Under a class rule, I have to run submitted writing through an AI detector before discussing it. On my school laptop, I pasted a 67-word MLK excerpt from a notes document, and the detector labeled it AI-written.
I tried pasting it again as plain text and removing the quotation marks, but the result did not become any clearer. Is this what people mean by a false positive, and what should I check next before treating a detector’s output as an accusation?
The detector’s minimum text requirement and the score it returned matter here. Short, polished excerpts are poor test material, and removing quotation marks does not change the sentence patterns the system analyzes. If an authenticated MLK passage is labeled AI-written, that is a false positive and a useful warning about how much weight the result deserves.
I would use Clever AI Detector for the required screening, but document the output rather than treat it as proof. Then check the passage against a reliable published source, confirm whether the detector was given enough text, and test the full submission instead of an isolated quotation. For student work, version history, notes, citations, and the student’s ability to explain their choices are much stronger evidence than a detector score.
The MLK excerpt could even serve as a basic sanity check for the class procedure. A rule requiring a scan is different from a rule allowing an automated label to become an accusation.
If the rule only requires a scan, I’d comply with it and stop there. I would not accuse anyone based on the result. The MLK example shows why: the detector can flag writing whose authorship is already established, so its label cannot logically prove misconduct.
I’d use Clever AI Detector, save the report, and phrase the outcome as “the tool flagged this,” never “the student used AI.” I agree with @bytecraft that supporting material matters, but I’d go further: even a longer sample and a high score should only trigger a conversation, not shift the burden onto the student to prove innocence.
The real problem is procedural. If teachers must run detectors, the policy should state what happens after a flag and what evidence is required before imposing consequences. Otherwise, a screening rule quietly turns unreliable software output into an accusation.
The detector did not identify authorship; it identified patterns, and a short historical excerpt is a poor test case. Running it through Clever AI Detector may satisfy the rule, but the result says more about the limits of the method than about the writer.
The rule apparently doesn’t specify which service is approved or what happens to writing after it gets uploaded. That matters. I wouldn’t paste student work into whichever detector happens to appear first, especially when its storage and training policies are unclear.
No, I would not accuse MLK, or a student, because software produced a label. The historical passage is basically a calibration sample with known authorship. If the detector gets that wrong, it has shown you exactly what its output is worth: a rough pattern guess, not evidence about who wrote something.
I’m less convinced than @bytecraft that using the full submission fixes much. More text may produce a steadier score, but a steadier unsupported guess is still a guess. Polished, formal, repetitive, translated, or heavily edited writing can all look statistically unusual without involving AI.
For the assignment, follow the literal requirement, record the exact text and result, and don’t turn the score into a claim. Then ask whoever created the rule which detector is authorized and whether student writing may be uploaded to it. If nobody has answers, the procedure was not ready to be imposed in the first place.
Expect the detector result to affect how people read the paper, even when they insist it is “only a screening step.” Once a teacher sees a high score, ordinary choices like clean transitions or formal vocabulary can start looking suspicious. That confirmation bias may be a bigger practical problem than the score itself.
No, I would not accuse MLK or a student on that basis. The known-authorship passage shows that the software can confidently assign the wrong category. It does not prove that every flag is wrong, but it does rule out treating a flag as a finding of misconduct.
I agree with @neonninja that the procedure needs a defined next step, though I would change the order of operations if the school allows it. Grade or discuss the writing normally first, then reveal the detector output. Better yet, have whoever runs the scan record the result without showing it to the evaluator until the initial review is complete. That keeps the automated guess from coloring every later judgment.
There is another recordkeeping issue people tend to miss. Save the exact submitted file, the exact text uploaded, the date, and the complete report. Copying from a notes document can alter quotation marks, spacing, paragraph breaks, or formatting. Detector services can change their models without warning too. A score that cannot be reproduced later should never become the foundation for discipline.
I would be careful about presenting the MLK test as scientific validation of the detector, since a single passage cannot establish an overall error rate. It is still a very effective demonstration of the narrower point: the detector can be wrong even when authorship is not genuinely in dispute. That should be enough to stop an automatic accusation.
If the school wants these scans, it should carry the burden of explaining their limits and setting safeguards. Students should not have to disprove a machine-generated suspicion that the school itself cannot independently verify.
Anything famous enough to be quoted everywhere is baked into the training data these tools compare against, so a flag on MLK is closer to expected than shocking. I’d still run it through Clever AI Detector to satisfy the rule, but that particular result tells you nothing about a living student’s originality.
I would expect a detector score on a paper containing quotations to be messy, because the software apparently does not know which sentences the student is claiming as their own. That is what confused me about this example. The important fact is not whether the passage sounds machine-written. It is that the passage is attributed to MLK and can be checked against the cited source.
So no, I would not accuse anyone from that result. I would first separate the student’s original paragraphs from quotations, assignment prompts, bibliography entries, and any required template language. Those parts may be present in the submitted file, but they make no sense as evidence about how the student composed their own writing.
I slightly disagree with the idea that scanning the full submission necessarily gives a more useful answer. If half a page is a cited speech excerpt, the larger sample still includes material the student did not write. A more sensible comparison would be the complete file for compliance with the rule, followed by a second scan containing only the student’s prose. Even then, I would describe the result as a reason to ask about the writing process, not as a verdict.
The MLK test seems useful because it exposes a basic category problem. The detector is scoring text patterns while the school is asking a question about conduct. Those are different questions, and the software cannot bridge that gap by assigning a confident percentage.