You read a glossy product description that feels oddly smooth and empty. You see a photo of a crowded event in which one person has too many fingers. A colleague sends a report that sounds confident but cites a source you cannot find. In each case, the same question arises: did a person make this, or did an AI?
This guide explains what you can and cannot tell. You will learn the common signs of AI-generated text and images, why those signs are clues and not proof, how to check claims, sources and context, why AI detection tools should be treated with great caution, and how to respond fairly when you suspect AI use. The aim is not to catch people out but to help you judge what you read and see.
Key takeaways
- No single sign proves that text or an image is AI-generated, so treat patterns as reasons to look closer, not as verdicts.
- Verifying facts, sources and origin is more reliable than guessing from style, and it works whether a human or a machine made the content.
- AI detection tools can be wrong in both directions, so never use one alone to accuse someone or make an important decision.
- Context matters: who published it, where it came from, whether it can be traced to an original and whether it makes sense.
- If you suspect AI use, ask questions and look at evidence fairly, because false accusations cause real harm.
Why spotting AI is hard
It would be convenient if AI content carried an obvious fingerprint. It does not, for several reasons.
The tools keep improving
Tell-tale flaws that were common in the past, such as clumsy sentences or malformed hands, are less common as tools improve. A checklist of signs can go out of date quickly, so use it as a starting point and not a rulebook.
People can sound like AI, and AI can sound like people
Plenty of human writing is formal, repetitive or bland, especially in business settings, and people writing in a second language may produce text that looks unusual to a native reader. Meanwhile, AI-generated text can be edited by a person until it reads naturally. The same is true of images, which can be retouched or combined with real photographs.
Mixed authorship is normal
Much content is partly human and partly AI: a person drafts, a tool suggests edits, a person revises. There is no clean line, which is why the question “was this AI?” often has no simple answer. A more useful question is usually “is this accurate, honest and fit for its purpose?”
Confidence is a trap
People are often sure they can spot AI, but gut feeling is a weak guide. Treat your hunches as hypotheses to test, not as conclusions.
Signs in AI-generated text
These patterns appear often in text produced by language models without careful editing. None is proof on its own. See our explainer on what a large language model is for the background on why these patterns arise.
| Pattern | What it looks like | Why it appears |
|---|---|---|
| Generic, smooth tone | Fluent text that says little and could apply to any subject | Models favour the most typical phrasing |
| Predictable structure | Every section or list follows the same rhythm and length | Models repeat patterns from similar texts |
| Hedging and padding | Phrases that add words without information | Safe, neutral wording is common in training data |
| Missing specifics | No real names, places, dates, numbers or personal detail | The model has no lived experience to draw on |
| Tidy conclusions | A neat summary that repeats earlier points | A common essay pattern |
| Confident errors | Wrong facts stated calmly, sometimes with invented sources | Models generate plausible text, not verified facts |
| Odd consistency | No typos or quirks, uniform sentence length | Output is generated, not typed and revised |
Invented details and sources
One of the stronger clues is a reference that does not exist: a study, a quotation, a legal case or a statistic that you cannot trace. This happens because models can produce convincing but fabricated details. It is not proof of AI by itself, since people also make mistakes, but it is a serious problem whoever caused it. Our guide to AI hallucinations explains why it happens.
Lack of a point of view
Human writers with real experience usually take a stance, include specific examples and show some personality. Text that circles a topic without committing to anything, or that lists many sides evenly without insight, may be machine-made, or may simply be cautious human writing.
Shifts in voice
If a piece suddenly changes style, perhaps from casual to very formal or from British to American spelling, it may include pasted material from several sources. This can happen with AI, with copied text or with several human authors.
Checks that work better than style
Instead of trying to guess from the writing style, test the content. These checks work regardless of who or what produced it.
- Verify the key claims. Pick the most important facts, figures and quotations, and confirm them in reliable, independent sources.
- Trace the sources. If references are given, look for the actual documents. Check that they exist, that the authors and dates match and that they say what is claimed.
- Look for specifics only an expert would know. Real experience tends to include practical detail, trade-offs and awkward exceptions.
- Ask for the underlying material. If it matters, ask the author for data, notes or an earlier draft.
- Check consistency. Do names, numbers and dates agree throughout? Contradictions are a warning sign.
- Test understanding. Ask the author to explain a point in their own words or to apply it to a new example.
For a fuller routine, see how to check AI output for mistakes, which is just as useful when reviewing someone else’s work as when reviewing your own.
Check the author and publisher
Does the author exist, with a verifiable history? Does the site have a clear owner, contact details and a record of accuracy and corrections? Content from anonymous, newly created sites with no history deserves extra care.
Compare against known work
If you are assessing a colleague’s or student’s writing, compare it with earlier work by the same person. Large unexplained changes in style, vocabulary or depth may be worth a conversation, though not an accusation.
