Engineering Lab Reports and AI Detection: What Is Actually Being Scored

Turnitin measures what it calls qualifying text, which it defines as prose sentences inside a long-form piece of writing. An apparatus list is not that. A numbered procedure is not that. A results table of values and units is not that. Once those come out, what is usually left in a lab report is the theory paragraphs, the procedure if your module asks for it as continuous prose, and the discussion. A percentage on a lab report is therefore mostly a statement about the discussion, which is also the part with the marks attached to it.

That has a blunt consequence for the first thing most people do, which is to go and rewrite the procedure because it "sounds mechanical". Depending on how your module asks for procedures, that text may never have been in the measurement, and it is a record of what you did in the lab. There is a better order to work in.

None of what follows predicts a number. Turnitin publishes the shape of the pipeline, not the thresholds.

The definition sits on Turnitin's report guide, and the sentence before the exclusions is the one people skip:

"Qualifying text (prose sentences contained in long-form writing format) means individual sentences contained in paragraphs that make up a longer piece of written work, such as an essay, a dissertation, or an article, etc. The model does not reliably detect AI-generated text in the form of non-prose, such as poetry, scripts, or code, nor does it detect short-form/unconventional writing such as bullet points, tables, or annotated bibliographies."

Two things in there matter for a lab report. The unit is a sentence sitting in a paragraph, not a line on a page. And the three worked examples of "a longer piece of written work" are an essay, a dissertation and an article. A lab report is not one of those three and Turnitin does not say which side of the line it falls on, because a lab report is not one kind of writing. It is five.

Add that up and the vendor's own summary of the arithmetic reads differently:

"This percentage is not necessarily the percentage of the entire submission. If text within the submission is not considered long-form prose text, it will not be included."

The odd part is the procedure row, because it is not a property of your experiment. Two students can run the same rig on the same afternoon, write up the same steps, and submit two documents with different amounts of qualifying text, purely because one module template asks for numbered steps and the other asks for continuous past tense. The gap between the percentage and the amount of highlighting on the page comes from exactly this, and how to read a Turnitin AI writing report works through it. The general version of the exclusion list is in what content does Turnitin AI skip.

People go looking for a policy page about STEM writing. We went looking too, and wrote down what was there.

On 27 August 2026 we read the rendered text of the three Turnitin pages that describe this report: the AI Writing Report guide, the AI writing detection capabilities FAQ, and the file requirements page. Searching all three:

Both controls came back non-zero, which is how you know the search hit real text and not an empty shell. The three hits for `subject area` are worth reading, because they are the only place a discipline shows up at all, and they are about who is in the training sample: the FAQ says the model was trained across geographies and subject areas, and names "anthropology, geology, sociology, and others" as less common areas included to reduce bias. That describes the training sample. No rule anywhere on those pages says anything about how a given discipline is scored.

So the only sorting rule published is prose against non-prose. Mapping that onto the five headings in a lab report is work you do yourself, and the table in the previous section is our reading of the published rule, not something Turnitin has confirmed about lab reports.

If your procedure section did end up in the qualifying text and it came back solid, the FAQ has a paragraph that is more useful to you than any explanation of the model:

"Sometimes false positives (incorrectly flagging human-written text as AI-generated), can include content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas."

What follows it is the half worth carrying into a conversation with a demonstrator:

"If our indicator shows a higher amount of AI writing in such text, we advise you to take that into consideration when looking at the percentage indicated."

Two of the three items on that list describe things a procedure is meant to do. Repeating a sentence frame for every trial is how you show that four runs were handled identically. Low structural variation is the point of a template. One caution about that paragraph. It is a list of properties false positives "can include", with no frequency attached to any of them. Turnitin does not say how often flagged text has those properties, and neither will I.

Before you change the register of a procedure, open your module handbook and read what it actually asks for. Some ask for third person and past tense in as many words. If yours does, rewriting into a livelier voice is a mark deduction you chose, in exchange for an effect on the score that nobody can promise you.

Two published limits collide in this genre. The file requirements say a submission needs "at least 300 words of prose text in a long-form writing format" before a report is generated at all. And the FAQ says short submissions do not simply get a rougher version of the same treatment:

"In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap. This means that some text that is a mix of AI-generated and original content could be flagged as entirely AI-generated."

