How to Use AI Coding Agents Without Losing Assessment Integrity

Assessment integrity is not the same as misconduct

A course can have weak assessment integrity even when nobody breaks a rule. For example, an agent may be able to complete most of a task while the submitted artifact provides too little evidence for the individual learning claim. Conversely, AI use can be permitted and an assessment can still support a defensible judgment when the task and evidence fit the outcome.

Misconduct is a separate question: did a student violate an applicable assessment or institutional rule? AI involvement by itself does not answer that question. A course-level design review asks whether its conditions and evidence support the learning claims; an individual conduct concern must follow the institution’s stated procedures.

Choose AI conditions to match the claim

AI prohibited may be appropriate when unaided performance is itself an intended outcome. A prohibition does not automatically make the task sound or establish that the student met the outcome. AI restricted requires clear boundaries about which tools or uses are allowed and how assessors should interpret them. AI permitted makes AI part of the assessment conditions; evidence must still support the learning claim. No one condition is right for every course.

If the claim is that a student can perform a particular reasoning task without AI support, controlled conditions may be relevant. If the claim is that the student can use an agent while detecting and verifying unsafe authorization behavior, banning AI may remove part of the intended capability. Explain conditions and disclosure expectations in advance. A disclosure can provide context about declared use, but does not itself establish learning or compliance; see Student Disclosure. For individual-task alignment, see Assignment Assessment; for tasks designed to retain their learning purpose when AI is available, see AI-Resilient Assignments.

Use evidence that answers the learning question

Evidence should be sufficient for the intended learning claim and proportionate to its stakes. Depending on the outcome, it might include the artifact, tests or other verification, selected development records, a controlled task, focused explanation, project or portfolio evidence, or disclosure. No assessment needs every type. A completed artifact or passing build does not automatically show every skill the course intended to assess.

Use evidence that is relevant and potentially complementary when the claim calls for it. Several weak or redundant signals do not become strong evidence by being counted together: a screenshot of a passing test and the log of that same run may reflect the same observation. Selected process records can inform questions about decisions, revisions or verification, but do not by themselves make an assessment valid or fair. For choosing records, see Process Evidence. A detector flag is likewise not learning or misconduct evidence; see AI Detection.

Constructed example: authorization behavior in a backend service

Constructed example — not classroom, student, customer or TraceYield evidence. A second-year HBO-ICT assignment asks students to design, implement and verify robust authorization behavior in a web service. AI implementation assistance is permitted. Students are expected to justify security-relevant decisions and verify important behavior.

An agent drafts an authorization check and the student modifies it. The submitted implementation works for the supplied tests, but an important negative scenario is missing. The student can explain some, but not all, of the reasoning. The artifact and passing tests support claims about the behavior they cover; they do not establish that the omitted case was considered or that the full verification outcome was met. The evidence does not itself show a rule breach.

The assessor can identify the remaining question—whether the student can recognize and verify the missing authorization case—and use a proportionate follow-up, such as asking the student to inspect or adapt a relevant test. The course should judge that evidence against its stated criteria. It need not resolve every uncertainty or infer who “really wrote” the initial code.

Assessment security and authenticity are not validity by themselves

Security concerns preserving intended conditions and reducing unauthorized assistance. A locked-down exam may be secure yet measure a capability that the course did not intend to certify. A realistic AI-assisted project may resemble professional work yet fail to provide enough evidence about an individual student. Authenticity is one design consideration, not a guarantee of validity. TEQSA’s discussion of assessment security is Australian sector guidance, not Dutch policy.

Review coverage across the course or programme

Not every learning outcome needs the same AI condition. A programme may combine AI-assisted projects, controlled individual work, practical tasks, portfolios, exams or focused follow-up when each component has a reason tied to an outcome. The system-level question is: where, and under what conditions, is each important learning outcome demonstrated? This does not prescribe a particular mix or require every programme to use both controlled and AI-assisted assessment.

Keep assessment conditions fair and workable

Where AI use is required or explicitly assessed, consider whether students have workable access to institutional accounts, comparable capabilities, usage limits, hardware and support, and whether accessibility needs are met. Exact equality may not be possible, but material differences should be visible in the design. If AI is optional, consider whether students have a viable alternative way to demonstrate the same outcome.

When evidence conflicts or a source is missing

Suppose an artifact is strong, verification is weak, process records are sparse and the student gives a plausible explanation. Do not average these into an integrity score. Ask which learning claim each item bears on and whether the evidence supports that claim. A focused follow-up, narrower judgment or another opportunity may be appropriate. If there is a separate concern about a rule, use the institution’s applicable procedure rather than treating conflicting learning evidence as a misconduct verdict.

Avoid making a consequential decision depend entirely on one detector, repository log, oral exchange, telemetry source or inaccessible tool. This does not mean every outcome needs redundant evidence. Consider alternate evidence paths where a source may be fragile or unavailable. An oral or practical follow-up can address a named uncertainty, but communication, language and accessibility conditions matter; oral skill should not accidentally become the construct. For a focused oral method, see Oral Defense.

Review the assessment system, not just the incident

For consequential judgments, shared criteria, assessor calibration, moderation or a second review may help staff interpret evidence consistently. These practices do not guarantee validity or fairness. A 2025 integrative review of higher-education moderation literature—broadly higher education, largely Australian and not specific to programming—highlights collaborative moderation, clear marking guides, staff discussion and resourcing as relevant considerations.

After an assessment, ask: What did a pass claim? Under which AI conditions was the outcome assessed? What evidence supported that claim? What happened when evidence was missing or conflicting? Were students working under materially comparable conditions? Did assessors apply the same criteria? These are review questions, not a formal integrity checklist. For programme-wide evidence across pieces of work, see Portfolio Assessment; for responsibility and understanding, see Student Ownership. If students work collaboratively, Individual Contribution addresses that distinct judgment.

Where captured, TraceYield may surface session or project context, interactions, steering, decisions, corrections, verification or patterns across work. Those records are optional context: they do not validate an assessment, ensure fairness, establish learning or authorship, detect misconduct, define permitted use, replace moderation or determine a grade. Evidence collection should serve the stated assessment purpose; institutional policy and procedures remain with the institution.