Signs in AI-generated images
As with text, flaws in images are clues, not proof, and they fade as tools improve. They are most useful when several appear together.
| Area | What to look for |
|---|---|
| Hands and limbs | Extra, missing or merged fingers; odd joints; limbs that do not connect logically |
| Faces and teeth | Asymmetry, mismatched earrings or glasses, too-perfect skin, strange teeth |
| Text and signs | Garbled letters, nonsense words, inconsistent fonts on signs, labels and clothing |
| Backgrounds | Warped lines, objects that blend into each other, repeated patterns, impossible architecture |
| Lighting and shadows | Shadows pointing in different directions, reflections that do not match the scene |
| Texture and detail | Overly smooth or waxy surfaces, uneven sharpness, odd hair edges |
| Physical logic | Objects that defy gravity, mismatched scale, cutlery or furniture that makes no sense |
| Crowds | Faces that look similar or distorted, people with odd proportions at the edges |
Zoom in
Many flaws appear only when you enlarge the image. Look at hands, eyes, jewellery, text, edges of objects and the far background. Do not rely on a small thumbnail.
Beware the over-polished look
An image that looks unusually perfect, with dramatic lighting, flawless skin and a staged feel, may be generated, or may be a professional photograph that has been retouched. Again, a clue and not a conclusion.
Real images can be edited and misleading too
A genuine photograph can be misleading through cropping, editing or a false caption. Deception does not need AI. Check the claim attached to an image, not only the image.
Checking where an image came from
For images, origin is often more revealing than appearance.
Reverse image search
Use a reverse image search tool to see where else the picture appears and when it first appeared. If a supposed breaking-news photograph shows up elsewhere under a different caption, or years earlier, that matters.
Look for the original source
Trace the image to the photographer, agency or account that first published it. Reputable sources usually credit their images and can supply the original file, and a lack of any traceable origin is a reason for caution.
Check the metadata and credentials, with care
Image files can contain information about how they were made, and some tools and organisations attach labels or provenance information. Such information can be missing or removed, so its absence proves nothing, and its presence should still be checked.
Cross-check with other evidence
If an image claims to show an event, look for other photos, videos or reports from independent sources showing the same scene. Real events usually leave several traces from different angles.
Ask the sender
If someone sent you the image, ask where they got it. A vague or evasive answer is useful information, and a clear, checkable answer is too.
Audio and video
The same principles apply to voices and video, which are increasingly easy to fake.
Warning signs
- Lip movements that do not quite match the words
- Unnatural blinking, facial movement or lighting on faces
- Audio that sounds flat, has odd breathing or lacks background noise
- Words or accents that seem slightly off for the speaker
- Blurring or warping around the face and hairline
Verify through other channels
If a call or message that sounds like someone you know asks for money, access or sensitive information, hang up and contact them using a number or channel you already trust. Voice cloning makes urgent requests more convincing, so agree a way to confirm identity with family, colleagues and your finance team.
Check the original source of a clip
For a video of a public figure saying something surprising, look for the full original footage from a reliable outlet or the person’s own verified channels. Be wary of short clips, missing context and content shared mainly to provoke anger.
The problem with detection tools
Many online tools claim to tell whether text or an image was made by AI. They are tempting, but they have serious limits.
Why detectors are unreliable
- False positives. They can label human work as AI-generated, which is particularly likely for formal writing, simple language and text by non-native speakers.
- False negatives. They can miss AI content, especially after it has been edited or paraphrased.
- Moving target. Generation tools change, and detectors may not keep up.
- Unclear scores. A percentage or label can look scientific while hiding how uncertain it is.
- Inconsistent results. Different detectors may disagree on the same text, and the same one may change its answer on a second try.
| Use of a detector | Reasonable? | Why |
|---|---|---|
| As one weak signal among several | Sometimes | It may prompt closer checking |
| As proof for an accusation | No | Errors can harm innocent people |
| As the basis for grades, discipline or dismissal | No | The stakes are too high for an unreliable signal |
| To screen large volumes before human review | With caution | Only if a person reviews and the process is fair |
Privacy of what you upload
Pasting someone’s text or image into an online detector shares it with a third party. Check the tool’s terms first, and do not upload confidential or personal material. Our guide to protecting company data when using AI tools explains what to look for.
Provenance labels
Some systems add labels or invisible markers to AI-generated content, and some platforms ask creators to disclose AI use. These can help where present, but they are not universal and can be absent or stripped, so they supplement your own checks and do not replace them.
Reading context and provenance
Often the surrounding facts tell you more than the content itself.
Ask the basic questions
- Who published it? Is the source known, accountable and consistent?
- When and where did it first appear? Can you trace it back?
- What is it trying to do? Inform, sell, persuade, provoke?
