You see a photo of a celebrity pushing a crypto scheme, a before-and-after that looks too smooth, or a news image of an event that never happened. Some of it is real. A fast-growing share is generated in seconds, and the tools keep getting better. You do not need to become a forensics expert to protect yourself, but a handful of checks catch most fakes, and knowing where the detection tools stop working matters just as much as using them.
Why these images travel further than the corrections do
A generated image costs nothing to make and nothing to delete, while a correction costs someone time and often embarrassment. Feeds reward whatever stops a scroll, so a striking or infuriating picture gets forwarded long before anyone checks it. The screenshot step does most of the damage, because it drops the label, the file details and the surrounding context in one go, and the receiver has no thread back to the source. The poster rarely has to lie either. Saying they found it somewhere is enough.
Why fake images are suddenly everywhere
Generating a convincing image used to take an artist, a studio and days of work. Now it takes a prompt and a few seconds on ordinary hardware. That collapse in cost is the whole story: when images became nearly free to make, they flooded social feeds, ad networks, dating apps and marketplace listings. The awkward part is that platform labelling is inconsistent, so you cannot count on a “made with AI” badge being present, accurate, or applied to the version you are looking at. A screenshot of a fake loses its label immediately. Practically, this means you should stop judging an image only by how it looks and start judging where it came from, who posted it, and what they gain if you believe it.
The eyeball checks that still catch most fakes
The old checklist was hands, teeth, jewellery and ears. Those still work sometimes, especially in older or weaker generators, but the good models now handle fingers far better than they did, so treat these as a quick first filter rather than proof. What has held up better is anything with structure and repetition. Look at the background: window frames that bow, tiles that drift off-grid, a fence that changes spacing, the pattern on a shirt that mutates halfway across the fabric. Look at text on signs, labels and packaging, which is often the giveaway on fake product photos and flyers. Then look at the people in the background of a group shot, where faces tend to turn soft, waxy or simply wrong the further they are from the main subject. If something feels off, slow down and zoom instead of scrolling past.
Why the tells keep getting narrower
Every tell you learn is one that generators get trained to fix. Hands were the classic giveaway, then they were not. Text on signs was reliable, then it was not. Each round of failures feeds the next round of training, so a checklist learned from last year’s images ages fast. The tells that survive longest are tied to physics and consistency rather than small rendering faults. Drawing one convincing hand is easy now. Keeping light, shadow, perspective and reflection consistent across a crowded scene is much harder. So treat any single tell as a reason to look closer, never as proof.
The physics tells that are harder to fake
Lighting and shadow consistency is the most reliable visual test you can run yourself. Work out where the light is coming from, then check that every shadow and every reflection agrees with that direction. In a real photo, a subject lit from the left casts shadows to the right, and the reflections in eyeglasses and glossy surfaces line up with the same source. In generated images these are frequently mismatched: two shadows pointing in different directions, a highlight on a cheek that has no matching light source, or a reflection in a car window that shows a street that is not in the picture. Reflections in eyes, jewellery, glossy tablet screens and wet surfaces are worth a careful zoom, because they carry a lot of information and generators often blur them.
Trace the image instead of staring at it
Reverse image search is still the highest-value move for most people, and it costs nothing. Run the picture through a reverse search engine as well as a normal web search with the claim typed out in plain words. You are looking for two things: the earliest appearance of the image, and whether the same image appears elsewhere with a completely different caption. Recycled images with swapped context are one of the most common tricks, where a genuine photo from years ago is relabelled as breaking news today. If a photo appears only on one anonymous account and nowhere in established coverage, that absence is itself a signal. On an emergency or disaster image, check whether a wire service, a local news outlet or an official account has published it. If nobody has, wait rather than share.
What metadata and file details can and cannot tell you
Cameras write details into every file they produce, such as the model, the exposure settings and a timestamp. It is tempting to read that as a birth certificate for the picture. It is not one. Social platforms usually strip it on upload, messaging apps often strip it on send, and a screenshot never carried it. Missing file details are normal and say nothing about whether an image is real. When they are present they are easy to copy onto a fake. Read them when you have the original file from the person who took it, and ignore them once an image has been through a feed.
What detector tools can and cannot do
Websites that promise an instant “AI or real” verdict are best treated as a weak hint, never a verdict. They are classifiers trained to spot patterns from a particular generation method, so they misfire in both directions: real photos get flagged, and lightly edited or re-compressed fakes slide through. Editing a fake, cropping it, running it through a screenshot or re-encoding it can strip the very artifacts a detector looks for. Meanwhile, provenance signals that actually mean something, like cryptographically signed capture data or watermarks embedded at generation time, exist but are far from universal, and many platforms strip file metadata on upload. So treat detector output as one signal among several, and never announce “a tool confirmed it” as fact.
