How candidates use AI in assessments and how explain-back checks catch it

AI use is legitimate when it follows the stated assessment rules and preserves the evidence the employer needs. The same assistance can be appropriate in one task and inappropriate in another. Define the intended skill first, then decide which tools support a realistic evaluation and which would perform the assessed work on the candidate's behalf.

For an analyst role, generating a draft formula may be permitted if the candidate must inspect assumptions and verify the result. If the exercise measures independent formula construction, that assistance changes what the submission establishes. Neither condition is inherently right for every role. The instructions should explain which one applies.

Distinguish preparation, task completion, and follow-up. A candidate may reasonably use AI to practice before an assessment. That does not automatically mean AI is permitted during a response or an explain-back check. Accessibility tools also need separate consideration rather than being treated as interchangeable with answer generation.

Use pre-employment testing to examine applied skill under defined conditions. A hiring team needs to know whether the candidate can make the relevant decisions, recognize errors, and take responsibility for an output. A blanket preference for unaided work may miss how the job operates, while unrestricted assistance may hide an essential skill. State the purpose and permitted conditions together.

Undisclosed AI use becomes a concern when the candidate's assistance conflicts with clear rules or required disclosure. A polished answer alone does not establish either condition. Review what the candidate was told, what assistance was allowed, and what the submitted work demonstrates before making an integrity judgment.

At a broad level, candidates may use AI for rewriting, analysis, code generation, or responses to assessment prompts. Some candidates understand the resulting work and can evaluate it. Others submit an output they cannot explain. The hiring implication is different, even when the finished submissions look equally fluent.

Keep separate records for skill evidence and possible policy concerns. A candidate can demonstrate weak reasoning without having violated a rule. A candidate can also demonstrate strong technical knowledge while raising an unresolved concern about a clearly restricted tool. Collapsing these into a single suspicion score makes the decision harder to explain.

• Check that the relevant rule appears before the task begins.
• Distinguish required disclosure from a complete prohibition.
• Ask for clarification when the evidence is incomplete.
• Apply the same review standard to comparable cases.

Avoid asking candidates to prove they never use AI. That question is broader than the assessment and offers little useful evidence. Focus on the stated conditions, the work, and the candidate's explanation of decisions relevant to the role.

Multiple-choice questions and take-home exercises become weak evidence when the employer treats a correct output as proof of independent ability. AI can help produce answers without establishing that the candidate understands the underlying problem. The issue is the inference drawn from the output, rather than the mere presence of a particular format.

Multiple choice can still check bounded knowledge when the questions match the role and the conditions are appropriate. It offers less visibility into how a candidate interprets ambiguity, checks an assumption, or handles a tradeoff. A selected option cannot explain what the candidate notices or why an alternative is rejected.

A take-home task offers richer work, but a finished document or code submission still leaves questions. Can the candidate identify its limitations? Do they understand the assumptions? Can they adapt it when a relevant condition changes? Adding an explanation step makes those questions observable.

Keep the exercise small and realistic. Provide fictional materials, relevant references, and a specific deliverable. Do not ask for a production feature or an unpaid project that benefits the employer. Our pre employment assessment guide connects task design to job requirements. A better assessment produces evidence the reviewer can inspect and clarify, rather than relying on polish or a final answer alone.

Proctoring may help establish some assessment conditions, but monitoring alone cannot establish understanding or resolve every integrity concern. Intrusive restrictions can also add friction for honest candidates. Evaluate what each monitoring measure contributes and whether a less intrusive process can answer the same job-related question.

A candidate may need assistive technology, a different input method, or an accommodation that changes normal interaction patterns. Home interruptions, unreliable connections, and ordinary pauses can also affect a session. Treating these events as evidence of dishonesty risks confusing environment and access needs with skill.

Before collecting monitoring information, explain what is collected, why it is relevant, who reviews it, and how long it remains. Give candidates a route to discuss accommodations or technical problems before starting. Review security information alongside the employer's own privacy requirements rather than assuming a monitoring feature is appropriate because it exists.

Ask whether reviewers actually use the additional data consistently. Collection that produces vague suspicion without an actionable review standard adds burden without clarifying the hiring decision. Prefer a defined question, such as whether the submitted reasoning holds up under follow-up, over an open-ended search for unusual behavior. Monitoring should support a proportionate process with human judgment and clear limits.

