What AI Detectors Cannot Tell You (and Why That Matters)

The Score Is Not What You Think

An AI detector returns a number, usually a percentage, and that number feels like a verdict. It is not. The score is a probability estimate, the detector's best guess about whether the text resembles patterns it learned to associate with AI. It does not, and cannot, tell you whether the text was actually written by an AI. The difference between a probability and a fact is the difference between a useful signal and a false accusation.

This distinction matters because detectors are increasingly used to make real decisions about real people: grades, job applications, publication acceptances. When a probability is treated as a fact, honest writers get punished for writing that happens to resemble AI output. Understanding what a detector can and cannot tell you is the only way to use it without causing harm.

What a Detector Does and Does Not Tell You

Questions a Detector Can Never Answer

• Did a human write this and then edit it carefully?
• Did a human write this without any AI at all?
• Which specific sentences are AI versus human?
• Did the writer intend to deceive anyone?
• How much effort or research went into the work?
• Whether the ideas are original to the writer

Reading a Detector Score Honestly

1. 1
   Treat the score as a weak signal, not a verdict

   A score suggests the text sits in an ambiguous zone. It does not prove anything about how the text was produced.

2. 2
   Remember what the score actually measures

   It measures resemblance to learned patterns, not authorship. Resemblance can come from careful editing, formal style, or genre conventions.

3. 3
   Consider what the detector cannot see

   It cannot see drafts, research, editing history, or intent. All the evidence of a genuine writing process is invisible to it.

4. 4
   Demand other evidence before acting

   If a decision hinges on authorship, no single score is enough. Context, drafts, and the writer's ability to discuss the work matter more.

The Authorship Question Detectors Cannot Solve

The fundamental limit of detection is that authorship is a question about process, and detectors only see product. A finished text contains no record of how it was made. The same paragraph could be raw AI output, carefully edited AI output, or a human's first draft that happens to share patterns with AI. The detector sees the paragraph and assigns a score. It cannot distinguish those three origins because the paragraph itself does not contain that information.

This is not a limitation that more training data or better algorithms will solve. It is structural. Authorship is a fact about history, and a piece of text does not preserve its own history. The only reliable evidence of authorship is external: drafts, timestamps, edit logs, source notes. Detectors will keep improving at pattern-matching, but they will never measure provenance, because provenance is not a property of the text.

The Responsible Way to Use Detection

Given these limits, the responsible use of detection is narrow. A detector can be one signal among several, never the sole basis for a decision. It can prompt a closer look, but it cannot substitute for that closer look. When a score leads to an accusation, the accusation must be investigated with evidence the detector cannot provide: the writer's drafts, their sources, their ability to explain their own work.

For writers, the practical takeaway is to keep your process visible. Save your drafts, keep your research notes, and be able to explain how you produced your work. A detector cannot prove you used AI, but you also cannot rely on it to prove you did not. What settles authorship is the evidence of a real writing process. The tools will keep arguing about patterns; the human record of how the work was made is the one thing that actually answers the question. Build that record, and no score can override it.

The Limits of Detection