Generative AI has made it easier than ever for candidates to get outside help during unsupervised online tests, and tools that try to recognise AI-generated content struggle to keep pace. This research takes a different approach: rather than asking which tool a candidate used, it looks at how they behave while taking the test, so it keeps working as new AI tools appear.
The report sets out Podium's four-layer integrity framework of Design, Deter, Detect and Defend, and tests it in five studies: a controlled experiment, and field evidence from more than 106,000 GCAT administrations across multiple countries, industries and clients.
The Integrity Framework
Four layers, each aimed at a different way of getting outside help. The strongest protection comes from running all four together.
Design
Dynamic Item Pools
Guards Against Item sharing, answer-key harvesting and repeat exposure.
Invisible to Candidates
Deter
Assessment Integrity Message
Guards Against Rationalising or normalising cheating.
Shown Before the Test
Detect
Confidence Score
Guards Against Consulting AI tools or other resources on the same device.
Runs Silently
Defend
Facial Validation
Guards Against Proxy test-taking, help from another person and second-device use.
Visible to Candidates
Key Findings
Five studies, from a controlled experiment to more than 106,000 live GCAT administrations.
Study 1
Does the Confidence Score Work?
In a controlled experiment, the Confidence Score told candidates who used outside help apart from those who didn't (.810 when cross-validated; d = 1.47). Being skilled with AI tools did not help candidates evade it.
Controlled experiment, N = 372
.822AUC for telling assisted and independent sessions apart
Study 2
Does Deterrence Messaging Work?
Showing the Assessment Integrity Message before every test significantly improved behavioural integrity (p < .001) and cut the share of flagged sessions by 5 to 7 percentage points.
3,072 matched sessions from 16 clients
+5.2points on the average Confidence Score after the message
Study 3
Does AI Proctoring Work?
Sessions with Facial Validation showed significantly higher behavioural integrity than matched sessions without it (p < .001), making it the strongest single layer of the framework.
2,194 matched sessions in Australia
+4.5points on the average Confidence Score with Facial Validation
Study 4
Does It Operate Fairly?
Confidence Scores were consistent across gender, age and ethnicity groups. Four-fifths-rule checks found no gender- or age-based adverse impact at operationally relevant thresholds; two isolated ethnicity-based cases occurred at single clients with small samples, not a systemic pattern.
Gender, age and ethnicity
0 of 64group comparisons reached an effect size of 0.30
Study 5
How Common Is Cheating?
Roughly 1 in 14 to 1 in 33 unmonitored candidates gain score-altering help. Screening 1,000 candidates without Facial Validation would let 30 to 70 inflated scores into the selection pipeline.
Estimated from field data
3–7%of unmonitored candidates gain score-altering help
Integrity flags inform people. They never make the decision.
Neither the Confidence Score nor Facial Validation is used for automated decision-making. Every flagged session is routed to human review, and any decision to act on an integrity concern always involves a qualified assessor.
Read and Cite the Report
Shaban, T., Rose, M., & Reed, F. (2026). Behavioral, model-agnostic integrity in online ability testing: Evidence from deterrence messaging, behavioral telemetry, and AI proctoring [Technical report]. Podium Systems Limited. https://doi.org/10.17605/OSF.IO/CKNRA