The honest starting point
It is often impossible to tell for certain whether a piece of text or an image was made by an AI tool. Clues exist, and they can raise or lower your suspicion, but none is proof. The tools improve, people edit their output, and ordinary human work can look “machine-like”. Treat everything below as prompts for closer checking, not as a verdict.
Clues in images
Image generators leave certain kinds of flaws, explained in the article on image generators. Things worth a second look:
- Hands and fingers. Extra, merged or oddly bent fingers have been a classic sign, though newer tools make it less common.
- Text and logos. Signs, labels and book covers may show lettering that looks almost right but is not real words.
- Inconsistent details. Mismatched earrings, glasses that melt into the face, a collar that changes pattern, or teeth that blur together.
- Backgrounds that do not make sense. Railings that stop, windows of different sizes, crowds of faces that look unfinished.
- Reflections and shadows that disagree with the light source.
- A glossy, overly smooth look. Skin without pores, perfect symmetry, or a dreamy finish applied to everything.
- Context. Does the picture have a credible source, date and location, or did it simply appear in a feed?
Real photos can show several of these too. Heavy editing, filters, compression and low resolution can all smooth skin or blur details. Equally, a generated image can be fixed by a person in an editing program until none of these signs remain.
Clues in writing
Text from chat tools often has a recognisable feel, though it is a tendency and not a rule.
- Even, polished prose with little personality, and a similar sentence rhythm throughout.
- Generic statements that sound sensible but say little, with few specific names, dates, places or first-hand details.
- Neat structure such as an introduction, a balanced list and a tidy summary, even for a casual question.
- Stock phrases and hedging repeated across paragraphs.
- Confident errors, such as invented quotations, references that do not exist or details that contradict each other. These are linked to the behaviour described in the article on AI hallucinations.
- Odd consistency. No typos at all, or an unnatural lack of any opinion.
None of this is definitive. Many people write in a formal, tidy style, especially in a second language or in professional settings. A human editor can also tidy machine-written text until it reads naturally, and a user can ask a tool to write casually, with mistakes.
Why detection tools are unreliable
Websites and programs claim to measure whether text was written by AI. Their results should be treated with great caution.
They guess from statistical patterns. A detector looks at features such as how predictable the wording is. But predictable, plain writing is also what many humans produce, and a careful, simple style can be flagged as machine-made.
They make both kinds of errors. They can label human writing as AI (a false positive) and AI writing as human (a false negative). A false positive can have serious personal consequences for a student or employee accused wrongly. Some providers of detection software have themselves acknowledged limits and have withdrawn or qualified their tools.
Edits and paraphrasing defeat them. Light rewriting of generated text can change a detector’s result.
Short texts are harder. There is little pattern to measure in a few sentences.
Writers of non-native English may be flagged more often, according to some published concerns. Because the evidence is still developing, the safe conclusion is only that error rates may not be even across different kinds of writers.
The tools keep changing. A detector tuned to one generation of models may do poorly on the next.
Because of all this, a detector score on its own is not good evidence. A percentage figure looks precise but should not be read as the probability that someone cheated.
For images, similar problems apply. Some tools and watermarks can flag content made by particular generators, and some cameras and editing programs attach information about how a file was made. These signals can be removed when a file is screenshotted, resized or re-uploaded, so a missing label does not show that a picture is real.
Better habits than relying on a score
- Check the source. Who posted this, when, and where did they get it? An original publisher or an identifiable person is more informative than a style judgement.
- Look for corroboration. For a dramatic photo or claim, see whether reliable outlets or independent witnesses report the same event.
- Search for the image’s history. A reverse image search can show whether a picture appeared earlier, in another context.
- Verify the facts in a text. Look up quotations, figures and references. A text with checkable, correct specifics is more trustworthy regardless of who or what wrote it.
- Ask the author. In school or work, a conversation about how someone produced a piece, their notes and drafts, often says more than a tool does.
- Slow down when emotion is high. Content designed to provoke outrage or urgency deserves extra checking. This is also a common feature of online scams, including those that use voice cloning or fake pictures to build trust.
Where it matters most
The stakes are highest when someone’s reputation, grades or money are involved. Accusing a person of using AI on the strength of a detector alone is risky and, in many cases, unfair. Organisations that need to address the issue generally do better with clear written policies on acceptable use, drafts and process evidence, and a fair route to respond.
Remember also that using an AI tool is not automatically wrong. Whether it is acceptable depends on the rules of the setting, and these differ widely. A more useful question than “was this made by AI?” is often “is this accurate, and was it made in a way that fits the rules here?”