Turnitin AI False Positives in STEM: Engineering and CS Writing

If your engineering lab report, computer science write-up, or physics problem set explanation was flagged by Turnitin as AI-generated, you are dealing with a specific technical writing problem, not a cheating problem. The features that make good STEM writing (short declarative sentences, standardized structure, formulaic transitions, precise vocabulary) are the same features that push detector scores upward.

How AI detectors read STEM writing

Turnitin's AI indicator, like most detectors, works from two statistical signals: perplexity (how predictable the next word is) and burstiness (how much sentence length and complexity varies). Human writing typically has higher perplexity and more variation. Language-model output tends toward lower perplexity and flatter sentence rhythm.

STEM writing pushes both signals in the wrong direction. Lab reports follow a rigid Introduction, Methods, Results, Discussion structure. Methods sections are, by convention, terse and formulaic. Results sections describe numbers plainly. Code comments and pseudocode use restricted vocabulary. The prose is, by design, low-variance. For a deeper comparison of how detectors read different genres of student writing, see our breakdown of detector accuracy across STEM and humanities.

What the peer-reviewed research actually shows

The most systematic evaluation of AI detectors to date is Weber-Wulff et al. (2023), published in the International Journal of Educational Integrity. The team tested fourteen detection tools, including Turnitin, and concluded that none met the reliability threshold that would justify using their scores as standalone evidence in academic misconduct proceedings.

The 2023 study did not publish per-discipline breakdowns for STEM writing specifically, and no peer-reviewed dataset that we can cite establishes a precise false-positive rate for engineering lab reports or computer science documentation. What the research does establish is that detector scores correlate with stylistic features (predictability and low variance) rather than with authorship, and that those features are more pronounced in technical writing than in narrative prose.

Turnitin itself has told institutions the AI indicator is not intended as sole evidence of misconduct, and has acknowledged elevated false-positive behavior on documents with less than 20% flagged content. Our post on what Turnitin has publicly said about its false positive rate collects those statements in one place.

Why engineering and computer science writing gets flagged

The specific features of engineering and CS writing that overlap with language-model output are worth naming precisely, because you will need to name them in a response letter.

• Standardized section structure. IEEE, ACM, and lab-manual templates enforce a fixed order of headings and content types. Detectors trained on varied prose read this predictability as machine-like.
• Passive voice in methods sections. "The solution was heated to 80°C" is the convention. Passive constructions with restricted vocabulary produce low perplexity.
• Formulaic transitions. "As shown in Figure 3," "Table 2 summarizes," "The results indicate that": these phrases appear in almost every technical report and appear frequently in LLM training data.
• Narrated pseudocode and algorithm descriptions. Explaining a sorting algorithm or a loop invariant in English produces short, patterned sentences that resemble LLM explanations of the same code.
• Definitions of standard terms. Defining a P-value, a Reynolds number, or Big-O notation produces prose that matches the many textbook definitions LLMs were trained on.

Evidence that clears these cases

Process evidence closes STEM AI-detection cases more reliably than arguments about detector accuracy alone. Reviewers who see the writing being built over time, in ways that a language model cannot produce, tend to withdraw the flag.

For code assignments specifically, Git history is decisive. A commit log with dozens of small commits (adding functions, fixing bugs, refactoring) is not something a language model produces in one pass. If you have been committing throughout the assignment, that log is the strongest single piece of evidence you can present.

Using this in a response to the accusation

A written response to a STEM AI-detection flag should do three things in order. First, request the specific score, the tool version, and any human-review notes in writing. Second, explain why the genre of the assignment produces low-perplexity, low-burstiness prose independent of authorship (citing Weber-Wulff et al., 2023, and Turnitin's own guidance). Third, submit the process evidence: notebook, Git log, version history, raw data.

The procedural rights FAQ covers what you are entitled to request before a hearing. If you are preparing the written response now, NotBot generates a personalized defense package that names the detector, explains the genre-specific false-positive risk in STEM writing, and organizes your Git history, notebook, and version-history evidence into the format reviewers expect. It is drafted in about a minute.

If your case involves suspension, degree revocation, or visa consequences, consult an education law attorney before your hearing in addition to preparing the written response.