When AI or other forms of assisted cheating become harder to detect, assessment teams usually face three choices: add more detection, redesign the task, or combine both. The evidence points to a hard truth: detection can help triage, but it rarely fixes a task that is already easy to outsource or fake. Redesign is often the stronger long-term response when the assessment is meant to prove unaided performance or authentic judgement.

This page compares two broad security strategies. Detection looks for suspicious behaviour, suspicious response patterns, or signs that AI or another helper was used. Redesign changes the task, conditions, or marking so that misuse becomes less valuable, more visible, or no longer relevant to the construct being assessed;;.

Many programmes add detectors when they really need a validity decision. If a candidate can still complete the assessment through hidden assistance, then more alerts may simply create more review work without improving the meaning of the score. For assessment leaders, the real question is whether the security problem is best solved by catching misuse or by changing the assessment so misuse no longer pays off.

- **Detection**: identifying possible misuse after, during, or around the assessment. - **Redesign**: changing the assessment format, task, or rules to better match the intended construct. - **Triage**: using signals to decide what deserves human review. - **Authenticity**: whether the work reflects the intended candidate performance. - **Construct protection**: keeping the assessment aligned to what it claims to measure.

The stronger evidence suggests that detection is inherently limited. AI-text detectors can be weakened by paraphrasing and other attacks, so they should not be treated as proof. The answer-watermarking idea is more promising because it tries to design detection into the workflow, but it still depends on prompt design and use context. The E-Assessment Association material is consistent with a broader design-led view: revise, relocate, and rethink assessment practice rather than assuming better policing will solve the problem.

The contest is not whether detection has any value. It does. The question is how much investment it deserves relative to redesign. Vendors and some practitioner tools tend to frame detection as a practical defence, especially where programmes need immediate controls. The open question is whether those controls can keep pace with changing AI behaviour, or whether they mostly shift the burden onto reviewers and appeals teams. There is also disagreement about what counts as enough redesign. For some assessments, a small change in prompt structure or task context may help. For others, the issue is deeper: the whole assessment format may need to change if it is to evidence authentic performance.

- endless escalation of detectors with diminishing returns - false positives and burden on review teams - candidate distrust if monitoring becomes the main response - poor alignment between assessment format and intended construct - overconfidence that a flagged score is already a proved misconduct case

1. Ask what the assessment is meant to prove. 2. Identify the most plausible misuse route. 3. Decide whether the route can be closed, narrowed, or made visible. 4. Use detectors only where they add triage value. 5. Redesign when the same misuse keeps recurring. 6. Recheck validity after any major rule or format change.

| Option | Best strength | Main weakness | Use when | |---|---|---|---| | **More detection** | Faster to deploy | Can be brittle and review-heavy | You need an interim control | | **Task redesign** | Reduces the value of misuse at source | Requires revalidation | The assessment can change without losing its purpose | | **Combined approach** | Balances short-term and long-term risk | Needs coordinated governance | The programme has time to improve steadily |

A university finds that AI-written essays are passing through text-detection software with mixed results. Rather than only tightening thresholds, it changes the task so students must justify decisions using local case material, draft reflections, and viva-style follow-up. That reduces the value of generic AI output and gives the institution better evidence of the student's own judgement.

- TCN note on watermarking-based detection embedded in answer-generation workflows. - TCN note advocating revise, relocate, and rethink strategies for undetectable cheating. - TCN note on the limits of AI-text detection under paraphrasing attacks.

The market is split between detector vendors, assessment platforms, and specialist consulting. Vendors often emphasise speed, automation, and coverage; redesign-oriented sources emphasise assessment meaning, construct alignment, and governance. For buyers, the useful question is not which product detects more, but which approach best protects the intended evidence.

### Is detection useless if AI keeps changing? No. It can still help triage and deter some misuse, but it should not be treated as the whole answer. ### When is redesign a better move than better detection? When the same misuse pattern keeps appearing, or when the assessment can only work if you assume candidates will not use AI. ### Can I combine watermarking with redesign? Yes. That is often the most sensible option when a programme needs both immediate protection and a more durable long-term design change.

Tim Burnett (Admin)

`Test Community Network. "Detection versus redesign in test security." TCN Wiki. Last reviewed 2026-05-05. https://www.testcommunity.network/wiki/test-security-detection-vs-redesign`

- TCN note on watermarking-based detection embedded in answer-generation workflows. - TCN note advocating revise, relocate, and rethink strategies for undetectable cheating. - TCN note on the limits of AI-text detection under paraphrasing attacks.