Checking an AI answer with ACCUSE

Recommended for: Any subject where students use AI to answer questions; best after they have once tried to judge an answer with no structure at all

Objective: Students can run an AI answer from their own field past six criteria in order, and can support each verdict with evidence from the text or from a source.

Time needed: 40 minutes · Group size: 12 to 120 students; checking in groups of three

Background for the instructor

Ask students whether an AI answer is any good and you usually get one verdict on the whole text: "seems right" or "too general". That is an impression, not an assessment. A language model is optimised for fluency rather than truth; it sounds equally certain about a correct claim and an invented source. A fixed order of checks takes the judgement away from the first impression.

Illingworth gathers those checks into the mnemonic ACCUSE. The name is the instruction: treat the text as something that has to answer for itself. The six points, in our wording and turned towards teaching:

• A - Accuracy. Is it true? Does every claim trace back to something checkable, or does it rest on assumption?
• C - Clarity. Does it say what it says? Or does jargon hide the absence of a thought?
• C - Compliance. Does it meet the rules that apply here - disciplinary conventions, privacy, this course’s own agreements?
• U - Usefulness. Does it move someone forward, or restate what the reader already knew?
• S - Strength. Does the reasoning hold? Do the arguments carry the conclusion, or only sound confident?
• E - Efficiency. Does the text do its job without ballast? What can go without loss?

The order is where the value is. Students tend to stop at the first letter - is it true? - while the four that follow often yield more, because AI answers are rarely false and frequently useless, weak or padded.

This activity is the systematic version of something the collection does more loosely elsewhere. Five critical questions before using AI is its counterpart at the front end: that one is about the decision to consult the model, ACCUSE about what comes back. For something lighter first, AI-Generated Summary Check or Summary Variation Analysis each practise one of the six letters without the whole framework. Comparison with Authoritative Sources produces the kind of evidence step 3 below asks for, and Feedback and Grading Analysis turns the same check on an AI answer that is about the student.

Procedure

1. Bring one AI answer of half a page to a page on a real question from your subject. Bringing it yourself works better than having students generate one on the spot: everyone then judges the same text.
2. Groups of three. Each member is assigned two of the six letters, so nobody does all six thinly.
3. Fifteen minutes individually: for each assigned letter, a verdict plus evidence - a sentence from the text, or a source that contradicts the claim. A verdict without a pointer does not count.
4. Ten minutes in the group: combine the six findings into one three-sentence verdict. Where do the letters disagree? A text can be accurate and useless at once.
5. Then one decision per group: usable after editing, or discard? And if editing, which letter goes first.
6. Plenary, ten minutes: which letter yielded most, and which could not be judged without subject knowledge?

Discussion questions

• Which letter stalled because you did not know the material well enough yourself? What does that say about when you can check AI at all?
• Was the text accurate and worthless anywhere? Which two letters collided there?
• What did the confident tone do to your first verdict, before you worked through the letters?
• Would you dare run these six checks over your own work?

Assessment opportunity

Fold one letter into the marking of an assignment where AI is allowed: ask the student to show, with evidence, that they checked their own AI use on that point. It is more concrete than a general declaration and not something a model can fill in. For a test, hand students a prepared AI answer with one built-in fault per letter.

Extension options

Reverse the roles: have students produce an AI answer that passes on five of the six points and deliberately fails on one, and let another group guess which. That takes more understanding than the checking does. In subjects with a professional code, Compliance is the richest letter - fill it in with the code that applies in your field. If your subject works with longer AI outputs, the disciplinary expertise line has activities that apply the same stance to a single genre: an AI-written protocol, a data analysis or a lab notebook. ACCUSE prepares the ground for those.

Reference

• Based on Sam Illingworth, Keep Your Voice: A Slow AI Guide to Writing with AI and Staying Human (2026), theslowai.substack.com. The ACCUSE mnemonic and the six points to check come from it; the questions under each point and the activity itself are ours.
• 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.