- What is the evidence? Are claims supported and checkable?
- Who benefits if I believe it? Strong motives call for extra scrutiny.
- Do independent sources agree? Real facts tend to show up in several reliable places.
Watch for emotional pressure
Content designed to trigger outrage, fear or urgency spreads fast, whether human-made or AI-made. If something makes you want to share it immediately, that is precisely the moment to pause and check.
Be careful with reviews and testimonials
Generic, repetitive reviews with no concrete detail may be fake or generated. Look for specific, varied comments, and for patterns such as many similar reviews posted close together. Treat a mass of identical praise with suspicion.
SEO-style pages
Some websites publish large volumes of thin, generic pages to attract search traffic. Signs include shallow content, no named author, repetitive structure and no original information. If you publish content yourself, our guide to writing blog posts with AI without generic content shows how to avoid becoming one of them.
Responding fairly
If you suspect that a colleague, student, freelancer or supplier has used AI, how you respond matters as much as what you find.
Start with a conversation, not an accusation
Ask neutral questions: how was this produced, what sources did you use, can you walk me through your approach? Many people use AI tools openly and legitimately, and the real issue may be unclear rules, not dishonesty.
Check against the rules that apply
Was AI use actually prohibited, or only unspecified? Does your organisation or institution have a policy? Without clear rules, it is unfair to punish people for breaking them. If you manage a team, our guide to writing an AI use policy for your team will help you set expectations in advance.
Focus on the problems that matter
Often the real concerns are accuracy, originality, undisclosed assistance or breach of confidentiality. Address those directly with evidence: the fabricated source, the copied passage, the factual error.
Document and be proportionate
Keep a record of what you checked and what you found. Respond in proportion to the evidence, and give the person a chance to explain. Decisions with serious consequences, such as academic penalties or dismissal, should follow a formal, fair process and not rest on a hunch or a detector score. This is general information, not legal or employment advice.
Consider the person
People writing in a second language, people with particular writing styles and people who use assistive tools may be wrongly suspected. Be mindful of fairness and avoid assumptions.
If you use AI yourself
The same lessons apply when you are the one producing content.
- Add real substance. Specific examples, first-hand experience and your own judgement make content more useful and more clearly yours.
- Check every claim. You are responsible for what you publish.
- Disclose where it matters. If your audience would care, or rules require it, say how AI was used.
- Keep your process. Notes, drafts and sources show how the work was made.
- Respect others’ rights. Do not pass off generated or copied work as something it is not.
Quick checklist
For text
- Are the key facts and quotes verifiable in independent sources?
- Do the cited sources exist and say what is claimed?
- Does it contain specific, lived detail, or only smooth generalities?
- Is the author real and traceable?
- Have I avoided relying on style or a detector alone?
For images
- Have I zoomed in on hands, faces, text, edges and background?
- Do light, shadows and reflections agree?
- Have I run a reverse image search and looked for the original source?
- Do independent sources confirm the event it claims to show?
- Is the caption honest, whatever the image is?
If I suspect AI use
- Have I checked what rules apply?
- Have I asked neutral questions first?
- Do I have specific evidence of a real problem?
- Have I recorded what I found and kept the response proportionate?
Frequently asked questions
Can I reliably tell if text was written by AI?
Not reliably from style alone. Some patterns are suggestive, but they also appear in human writing, and edited AI text can look entirely natural. Checking facts, sources and context is more dependable.
Are AI detectors accurate?
They can be wrong in both directions and should not be treated as proof. Use them, if at all, as one weak signal, never as the sole basis for an accusation or a serious decision.
What is the fastest way to check an image?
Zoom in on hands, faces, text and background details, then run a reverse image search and look for the original source. Cross-check the claimed event with independent reports.
Does the absence of flaws mean something is real?
No. Modern tools can produce images and text with very few visible flaws. The lack of obvious errors is not evidence of authenticity, so provenance and verification matter more.
Is using AI to write always dishonest?
No. Using AI as a tool can be perfectly legitimate, depending on the rules and expectations in your setting. The problems arise from undisclosed use where disclosure is expected, from inaccuracy and from passing off others’ work.
How do I protect myself from fake voice calls and messages?
Do not act on urgent requests for money or access based on a voice or message alone. Verify using a separate, trusted channel, and agree a confirmation method with people you deal with regularly.
Conclusion
Spotting AI-generated text and images is less about detective tricks than about good habits of verification. Treat stylistic signs as clues, test the claims and sources, trace where content came from, be sceptical of detection tools and of your own certainty, and remember that mistakes and deception are not unique to machines.
When you do suspect AI use, respond fairly, with questions and evidence, not accusations. And if you use AI yourself, add real substance, check your facts and be open when it matters. To build the checking habit, read how to check AI output for mistakes, and to understand why these tools behave as they do, see what is a large language model.
