Why this is hard
People want a quick test for whether a paragraph or picture came from a machine. The honest answer is that no simple test is reliable. Generators are designed to imitate human work, they improve over time, and people routinely edit their output. Clues that worked last year may not work now. What follows is a set of hints, not proof.
Clues in text
Machine-written text often has certain tendencies, especially when it has not been edited.
- Smooth but generic. The writing is grammatical and well organized, yet says little that is specific. It may circle a topic with broad statements and few concrete examples, names or personal details.
- Repeated structure. Paragraphs may follow the same pattern, or openings and closings may be formulaic.
- Hedging and balance. Language models often present “on one hand, on the other hand” framings, even where a clear answer exists.
- Confident errors. Invented facts, citations that do not exist, or quotations that cannot be found are a well-known sign, tied to AI hallucinations. Checking a few references is more informative than judging the style.
- Odd consistency. Very even tone and sentence length, with no digressions, slang or quirks of a real voice.
- Missing lived experience. Claims of personal experience with nothing checkable behind them.
None of these is conclusive. Careful human writers can be formulaic, and a person editing machine output can remove most of these tendencies. Many non-native English writers also produce simple, regular prose that can look machine-like to a reader or a tool, which leads to unfair accusations.
Clues in images
Image generators leave their own traces, although they are fading as the tools improve.
- Hands, teeth and ears that have odd proportions or too many or too few parts.
- Garbled text on signs, shirts or book covers, with letter-like shapes that spell nothing.
- Inconsistent details, such as earrings that do not match, glasses that merge into a face, or patterns that change across a garment.
- Impossible physics: shadows or reflections pointing the wrong way, objects that blend into the background, or stairs that lead nowhere.
- A glossy, over-smooth look, particularly on skin, with an unusually perfect finish.
- Background oddities. Crowds with distorted faces, repeating shapes or warped architecture.
How these happen is explained in how image generators work. As systems improve, many of these flaws vanish, and a high-quality image with a careful prompt and some editing may show none of them. Real photographs can also look strange because of lighting, compression or editing, so oddness alone does not prove anything.
Detection software and its limits
Many tools claim to flag AI-written text or images. They typically estimate how likely content is to be machine-made, based on statistical patterns. Experience and published discussions show recurring problems:
- They can produce false positives, labeling real human work as machine-made, with serious consequences for students and employees who are wrongly accused.
- They can produce false negatives, missing machine content that was lightly edited or paraphrased.
- Results change from one tool to the next and from one version of a generator to the next.
- A single score is a probability, not a verdict, and it should not be the sole basis for a punishment or accusation.
For these reasons, many institutions advise treating detector output as one weak signal at most, not as evidence on its own.
Better approaches than guessing
Instead of judging how something looks, ask where it came from.
- Check the source. Who published it, and do they have a track record? Is the claim reported by independent outlets?
- Look for provenance information. Some tools and cameras attach labels or signed records showing how a file was made or edited, and some generators add invisible marks. These are helpful when present, but they can be missing or stripped, so their absence proves nothing. File metadata sometimes holds clues, though it is easy to remove.
- Search for the original. A reverse image search can show whether a picture appeared earlier in another context or has been altered.
- Verify the claims. For text, confirm names, quotations, dates and figures against reliable sources. For images of news events, look for other photos or reports from the same moment.
- Ask the author. For schoolwork or professional writing, a conversation about how the piece was produced, with drafts or notes, is more informative than a score.
- Consider the motive. Content designed to provoke strong emotion, sell something or push people to act fast deserves extra care, whether or not a machine made it. Many online scams now use generated images and text.
Why it matters
Fake or synthetic content can mislead people, impersonate others and make it easier to produce scams or false news at scale. At the same time, over-suspicion has a cost: real photos and real writers can be dismissed as fake. Aim for a middle path of reasonable doubt and active checking.
Disclosure norms are also developing. Many platforms and publishers ask creators to label synthetic content, and being open about your own use of these tools is part of responsible AI use.
A balanced rule
Treat style clues as weak hints, treat detector scores as unreliable, and put your effort into verifying sources and claims. If something matters, such as a news image, a financial request or an accusation against someone, confirm it through independent means before acting or sharing.