LunaPrompts
When a recruiter suspects AI cheating, the next step should be a documented review of the interview rules, the observed behavior, and the candidate’s work. An alert or an unusual answer is a reason to investigate. It does not establish misconduct by itself.
A clear process helps the team handle both genuine concerns and innocent explanations. It also stops a stressful interview from turning into an improvised interrogation.
Why HR teams are revisiting interview integrity
In a September 15, 2026 article, Greenhouse describes risks including synthetic identities, scripted interview assistance, and coordinated applications, while emphasizing structured hiring and human oversight. The relevance for recruiters is broader than one tool: different risks need different evidence. Greenhouse’s candidate-fraud guide.
This article focuses on reviewing suspected unauthorized assistance during an assessment or interview. It is an operational starting point to adapt with your organization’s owners, not a claim that every AI-assisted answer is fraudulent.
Start by distinguishing the problem
| Situation | What the team needs to establish | Appropriate starting point |
|---|---|---|
| Permitted AI use | Whether the candidate follows the stated task rules | Assess the work and explanation |
| Suspected unauthorized assistance | Whether evidence indicates a breach of a disclosed rule | Review the session and task context |
| Unclear authorship | Whether the candidate understands and can adapt submitted work | Ask submission-specific follow-ups |
| Possible identity mismatch | Whether the participant is the expected applicant | Use the approved identity process |
Do not collapse these into one “cheating score.” An identity question is different from an authorship question. A strong answer generated with permitted AI may be exactly what the assessment was designed to evaluate.
Put the tool policy in the invitation
Before the assessment, state which tools are permitted, which are restricted, what must be disclosed, and what evidence the process collects. Give candidates a way to ask about accessibility tools or unclear instructions before starting.
A policy should describe the behavior, not rely only on a list of product names. Tools and versions change. For example, explain whether receiving live suggested answers during an unaided interview is permitted, and distinguish it from an approved coding assistant used in a practical task.
For mixed interview loops, repeat the rule at each stage. The candidate should not have to infer that AI is allowed in the take-home but restricted during a later conversation.
A five-step review process
1. Record what actually happened
Write a factual observation with the relevant time and task context: “The session flagged a possible assistance event during question three.” Keep the raw signal separate from the interpretation.
Avoid notes such as “looked dishonest.” Gaze, nervousness, pauses, and speaking style are weak foundations for a misconduct conclusion. They can have many explanations unrelated to unauthorized assistance.
2. Check the configured environment and rules
Confirm which assessment mode was used, what the candidate was told, and what the available signal actually measures. Was a tool allowed? Was an accessibility arrangement agreed? Did a connection failure affect the recording?
If an instruction was ambiguous, record that limitation. A reviewer should not quietly replace an unclear policy with a stricter one after the session.
3. Revisit the candidate’s work
Choose a specific part of the submission and ask the candidate to explain or adapt it. For a code task, ask about a failed test or an implementation choice. For a prompt, change an input and ask how the output should change. For a design, introduce a new constraint.
Use questions of comparable difficulty across candidates. A useful follow-up seeks understanding; it does not demand that someone reproduce an entire solution under surprise conditions. Our AI Viva guide gives examples of grounding questions in submitted work.
4. Give the candidate a chance to clarify
Describe the concern neutrally and explain what information would help resolve it. For example: “We need to clarify which tools were active during this part of the assessment. Can you walk us through your workflow?”
Use the organization’s process for sharing evidence, discussing accommodations, and handling disputes. Avoid asking a candidate to expose unrelated personal information or accounts as an improvised verification step.
5. Have an accountable reviewer decide
Separate the technical assessment result from the integrity review. Record the evidence considered, alternative explanations, the candidate’s response, and the reason for the outcome. Where the available evidence is inconclusive, say so rather than turning uncertainty into a fact.
The outcome could be no further action, clarification of instructions, a proportionate follow-up, or action under the organization’s established policy. The reviewer owns that decision; a software alert does not.
A fictional example of an inconclusive flag
An assessment produces an assistance alert during a coding question. The candidate had permission to use an AI coding assistant but was required to explain generated changes. The initial report does not show whether the alert relates to permitted activity.
The reviewer checks the configuration and asks about a specific change in the submitted function. The candidate identifies a boundary case, explains the generated suggestion, and adds a test. That work is relevant evidence of understanding. It does not automatically explain every telemetry event, so the reviewer records the remaining uncertainty separately.
The important detail is that the team did not treat “AI was present” as proof that a rule was broken. This example illustrates a review approach; it is not a LunaPrompts customer incident.
What LunaPrompts adds to the review
LunaPrompts surfaces identity, session, and authorship signals, including indicators associated with real-time assistance tools such as ParakeetAI and Cluely. Selected secure assessment environments can provide additional desktop context. Coverage depends on the configured environment, tool, and version; it is not a guarantee of detecting every hidden overlay or attempt.
AI Viva adds a way to explore the candidate’s reasoning about their work. Together, the submission, available session evidence, and follow-up can help a reviewer investigate a concern. Ask the team to demonstrate the exact workflow during an assessment integrity pilot, and review the integrity feature guide.
Test your process before a hiring campaign
Run a controlled pilot with consenting internal participants. Include permitted AI use, a simulated policy violation, a technical interruption, and an ambiguous situation. Check which signals appear and whether reviewers interpret them consistently.
Record missed test cases, false alerts, review time, and unresolved cases. Do not report accuracy without a defined test set and known outcomes. Give the team a clear escalation path before increasing candidate volume.
Questions recruiters ask
Is using ChatGPT during an interview always cheating?
No. Whether assistance is permitted depends on the disclosed rules and the capability being assessed. Make that distinction explicit before the task begins.
Can eye movement or a long pause prove hidden assistance?
No. Those observations are ambiguous and should not determine an outcome on their own. Focus on relevant evidence, a fair explanation process, and the stated rules.
Should an integrity alert automatically reject a candidate?
Use a human review process with an accountable owner. Inspect what the alert means, the task context, and the candidate’s explanation before acting.
How does this fit with AI-led interviews?
Explain both sides of the policy: how the employer uses AI and what the candidate may use. Our candidate experience guide covers invitations, support, and handoffs.