Norming and bias in online tests when candidates use AI along the way
ISE's Student Recruitment Report 2026 shows that 92 percent of UK employers suspected or confirmed misconduct in selection, most often unauthorised AI use in online assessments and interviews. The article looks at what this means for norm groups and fairness in testing, what employers are doing about it, and which questions to ask the vendor.
What the new survey shows
The Institute of Student Employers, ISE, published its Student Recruitment Report 2026 on 14 October 2026. The report draws on 155 UK employers that between them received more than 1.7 million applications and made 31,577 hires in the 2025/26 recruitment cycle. ISE describes it as the largest survey of student employers in the UK.
The figure that has drawn the most attention concerns misconduct. 92 percent of employers suspected or confirmed candidate misconduct in assessment and selection, up from 85 percent in 2025 and 78 percent in 2024. The share of employers with no concerns at all fell from 22 to 8 percent. The most common form of misconduct was unauthorised use of AI during interviews and online assessments.
The report also shows record competition, with 141 applications per graduate role and 96 per apprenticeship. Only 16 percent of employers use AI in their own recruitment, and 52 percent have no plans to do so.
Why this undermines the norm group
A test score only means something when it can be compared with something else. That something is the norm group, the people who took the test earlier under known conditions. A vendor's documentation normally describes who they were, when they were tested and how.
This is where the problem starts. If part of today's applicant pool completes the items with help from a language model while the norm group completed them unaided, two different situations are being compared. The score looks normal on screen, but it no longer measures the same thing. This is an interpretation of ISE's figures rather than a conclusion the report draws itself, but it follows from the fact that the most common form of misconduct takes place in online assessments.
The consequence is not that the test becomes useless. The consequence is that percentiles and cut-off scores set on an older norm base can give a distorted picture of how strong a candidate really is relative to the field.
Bias in a new disguise
When people talk about bias in testing, most think of differences between demographic groups. The ISE report points to a different kind of unfairness. If some candidates follow the rules and others do not, they compete on unequal terms in the same pool. The candidate who completes the test alone is effectively penalised for doing the right thing.
Employers are not against AI as such. 64 percent now accept candidates using AI in the hiring process, up from 46 percent a year earlier. That acceptance mainly covers CVs, cover letters and application forms. During online assessments, interviews and assessment centres, AI is usually not allowed. The line is drawn exactly where the comparability of the test is at stake.
What employers are doing about it
The report shows a clear shift from policing to guidance. 82 percent of employers now give candidates guidance on acceptable and unacceptable AI use, up from 55 percent in 2025. 77 percent are reviewing or redesigning their selection process.
The specific measures ISE mentions are more personal questions, time limits and randomised questions. To cope with volume, some employers also use automated eligibility screening, shorter application windows, early closure of campaigns or a cap on the number of roles a candidate can apply for.
One figure deserves attention from anyone considering game-based assessments. According to ISE, only 10 percent of game-based assessments are fully AI-driven, and that is the area where AI has advanced furthest.
Questions to ask the vendor
ISE's figures are British and concern graduate recruitment, but the mechanism is the same in Denmark as soon as a test is taken unsupervised. That raises some concrete questions to ask before buying or renewing a test.
When was the norm group collected, and under what conditions? Are there newer norms from the period in which language models have been freely available? Does the test draw from an item bank so that two candidates do not see the same items? What time limits apply, and are they justified in the documentation? And does the vendor address how validity is affected if a candidate receives help along the way?
The answers do not only determine whether the test is good. They determine whether the numbers in the test report can be read the way they were intended.
Kilde: https://ise.org.uk/knowledge/insights/615/graduate_jobs_stabilise_but_ai_use_disrupts_recruitment/