AI Detection Accuracy: STEM vs Humanities Writing Compared

AI detectors do not treat all writing equally. The same tool that clears a lab report can flag a literature essay, and the reason is not that one student cheated. It is that the statistical signals detectors rely on behave differently in STEM writing than they do in the humanities. If you are trying to understand why your paper was flagged when your labmate's was not, this is where to start.

How detectors read writing across disciplines

Most commercial AI detectors, including Turnitin, GPTZero, and Originality.ai, score text using two headline signals: perplexity (how predictable each word is given the words around it) and burstiness (how much sentence length and complexity vary across a passage). Low perplexity plus low burstiness reads, statistically, like a language model.

That framing is discipline-blind. It does not know whether it is scoring a Milton essay or a materials science methods section. But those two things do not produce the same statistical fingerprint, and detectors were largely trained on general-domain internet text. For a broader account of what these tools measure and miss, see the research on AI detector false positives.

Why STEM writing trips detectors

Technical writing is structurally the closest human genre to what large language models produce by default. Lab reports, methods sections, and engineering documentation reward exactly the traits detectors read as "AI":

• Predictable phrasing. Standard constructions like "the samples were prepared according to" or "results indicate that" recur across the discipline because clarity is the priority.
• Low lexical variation. Technical terms have to be repeated exactly. You cannot swap in synonyms for "polymerase" or "eigenvalue" without changing the meaning.
• Uniform sentence rhythm. Passive voice, parallel structure, and short declarative sentences are conventions of the genre, not stylistic quirks.
• Formulaic sections. IMRAD structure (Introduction, Methods, Results, Discussion) produces highly predictable transitions.

The result is that a well-written lab report and a competent AI-generated one look statistically similar. This is not a new problem: engineering and biology students have been reporting flags on genuine work throughout 2024 and 2025.

Why humanities writing trips detectors, differently

Humanities writing gets flagged for a different reason. When a student writes formally, cites carefully, and uses period-appropriate or discipline-specific vocabulary, the prose can register as low-perplexity by detector standards, especially in short passages. History, philosophy, and literature essays flag frequently, and the pattern is documented across multiple institutions we have covered, including Berkeley philosophy essays and Michigan history essays.

The difference from STEM is where the false positive lives. STEM writing is flat because the genre requires it. Humanities writing is flat in patches: a paragraph paraphrasing a secondary source, a careful thesis statement, a formal transition. The rest of the essay may burst with variation, but detectors that score at the sentence or passage level can still surface those patches as "AI-detected content."

What the research actually shows about accuracy across genres

The clearest published finding is from Weber-Wulff et al. (2023) in the International Journal of Educational Integrity, which tested fourteen detection tools across a range of texts and concluded that none performed reliably enough across conditions to serve as institutional evidence. The paper explicitly noted variation by writing style and subject matter as a source of instability.

A separate line of research from Sadasivan and Feizi at the University of Maryland (2023) argued that AI-generated text cannot be reliably detected in principle once light paraphrasing is applied, and that pushing detector sensitivity up to catch more AI text inevitably pushes the false positive rate up as well. In practical terms, that trade-off hits STEM writing harder, because STEM prose already sits closer to the AI baseline.

The Liang et al. (2023) Stanford paper in Patterns is often cited for its findings on non-native English writers, but the underlying mechanism, that clear and structurally simple writing scores as AI, applies to technical writing too.

The UCSB retraction and what it revealed

In one recent case at UC Santa Barbara, a professor publicly retracted an AI cheating accusation after the student produced browser history and dated drafts showing the writing process. The retraction did not turn on which department the student was in. It turned on process evidence that the detector score could not account for. We covered the full case in our writeup on the UCSB retraction pattern.

That case matters for the STEM-versus-humanities question because it shows what actually reverses a detector-based accusation, and it is not a counter-argument about genre. It is documentation.

What this means for your defense

If you are a STEM student flagged for a lab report or methods section, the argument to make is that your writing sits in a genre detectors are known to misread, and that your process evidence (drafts, data notebooks, citation trails, code commit history where relevant) shows authorship. If you are a humanities student flagged for a history or literature essay, the argument is that formal academic prose triggers detectors on stylistic grounds that have nothing to do with authorship, and that your research trail supports the same conclusion.

In both cases, the effective move is the same:

1. Ask in writing which detector was used, the specific score, and whether human review occurred.
2. Preserve version history, browser history, notes, and any prior drafts immediately.
3. Cite the research (Weber-Wulff 2023, Sadasivan and Feizi 2023, Liang 2023) that shows detector accuracy is unstable across text types.
4. Point out that your institution's policy almost certainly requires more than a detector score as evidence.

Our procedural rights FAQ covers what you are entitled to request before a hearing. If you are already past a finding and preparing to appeal, the appeal package is built around the procedural grounds that matter at that stage. NotBot generates a personalized response that names the detector, cites the research that applies to your discipline, and documents the writing process your accusation requires you to defend.