourth is a decision changing course conditions. A model may advise deducting points for assistance that was not previously restricted. Such a rule cannot be silently applied retroactively. The fifth is data that may not be sent to the chosen environment. Here, stopping is part of a responsible process rather than a technical inconvenience to bypass.

Stopping does not always mean abandoning AI forever. The task can change: work with hypothetical answers, edit already approved feedback, or check only the author's instruction. Do not transfer an unfinished process to a new tool without analyzing the cause. If the obstacle concerns data rights, another service does not automatically remove it. If the problem is an unclear rubric, a new model will not make it transparent by itself. Change the basis of the work first, then repeat the permitted step.

How to Explain Use to Students

A short explanation can name the specific role: 'AI helped edit the wording of feedback after the instructor's review.' This differs from 'AI assessed your work.' Neither exaggerate nor conceal the tool's involvement. Participants must understand who makes the decision and how to ask about a factual error. If the course provides an appeal procedure, refer to the current rule. Do not create a new procedure in one comment merely because AI proposed it as a considerate closing.

After a student's inquiry, review the basis rather than defend the machine text. They may have identified a missed excerpt, misreading, or changed criterion. Checking requires the work and announced task conditions. If the conclusion is corrected, explain the change through specific evidence. People then see that assessment is a process that can be justified. Trust comes from the ability to check a decision and correct it when its basis was incomplete, rather than a promise of infallibility.

A Boundary You Can Explain

Before use, write one short sentence describing the tool's role. If it does not make clear what remains with a person, refine the process. In different tasks, AI can be an editor, a source of options, or an assistant for locating excerpts. The final decision must rest on announced criteria and actual evidence. Missing data is not compensated for by a confident prediction, and responsibility does not disappear through automatic wording.

In WeStudy, connect task conditions, criteria, and approved feedback with the corresponding work. The WeStudy LMS helps organize learning, while the course team defines AI's boundaries. A supporting tool is appropriate where its result can be checked and explained. Where direct observation, proper authority, or a decision about a person without available grounds is involved, an automatic conclusion must not replace responsible assessment.

Sources and Further Reading

• Carnegie Mellon's Eberly Center suggests considering motivations, quality, data, and communication