How to reduce AI assistance in online assessments

Decide what independent work means

Reducing unauthorized AI assistance starts with a decision about the job, not a detector. Some tasks genuinely require a person to recall information, reason independently or respond without outside help. Others are normally completed with documentation, colleagues and AI tools. An assessment should explain which capability it is trying to observe. Otherwise, a candidate may follow normal workplace habits while the employer interprets that behavior as a rule violation.

Write a short tool policy for each stage. State whether search, calculators, code completion, translation and generative AI are allowed. Explain whether the candidate may take notes and whether those notes can be digital. Give a concrete example when a boundary might be unclear. “You may use the provided reference sheet, but not an external assistant” is more useful than “complete this honestly.” Make the policy available before an invitation becomes a timed activity.

A useful process can include both unaided and assisted tasks. For example, a junior developer might explain a small function without assistance, then use documentation during a separate debugging exercise. The second exercise can assess how they verify a suggested solution. This distinction makes the evidence easier to interpret and avoids pretending that every real workplace task happens in isolation.

Design a task that reveals thinking

A task that can be answered with a generic paragraph is difficult to interpret. Give candidates specific information and ask them to make a decision from it. For a support role, provide a short order history and a return policy. Ask for the next action and a brief explanation. For an analyst, provide a small table with one ambiguous value and ask what needs checking before a conclusion is shared.

The aim is not to make a puzzle unnecessarily obscure. The candidate should have enough information to produce a defensible answer. What makes the task useful is the connection between the evidence and the decision. A strong response identifies relevant facts, acknowledges missing information and explains a proportionate next step. A fluent answer that ignores the supplied facts becomes easier to question without invoking an AI detector.

Randomized item pools can reduce repeated exposure, but randomization does not automatically make versions equivalent. Each version still needs comparable coverage and difficulty. Per-question time limits can constrain outside assistance, but they also introduce speed as part of the measurement. Only use that constraint when it is appropriate, and preserve a process for accommodations. An assessment should not become a test of panic under unexplained pressure.

Treat browser signals as context

A browser can observe some events, such as losing focus or receiving pasted text. It cannot reliably tell you everything happening around the candidate. A second device may be invisible. An overlay may not produce a detectable event. A focus change might come from a password manager, an operating-system message or an accessibility tool. These technical limits matter when explaining the product and interpreting a result.

An integrity report should therefore describe what was observed before suggesting what it might mean. “The assessment lost focus for twelve seconds” is a narrower statement than “the candidate used AI.” The first can be checked and discussed. The second introduces a conclusion that the event alone does not support. Timing patterns and answer similarities also need context rather than automatic accusations.

The proposed HireValid integrity approach combines several signals and keeps decisions with a person. Its website demo uses illustrative weights. It does not represent validated detection performance or monitor your browser activity. Before a production rule is used, the team needs to understand what it measures, how often innocent behavior triggers it and what explanation a reviewer will receive.

Review a concern without making an accusation

Imagine that a candidate pastes a paragraph into a written response. There are several possible explanations: they drafted it elsewhere, used an assistive tool, copied an external answer or misunderstood the rules. Start by checking the instructions they received. If the policy allowed drafting in a separate document, the event is not evidence of breaking that policy. If the rule was unclear, improving the instruction may be the most important action.

A neutral follow-up might ask the candidate to explain their approach and how they prepared the response. A short discussion of the actual content can be more informative than asking whether they cheated. For a technical task, ask why a particular condition was included or what would happen with a different input. Keep the question relevant and comparable with the process used for other candidates.

Record the observed event, the applicable rule, the explanation and the evidence considered. Do not record an unsupported diagnosis of dishonesty. If you decide the evidence remains insufficient, use another proportionate assessment stage rather than inventing certainty. A hiring record should distinguish an unresolved concern from a demonstrated failure to follow a clearly communicated instruction.

Use AI-text likelihood conservatively

A text classifier estimates patterns in writing. It does not observe who wrote the passage or what tools were used. Polished prose, simple wording or an unusual style can have many causes. A classifier output should never be presented as proof that a candidate used a particular system. It is especially risky to turn a weak probability into a confident label shown beside a person’s name.

If an AI-text signal is included at all, define its limited role in the review process. Consider whether it adds useful information beyond a direct discussion of the candidate’s answer. Require a reviewer to see the original task and response. Keep the model or rules version so the meaning of the result can be understood later. Explain to candidates where automated assistance is used in scoring or summarizing.

A separate issue is AI-assisted scoring of the work itself. A rubric-based score can still be wrong even when the candidate followed every rule. Review the rubric, check the rationale and provide a route to correct an error. Detection and scoring are different uses of AI and deserve separate evaluation rather than one broad claim about “AI-powered assessment.”

Keep monitoring proportionate

More collected data is not automatically better evidence. Before enabling camera checks, ask what risk they address, what information is needed and whether a less intrusive method could answer the same question. Candidates should know the setting before beginning and have a clear route to request an appropriate alternative. Privacy, accessibility and legal review belong in the design, not in a footer added later.

Retention also needs a purpose. A team should not keep images indefinitely because storage is cheap. The HireValid brief proposes shorter retention for images than for general assessment records, but the legal schedule is not finalized. A live process needs actual deletion behavior, restricted access and a way to respond to requests. The candidate guidance explains the proposed experience without implying that the marketing demo collects this information.

A practical review checklist

Before your next hiring round, review one assessment from the candidate’s perspective. Can you explain the skill being measured, the allowed tools, the reason for any time limit and what monitoring is enabled? Does each integrity event describe an observation? Can a reviewer investigate it without jumping to a conclusion? Is there a consistent way to ask the candidate for context?

Then test the workflow with ordinary interruptions: a connection drop, a browser notification or a request for extra time. Write down where the instructions or recovery behavior are confusing. The objective is not to manufacture a flawless integrity number. It is to collect useful evidence under understandable conditions, acknowledge uncertainty and make a human decision that can be explained.

Key takeaways

• Define permitted tools for each stage before the candidate starts.
• Use specific, job-relevant tasks that reveal the reasoning behind an answer.
• Treat browser events and AI-text likelihood as limited signals.
• Review concerns in context and provide a fair opportunity to explain.
• Combine independent work, relevant follow-up questions and appropriate privacy safeguards.

Explore a general cognitive practice item, the junior developer bundle or the glossary explanation of an Integrity Score.