When provenance signals actually help
Provenance is a record of where an image came from, carried with the file rather than guessed from its pixels. Content Credentials and watermarks written in at generation time are the two you might meet. When one is intact and checks out, it is the strongest evidence available, because a camera or a tool signed it instead of a model inferring it. Coverage is the catch. Not every camera and not every generator writes these marks, and plenty of publishing steps strip them. A valid signal is evidence. Its absence proves nothing.
When the fake is aimed at your wallet
The images matter because of what they carry. Fake endorsements from celebrities, doctors and business figures are used to sell investment schemes, supplements and apps. Fake listings push renters and buyers into sending deposits for property that is not for sale. Fake profile photos drive romance and job scams, and cloned voices of a relative in trouble are used to rush people into transferring money. A few habits block nearly all of it. Never act on money or credentials in a hurry, whatever the message says about urgency. Verify through a channel you looked up yourself, such as the official app or a number on the company site, not the contact in the message. For family emergency calls, agree a code word in advance and call back on a number you already have. And never pay by gift card, crypto or instant transfer to someone you have not met. Report it to the platform and to your local consumer protection body as well.
Checking an image without accusing anyone
Real photos look odd for ordinary reasons. A phone’s automatic processing, heavy smoothing, a strong filter, low light and a hurried crop all produce the unevenness people read as artificial. Keep the aim narrow, then. You are deciding what to do with your own attention and your own money, not proving a stranger a liar. Ask what you would gain if you were right, and what it would cost the other person if you were wrong. Saying you cannot tell is a complete answer, and a public callout built on a hunch is a real harm.
Living with not knowing
Some images will never be resolved, and that is a normal result rather than a gap in your method. Detection is a contest against an opponent that keeps training, so probably is the honest answer most of the time. Build habits that do not need a verdict. Do not act on money or personal details on the strength of an unverified image. Do not forward a picture you cannot source. When a story rests on one photo, wait for a second outlet. None of that requires you to know how the pixels were made.
Pros and cons
What works well
- Reverse image search and a plain-language search of the claim: free, fast, and it catches recycled and mislabelled photos, which are the most common fakes you will actually meet.
- Zooming in on hands, background patterns, signage text and reflections: catches sloppy and mid-quality generations without any tool at all.
- Signed capture data and embedded watermarks (provenance standards such as Content Credentials and generation watermarks): strong evidence when present, because they are hard to fake.
- Cross-checking the source: a photo published by a news wire or official account with a named photographer carries accountability a random post does not.
What to watch out for
- AI detector websites: unreliable in both directions, easy to defeat with cropping or re-encoding, and not evidence you should quote at someone.
- Hands, teeth and ear checks alone: far weaker than they were a couple of years ago and increasingly unreliable on good generators.
- File metadata: platforms routinely strip it, so missing camera data proves nothing either way.
- Watermarks and labels: not applied by every tool, and usually lost the moment an image is screenshotted or re-uploaded.
Verdict: best for, and who should skip it
Best for: anyone who clicks, shares or buys from social feeds and ads, which is most of us. The reverse-search habit and the money-message rules do more work than any app you can install. Skip it as a debate-winning tool: you can rarely prove a specific image is fake to someone who wants to believe it, so aim to protect your own decisions first. The main trade-off is time, since a careful look takes maybe thirty seconds that scrolling does not.
Questions
Can an AI detector tell me for certain whether a photo is fake?
No. Detectors are probability guesses based on patterns from known generators. They flag real photos and miss edited fakes, and simple steps like cropping, screenshotting or re-saving can defeat them. Use them as a hint, then fall back to source checks and provenance.
The image has no camera metadata. Does that mean it is AI?
Not at all. Most social platforms and messaging apps strip metadata on upload, and screenshots never carry it. Missing EXIF data is normal and tells you almost nothing on its own.
Are videos harder to check than photos?
Usually yes, because you are watching a sequence rather than a single frame. Watch for audio that does not match lip movement, hands and teeth across many frames, hair that blends oddly into the background, and backgrounds that shift when they should not. For calls claiming to be someone you know, hang up and call back on a saved number.
What should I do if I spot a fake ad or profile?
Report it to the platform, warn the person being impersonated if you can reach them through a verified channel, and tell the people who might be targeted. If money has already moved, contact your bank immediately, since speed matters more than anything else for recovery.
Next step: Related reading on Nexgadgets: our plain-English guide to locking down your phone’s privacy settings, and our comparison of what on-device AI versus cloud AI actually means for the photos and messages you handle. Start with the privacy section if you want to tighten things up this week.
Nexgadgets provides general information, not professional advice. Always do your own research.
