The Detector and the Fourteen Essays
St Theresa's College, a sixth form in Malta, has referred fourteen of its 120 end-of-year essays to the examinations appeals panel after a commercial detector reported each as 'likely AI-generated', with confidence scores between 78% and 94%. The vendor's technical note states a 3% false-positive rate on English-language samples and warns that the rate is 'substantially higher' for writers whose first language is not English. Eleven of the fourteen flagged students are Maltese–English bilinguals; one has a documented Specific Learning Difficulties profile on file. The college's plagiarism regulations place the burden on the student to 'demonstrate authorship', and the panel must decide all appeals within ten days. For nine students the only surviving evidence is a shared cloud document with a version history: three timestamped revisions over six hours, the earliest containing an outline and two paragraphs. The vendor will not disclose its detection method, citing trade secrets.
1. Using the figures given, assess whether the panel would be justified in concluding that a flagged student probably used AI. Show your reasoning and state explicitly which assumptions your answer depends on.
Read the worked answer
Justification is not truth. A belief is justified when it is proportionate to the evidence available to the believer, so the panel could be justified and still wrong. The question is therefore: what does the evidence make probable?
The detector output is a likelihood, not a posterior. Suppose 10% of the cohort actually used AI, and the detector catches 90% of such essays while falsely flagging 3% of genuine ones. In 1,000 essays: 100 AI-written, 90 flagged; 900 genuine, 27 flagged. So P(AI | flagged) = 90/117 ≈ 77%. Now vary the prior: at 4% prevalence, 36/(36 + 28.8) ≈ 56%; at 1% prevalence, 9/(9 + 29.7) ≈ 23%. The posterior is not determined by the detector alone - it depends on a base rate the panel does not know and cannot responsibly guess.
The vendor's own warning matters more than the headline rate. If the false-positive rate is higher for bilingual writers, the likelihood ratio for exactly the students actually flagged is weaker than 0.90/0.03. The '92% confidence' figure is a property of the classifier's output, not the probability that a student cheated - conflating the two is the central error here.
Threshold should track consequence. A misconduct finding is an action with serious costs, so the evidential bar should rise with the gravity of the outcome. On these facts the panel is entitled to investigate; it is not entitled to conclude.
2. Is there an epistemic injustice in the college's procedure, and if so of what kind?
Read the worked answer
Testimonial injustice. Miranda Fricker's concept fits closely: a speaker suffers testimonial injustice when their credibility is deflated by a feature that is not truth-conducive. Here the deflating feature is being a bilingual or dyslexic writer - precisely the group the vendor flags as unreliable evidence - and the student's own account of drafting is discounted against a machine's.
Credibility excess in the other direction. Fricker's framework is symmetrical: the detector receives a credibility it has not earned, because it is opaque, expensive and 'scientific'. The panel cannot cross-examine an algorithm whose method is protected as a trade secret; opacity converts a statistical guess into an unchallengeable witness.
Structural rather than individual injustice. The burden-shifting rule means the student must disprove an accusation whose basis she cannot inspect. That is not a failure of any one teacher's bias but a defect in the procedure's design, and it produces a systematic pattern: the same demographic is disadvantaged every time. Note that eleven of fourteen flags landing on bilinguals is itself evidence that the instrument is measuring style, not cheating.
A likely objection. One might say a student can always produce drafts. But where the draft history is thin and the accusation unfalsifiable in principle, the 'opportunity to answer' is formal, not real.
3. Recommend one procedural reform and identify the philosophical principle that justifies it. What could be said against your reform?
Read the worked answer
Reform: shift the burden back and require corroboration before a finding. No misconduct finding on detector output alone; a flag triggers human blind review, disclosure of the method to the panel (under confidentiality if necessary), and a second independent line of evidence - a viva, a supervised rewrite, or the version history.
Justification. Two principles converge. First, the presumption of innocence functions as an epistemic device, not merely a courtesy: because error is asymmetric - a false finding damages a student irreparably, a missed case costs the institution little - the burden belongs with the party that can bear it and can inspect its own evidence. Second, Popperian fallibility: an accusation that cannot in principle be checked is not a finding but an assertion, and institutions may not act on unfalsifiable claims when serious interests are at stake.
The Bayesian point reinforces this. A threshold set at, say, 95% posterior probability cannot be met from a 3% false-positive instrument with an unknown prior, so the reform is not softness but calibration.
Against the reform. One can argue that a two-source rule lets sophisticated cheaters through, since AI use leaves little trace, and that the institution owes honest students a credible deterrent. That is a genuine cost, but it is a cost in detection, not a licence to convict on unreliable evidence - and the deterrent value of false findings is negative, since it teaches students that the process is arbitrary.