A tool that sounds more certain than it is
Many people reasonably want a way to tell whether a text was written by a person or a machine. A teacher wants to know if an essay is a student’s own work. An editor wants to know if a submission was generated. Software sold as AI detectors, which promises an answer often given as a percentage, is appealing.
The difficulty is that this is a hard problem, and the tools do not solve it reliably. A detector’s output is a guess, and treating that guess as proof can cause real harm.
How detectors typically work
Companies rarely disclose exact methods, but most text detectors fall into a few broad types.
Trained classifiers. A machine learning model is shown many examples of human writing and of machine-generated writing and learns to tell them apart. It then outputs a probability that new text belongs to one group or the other. Its accuracy depends heavily on how similar the new text is to what it was trained on.
Statistical predictability. Language models tend to choose words that are probable in context. Some detectors measure how predictable a text is to a reference model, reasoning that very predictable text is more likely to be machine-made, and text with more surprising word choices is more likely human. The idea is plausible, but it measures a style tendency, not authorship.
Watermarks. In principle, a generator could build a hidden statistical pattern into its output, which a matching checker could then spot. This only works for text from systems that use it, can be weakened by editing or paraphrasing, and does not help with text from other tools.
Provenance records. Attaching information about how a file was made is another approach, but it depends on every step of the chain preserving it, and plain text copied between programs readily loses it.
Why false positives happen
A false positive means human writing flagged as machine-written. Several reasons make this likely.
- Plain, formulaic writing looks predictable. Clear, simple prose, standard structures and common phrasing can resemble what a language model produces.
- Non-native writers can be affected. People writing in a second language often use a narrower range of vocabulary and familiar constructions, which some detectors may misread. Researchers have raised this as a fairness concern.
- Formal genres. Technical reports, legal-style text, and school essays that follow a template can score as machine-like.
- Editing tools. Grammar checkers and rewriting assistants smooth text in ways that may change how it is scored.
- Short samples. A few sentences give a detector very little to work with.
- Famous or widely copied text. Material that appears often online can look highly predictable to a model that has seen it.
Why false negatives happen
A false negative means machine-written text passes as human. This is easy to achieve.
- Light editing. Rewording, reordering or adding personal touches can change the signals a detector relies on.
- Paraphrasing tools. Passing text through another rewriter often disrupts patterns.
- Prompting for style. Asking a chatbot to write casually, with errors or in a particular voice changes its statistical fingerprint.
- New models. Detectors trained on older systems may be less accurate on newer ones, and the field moves faster than the checkers.
- Mixed authorship. A human draft polished by a machine, or a machine draft rewritten by a human, does not fit neatly into either category.
Improvement on one side tends to push the other. A detector tuned to catch more machine text will usually flag more human text, so claims of high accuracy deserve questions about how they were measured, on what texts, and with what error rate on genuine human writing.
The deeper problem
Even a perfect detector would answer only whether the text looks machine-like, not whether the person did anything wrong. Using an assistant to fix grammar, brainstorm or translate is treated differently in different classrooms and workplaces. Machine-written text is not a single category with a clear edge, because the line between assisting and writing is blurry. There is also no reliable independent way to check a score; there is no ground truth in the document itself.
Images, audio and video raise similar challenges. Tools aimed at spotting synthetic media face an arms race with generators, and the same caution applies to claims around voice cloning checks and similar tools.
What this means for different readers
Students. If you are accused on the basis of a detector score, you can ask to see the evidence and explain your process. Keeping drafts, notes, version history and sources can help show how you worked. Follow your institution’s rules about permitted tool use, and when unsure, ask.
Teachers and editors. A score is not evidence by itself. More meaningful approaches include talking with the author about their work, comparing it to earlier writing, asking for drafts, setting assignments that include in-class or process-based components, and being clear in advance about what tool use is allowed. These matter all the more given wider discussions of technology in education.
Readers. A suspicion that something is machine-written is not proof. Checking claims and sources is a better use of attention, since the main risk of generated text is that it may be wrong, not that it is generated, a point relevant to online misinformation.
Using detectors, if at all
If you use one, treat it as a prompt to look more closely, never as a verdict. Test it on text you know is human. Do not rely on a single tool. Avoid consequences, such as penalties, that rest on a score alone. And remember that fluent, plain writing by a real person is entirely capable of looking like a machine’s to a statistical tool.