A lab report at 1,500 words total is a candidate for that zone once the apparatus list, the results tables, the sample calculations and the reference list have come out of the count. Strip a report down to the discussion alone and pass it through a self-check, which is what people do when they are worried, and you may be measuring 400 words rather than 1,500. Short documents and the all-or-nothing problem in Turnitin covers that behaviour on its own.

There is a second thing hiding underneath. Above 0% and below 20%, an asterisk stands in for the figure and no highlighting is attributed, so from where a student sits a real improvement and no improvement are indistinguishable (what does the asterisk (*%) mean on a Turnitin AI score).

An essay gives you slack. There is an introduction you can restate, a signposting paragraph you can cut, a conclusion nobody grades word by word. A lab report has almost none of that. Nearly every sentence outside the discussion is either a record of what happened or a value.

So it is worth sorting the file once, before touching anything.

That last row is the one worth reading twice. A deviation from the printed protocol is often the most valuable sentence in a lab report, and it is also the sentence most likely to look like an awkward aside worth smoothing out.

The flip side is that the discussion, the one section where rewording is genuinely rewording, is the section carrying the marks and the section where a marker already knows what you sound like. So it gets rewritten carefully and read back line by line against the results table, not skimmed. The general version of that sorting job, for research reports rather than lab reports, is in what you can rephrase in methods and results, and the thesis-scale version of a whole methods section coming back flagged is in Turnitin flagged my whole methodology section.

The other worry in this genre is not the AI number at all. Your aim, your theory and half your procedure may descend from a manual that every student on the module was handed, so thirty submissions contain the same sentences. That is a matching-text question, and it lands somewhere else:

"The Similarity score and the AI writing detection percentage are completely independent and do not influence each other."

Which means rewriting the procedure to move the AI number is aiming at the wrong report, and rewriting it to move the similarity number is a question you should put to your demonstrator before you act on it, because how much of a lab manual you are expected to reproduce is a module decision and not a writing decision. Turnitin AI writing report vs similarity report sets out what each one measures.

The thing that makes a lab report expensive to revise is not the writing. It is that every rewritten paragraph is a paragraph you have to check back against your data, so the cost of a revision is measured in paragraphs touched.

Deciding that scope up front is what HumanPen is for. You upload the document and say what is in it: either you mark the passages, or you import a Turnitin or iThenticate AI report and its flagged passages become the boundary, with the rest of the file preserved as it was. Paragraphs are the unit, so a selection that stops halfway through one gets widened to the full paragraph and drawn for you to confirm before anything starts. Billing follows the words that were actually rewritten. Where the conditions for it are met, passages that come back still flagged can go through again at no charge.

Its own FAQ tells you to review complex documents after download, and that line deserves to be read literally. For a lab report, reviewing means reading every number, unit and uncertainty in the revised text against the results table it came from.

Does the equation in my theory section count towards the AI percentage? Turnitin's published rule is that the model analyses prose sentences inside paragraphs, and it says it does not reliably detect AI writing in non-prose. The paragraphs around a display equation are prose. The equation itself is not the unit being scored. How Turnitin handles math equations and non-prose content goes into that in more detail.

My results table has a paragraph of interpretation in one cell. Is that in? A release note dated 9 August 2023 says the model is now able to process long-form prose text in tables, and adds that existing submissions containing tables must be resubmitted before they are reprocessed. So a paragraph in a cell behaves like a paragraph, which surprises people who moved text into a table for layout reasons.

Should I rewrite the procedure so it sounds less like a template? Read your module handbook first. If it specifies the register, changing it is a mark you have decided to lose in exchange for an outcome nobody can promise. If your procedure is a numbered list, it is also outside the definition of qualifying text as Turnitin publishes it.

My reference list got highlighted on an old report. The same August 2023 release note records that highlighting inside a bibliography was a bug, and that bibliographies are now excluded when the AI writing report is processed. Existing submissions have to be resubmitted before that takes effect, so an old report can still show it.

Can I use a low score to prove I wrote it? No, and the vendor says so first: the model "may not always be accurate (it may misidentify human-written, AI-generated, and AI-paraphrased text), so it should not be used as the sole basis for adverse actions against a student." That cuts in both directions.