Spotting drift in your own writing

Recommended for: Courses where students write argumentatively or reflectively; best placed midway through a writing assignment, once a draft exists

Objective: Students can point to the sentences in their own text that still carry a position after an AI pass and the ones that do not, and can decide with reasons which edits to undo.

Time needed: 45 minutes · Group size: Works from 12 to 150 students; the comparison happens in pairs

Background for the instructor

Most AI literacy is about error: is this true, did the model invent it, is the source real. This activity is about something else, harder to see because nothing goes wrong. A language model is trained on fluency, and fluent prose is average prose. Ask it to improve your text and you often get a version that is more correct and says less.

Illingworth calls this drift: the language keeps moving, but intention has stopped leading. His example is sharp enough to use in class. "The policy fails because it ignores the lived experience of part-time faculty, who teach 60% of undergraduate courses but have no say in curriculum decisions" becomes, after an edit, "The policy presents challenges as it may not fully account for all stakeholder perspectives." Nothing in it is untrue. There is simply nobody standing behind it any more.

Students tend to recognise this only when they see it in their own work. Hence a paragraph of theirs rather than an example of yours: the point is the difference between "this is better written" and "this is still what I meant".

Procedure

1. Ask students to bring a paragraph of 150 to 250 words from work in progress. Not an invented text - it has to be something they take a position in.
2. They have an AI tool "clarify and sharpen" the paragraph in a single pass. No further instruction: the lazy prompt is exactly what you want to examine.
3. Both versions side by side, sentence by sentence. One question per sentence: is my position still in here? Green if it is, amber if it has been softened, red if it has gone.
4. Swap in pairs. The partner reads only the edited version and writes one sentence saying what the author thinks. Does that match what the author meant?
5. Rewrite by hand only the red and amber sentences. Do not undo everything - some edits were simply better.
6. Close with two sentences in the written work itself: what the AI version did better, and what it took away.

Discussion questions

• Could your partner tell from the edited version what you think? If not, which sentence would have been needed?
• Which edits did you keep, and why were those genuine improvements?
• What made the softened sentences attractive? Politeness, caution, or simply that they sounded more professional?
• Reading this again in six months, which version would sound like you?

Assessment opportunity

Have students hand in both versions and the closing note alongside the finished work. Do not mark which version is better, but whether the student can point to what changed and account for it. That is assessable, and it is not something AI can do on the student’s behalf.

Extension options

Repeat at the end of the course with a new text and set the two reflections side by side: are there fewer red sentences, or has the student only got faster at spotting them? For language teaching there is a caveat: students writing in a second language sometimes lean on AI for phrasing, and rightly so. For them the question is not "did you edit" but "which edits touched the substance".

Reference

• Based on Sam Illingworth, Keep Your Voice: A Slow AI Guide to Writing with AI and Staying Human (2026), theslowai.substack.com. The notion of drift and the part-time faculty example come from it.
• The guide is written for writers rather than teachers, and carries no open licence. This activity - its shape, its steps and its questions - is ours; what we took is the idea, not the wording.