How do testing programs spot collusion when two candidates answer alike?
Right answers match for innocent reasons
Two candidates from the same class, taught by the same teacher from the same materials, will answer a well taught topic the same correct way. Any method that flags that is flagging education. Collusion analysis therefore leans on the errors.
Wrong answers carry a fingerprint. On a multiple choice item there are several ways to be wrong, and candidates working alone scatter across them. Two scripts that pick the same wrong option on item after item, including items most candidates answer correctly, are showing a pattern chance struggles to explain.
The number is a triage tool
Similarity indices compare an observed match rate against the rate expected from each candidate overall performance. The output is a probability statement about the pair, and it arrives with two honest limits. First, a busy exam produces thousands of pairs, so some pairs will look extreme by luck alone; the more pairs tested, the more false alarms the threshold must allow for. Second, the statistic cannot say who copied whom, or whether a third source fed both.
That is why mature programs set the threshold to generate reviews, not sanctions, and why the review pulls the context the number lacks: were they in the same room or sitting, did their start times overlap, do proctor notes mention either candidate.
Remote sittings add a wrinkle the hall never had: two candidates can be in different cities and still share a source, if both were fed answers from the same helper or the same leaked item set. Similarity there points at a supply problem, a compromised item bank or a paid answering service, and the remedy is an item bank investigation, not two expulsions.
The study group problem
Candidates who prepared together share mnemonics, worked examples, and occasionally the same misunderstanding, which the teacher may have taught them in good faith. Their wrong answers can rhyme without a word passing between them in the exam hall.
A fair process asks that question explicitly before concluding anything, and documents the answer. Programs that skip it end up retracting cases against study partners, and every retraction makes the next genuine case harder to bring. The credibility cost lands on the programme, not on the statistic.
Build the file before you need it
Decide the thresholds, the review steps, and the evidence bundle in the assessment policy while no case is live. A threshold chosen after a suspicious pair is found will always look chosen to convict that pair. Written in advance, it is a standard. The difference matters in an appeal, and appeals are where collusion decisions are truly decided.
Keep reading: How should exam programs handle AI assistants a candidate can run on a second screen?, Remote proctoring flags: reviewing alerts before accusing a candidate, and Exam rescheduling rules: fair cutoffs that reduce no shows cover the nearby cases in this series.
Frequently asked questions
What is answer similarity analysis?
A statistical comparison of two candidates responses that weighs matching wrong answers far more heavily than matching right ones. It estimates how likely the observed pattern is if the candidates worked independently, and flags pairs whose similarity is hard to explain by chance.
Can similarity scores prove collusion by themselves?
No. They cannot show direction, opportunity, or mechanism, and large exams generate extreme pairs by luck alone. Scores open an investigation that needs seating, timing, and proctor evidence before any finding is made.
Why do study partners sometimes get flagged?
Shared preparation creates shared errors: the same mnemonic, the same worked example, sometimes the same misconception from the same lesson. A fair review examines preparation history alongside the score before drawing conclusions.
What threshold should a program use for flagging pairs?
One set in writing before results are processed, calibrated to the number of pairs being compared, and tied to review rather than automatic sanction. Borrowing a threshold from another program without that calibration invites both false accusations and missed cases.