Spotting machine writing and machine images

Start with the right question

There are two different questions: “How was this made?” and “Is its claim true?” A human can write something false. An AI-assisted explanation can contain accurate facts. Deciding authorship does not replace checking content.

This course includes a text-clue game and a signal sprint before the competition round. They train you to notice what a passage says and what evidence you would need next. Neither qualifies you to prove that a classmate used AI.

Why a writing-style checklist fails

Neat paragraphs, repeated phrases, formal transitions, unusual punctuation, or a polished tone can appear in both human and generated writing. A person can edit a generated passage; a human writer can use templates or write in a second language. A clue that changes when someone replaces a few words is not dependable proof of origin.

Specific detail is not proof either. A generator can invent a childhood memory, and a human can write abstractly. Uneven texture or a distinctive voice may inform a guess, but no single stylistic cue settles the case.

Work through a text example

Suppose an unsigned post says a school is closing early tomorrow. Its writing is smooth and repetitive. Instead of guessing its author from those features:

1. Find who posted it and whether the account is an official source.
2. Look for a dated school notice supporting the closure.
3. Check whether the post refers to the right school and day.
4. If the claim cannot be verified, describe it as unconfirmed.

Even if the post was written by a person, the closure may be false. Even if AI helped draft it, an official supported notice can still convey a real closure.

Work through an image example

A dramatic photograph appears with a claim about a recent event. Strange hands or warped text can raise suspicion, but genuine photos can contain motion blur, unusual perspective, compression, or editing.

Look for the earliest available publication, the photographer or organisation, and other coverage of the same event. A reverse-image search, where you give a search engine the picture itself instead of words, finds other places the same picture has appeared. It may reveal that a real image is old and has been given a false caption. Some cameras and editing apps attach content credentials to a file: a record of how it was made and edited. That record, and any other evidence of where something came from and how it was changed, is called its provenance. Provenance helps when it is present and trustworthy, but its absence alone does not prove the image was generated.

Use detector outputs carefully

A detector's percentage is an output of that detector, not an automatic probability that a particular student cheated. Its error rates, test conditions, and suitability for the material matter. Do not use it alone to accuse someone or to make a decision with serious consequences.

If the learning task asks for a forced label, choose the best-supported option and keep your uncertainty in mind. In ordinary life, “I cannot determine the origin from this evidence” is often the correct conclusion.

Work through a case where appearance cannot settle it

Imagine receiving a picture of a school courtyard with an unusually straight row of identical flowers. That regularity gives you a question to investigate. It does not give you the answer. The flowers might be real landscaping, copied in an ordinary image editor, or generated.

First separate two questions: “How was this image made?” and “Does it accurately show the courtyard at the claimed time?” An ordinary photograph from another year can mislead without using generative AI. A clearly labelled generated illustration can be appropriate when nobody presents it as documentary evidence.

Look for the original post, its caption, the creator's explanation, and independent views of the location. If those are unavailable, a responsible result is “I cannot establish its origin or the date it depicts.” You can still identify the unsupported claim without identifying the production tool.

Apply the same reasoning to text. A tidy five-paragraph essay may be human-written; a rough draft may be AI-assisted. Discuss the evidence and the claim being made. Do not turn a style impression or detector score into an accusation about a person.

A useful final statement

Compare “This is definitely AI because it sounds polished” with “I could not verify the source, and its event claim conflicts with the school's notice.” The second statement identifies evidence and a meaningful problem. That is the habit to carry beyond the rehearsal.