Explain-back questions test whether the candidate understands, verifies, and can adapt the work they submit. Good questions refer to specific choices in that submission and connect them to the role's rubric. They can expose a gap between a polished answer and demonstrated skill, while leaving the source of that gap open to review.

For an analyst task, ask why the candidate excludes a particular record and what changes if its source becomes reliable. For a support response, ask which policy provision supports the proposed resolution and what requires escalation. For code, ask which test exposes the original defect and what edge condition remains unhandled.

Use a comparable follow-up structure across applicants while allowing the content to reference each person's work. For example, ask every candidate to justify a consequential choice, address a limitation, and adapt to a relevant change. Do not make follow-ups harder merely because a reviewer dislikes the original answer.

• Score the substance against explicit skill criteria.
• Allow candidates to acknowledge uncertainty or correct an error.
• Avoid rewarding confident presentation over sound reasoning.
• Provide an accessible format that preserves the intended measurement.

SkillsValidator uses role-specific work samples and explain-back follow-ups to connect the original submission with evidence of understanding. Explore how skills validation works to see how the task, explanation, and rubric fit together. An explain-back check supports skill evaluation; it is not a conclusive authorship test.

Integrity signals indicate where context may be useful, not whether a candidate cheats. Paste events, typing rhythm, time-on-task anomalies, and similarity to known AI output can help a reviewer select an issue to examine. Each signal also has possible explanations unrelated to misconduct, so interpretation requires care.

Pasting may reflect drafting in another editor or using an accessibility workflow. A pause may reflect reading, checking a reference, or handling an interruption. Similar wording can arise from common terminology or a constrained task. Even several signals together require examination of the circumstances and assessment rules.

Set a review sequence before the assessment runs. Inspect the task and permitted-tool policy, compare the submission with its explanation, and ask a focused clarification where needed. Record what is observed separately from what the reviewer infers. Give the candidate a reasonable opportunity to address a concern without disclosing other applicants' information.

In SkillsValidator, integrity signals go to human review and never automatic rejection. The Skills Report includes a score per skill, quoted evidence, the rubric line by line, and recommendation bands of strong, verify, and not shown. Those bands organize demonstrated skill for employer review. They should not be treated as findings of misconduct, and the employer remains responsible for the hiring decision.

A useful candidate policy states what tools are allowed, what must be disclosed, what remains the candidate's responsibility, and how review works. Put it in the invitation and task instructions before candidates invest effort. Keep rules consistent across applicants for the same role while handling accommodations through an appropriate process.

For a task that permits AI-assisted drafting, wording can be direct. You may use AI to draft or revise your submission. Disclose the assistance you use and explain which parts it affects. Verify the claims, calculations, or code you submit. Do not enter confidential information into external tools. You are responsible for understanding the final work.

State the follow-up conditions separately. During the explain-back step, answer from your own understanding without AI-generated responses. You may refer to your submission and the supplied materials. If you need an accommodation or clarification about permitted tools, contact the assessment team before starting. Publish an actual contact route alongside the policy.

Explain consequences without assuming intent. Concerns about tool use receive human review under the stated rules, and candidates have an opportunity to clarify relevant circumstances. Follow-ups assess understanding of the submitted work. Avoid claiming that a system reliably identifies every instance of AI use.

If independent performance is essential, replace the drafting permission with a clearly stated restriction and its job-related reason. Do not change the rule after submission. Have HR, accessibility stakeholders, and counsel review the policy as appropriate for your process.

A fair assessment process requires more than good questions. Reviewers need shared criteria, a way to resolve uncertainty, and a documented distinction between skill evaluation and integrity review. Employers remain responsible for job-related selection practices, appropriate accommodations, and the use of assessment results in hiring.

Before launching, have reviewers evaluate the same fictional submissions independently. Compare their reasoning against the rubric. If one reviewer rewards polished prose while another rewards verification, clarify the criteria. For engineering tasks, a coding assessment platform should help reviewers inspect correctness, testing, and tradeoffs separately.

Keep candidate clarification focused on observable issues. A useful question asks why a recommendation follows from the supplied evidence. An unhelpful question asks whether the candidate seems like someone who uses AI. Review equivalent concerns consistently and document why a decision follows from the assessed requirements.

Monitor candidate questions, accommodation handling, disagreement between reviewers, and potential adverse impact with qualified support and counsel where appropriate. Cognitive tests require particular care around validity and adverse impact review. This guidance is not legal advice or a compliance guarantee. Revisit the workflow when roles, tools, or task conditions change so the assessment continues to measure the work you need.