Checking If an Image Is AI-Generated: Hands, Text, Reflections, and Metadata Limits

Stop before you share. That's the first instruction. The image in front of you looks perfect, and the unease you feel is the signal. Generative models have reached the point where nobody can reliably spot fakes by eye alone. The working method is not to stare harder at the picture. Check it for a handful of visual tells that still trip up the models. Check its file for hidden metadata. Then get off the image entirely and read laterally about where it came from. No single method is enough. Any tool or guide that claims otherwise is wrong.

Hand Anomalies: The Tell That Is Shrinking Fast

Start with the hands. For three years they have been the single most reliable visual giveaway. Generative models learn from millions of photos, but hands are small, they move constantly, and they are almost never the subject of a training caption. The models never built a stable internal rule for how many fingers a hand should have. The result, in anything but the simplest poses, is extra digits, fused fingers, or a thumb that bends backward at a joint that does not exist. Look at the edges of the hand, not the center. A real photo has clean silhouettes. An AI hand has a lumpy outline where a finger merges into its neighbour, or a small nub where a sixth digit is trying to grow. Check the joints for unnatural angles. The model will happily draw a thumb that folds flat against the palm in a way no human thumb can manage.

Here is the catch. Midjourney v6, released in late 2023, improved hands dramatically over v5. Newer models like DALL-E 3 and Stable Diffusion XL have closed the gap further. A hand in a simple position, resting on a table or holding a cup, can pass a casual look. The errors cluster in complex poses: crossed fingers, a hand in mid-wave. Treat hands as your first test, not your only test. Clean hands tell you nothing on their own; you are looking at a newer model. The failure case is the image that has been through a filter or a re-compression pass. A slight blur hides the hand errors entirely. When the image has been compressed to death on a messaging app, do not trust your eyes on the hands. Find the original file or move to a different check.

Text Garbling: The Most Reliable Tell Left

When the hands fail you, look for text inside the image. Garbled text is the closest thing to a permanent flaw in synthetic images. The reason is structural: language models and image generators are separate systems. Even the best current models treat letters as shapes to be painted rather than symbols with meaning. Text that is not the central focus, a street sign, a poster in the background, a logo on a t-shirt, a headline on a newspaper, comes out as near-nonsense. You will see characters that are almost right. A mix of Latin script with a stray Cyrillic letter. A word spelled correctly but in a font that has no business being there. The most common failure is garbled letterforms: an 'r' that looks like an 'n', an 'a' that has lost its bowl, two letters fused into a single blob that is neither one thing nor the other.

This tell is shrinking more slowly than hands. Midjourney output still includes text failure in the v6 generation, but the failures moved from every word to only long or complex words. Newer fine-tuned models can render a short word like 'EXIT' or 'COFEE' correctly if the prompt demanded it. The trick is to look at secondary text, not the main subject. A prompt asking for 'a coffee shop' will nail the word on the cup. The sign on the back wall says 'EATERY' or 'ESPARESSO' or nothing at all. Check the fine print on posters, the labels on bottles, the engraving on a plaque. If the text is in a script you cannot read, an image set in Bangkok with Thai signage, you cannot eyeball this. Zoom in and compare the shapes against a known sample of that script. When the text is in English and it looks wrong, you have your answer. When the text is in a language you do not read, you have no answer from this test alone. Move to the metadata check or the reflection check.

Reflection Errors and Lighting Inconsistencies

Reflections are the third visual check. They are subtle enough that most people skip them. The rule is simple: light bounces, and it bounces in ways that follow physics. A reflection in a mirror, a window, or a pair of glasses must match the scene it is reflecting. Same light source direction. Same colour temperature. Same relative position. AI models generate the reflection from a statistical guess, not from a physical simulation. The guess is often wrong in ways the eye registers as 'off' without you knowing why. The classic failure is a reflection that does not line up. A window behind the subject shows a street. The reflection in the subject's glasses shows a room. The light source in the reflection comes from the left. The shadows on the subject's face come from the right. The reflection is too sharp, or too blurry, or it contains an object that is not in the scene at all.

Synthetic output from DALL-E and Midjourney both suffer this, though the specific errors differ. Look for symmetry tells too. They cluster with reflection errors: earrings that should match but do not, a collar lapel that folds one way on the left and the other way on the right, buttons misaligned down the front of a shirt. These are not lighting failures. They are the same class of error: the model is generating the shape of a thing without understanding its symmetry. Newer models are fixing the easiest cases. A single portrait with one light source will have consistent enough shadow direction to pass. The failure case is the image with multiple light sources. That is where the model's statistical guess has the most freedom to be wrong. A scene with a window, a lamp, and a screen all casting light, and the shadows do not obviously agree. That image deserves a second look.

Backgrounds, Geometry, and the Impossible Architecture

After hands, text, and reflections, check the background. That is where the model's understanding of physics breaks down most visibly. Generative models are excellent at the subject and notoriously sloppy at the edges. The background is where you find warped straight lines, repeating texture patterns, and architecture that cannot exist. A brick wall that should be straight lines will have a section where the mortar lines curve and merge. A tiled floor will have tiles that shrink or grow as they approach the edge of the frame. A building behind the subject will have windows that are not square, or a roofline that bends at an angle no structural engineer would accept. These are not artifacts you need a trained eye to see. They are visible in a side-by-side with a real photo. They become obvious the longer you look.

The deeper issue: the model is not trying to draw a real building. It is drawing 'a building' from statistical memory. The parts are plausible. The whole is impossible. The background check works best on images with architecture. It fails on images with no background at all: a tight headshot, a macro shot of an object, a product photo on a solid colour. If the background is empty or a simple gradient, skip this check and rely on the others. If the background has any complexity, look for the specific failure modes. The repeating texture that is too regular. The straight line that is not quite straight. The perspective that seems correct at the centre of the image but goes wrong at the edges. When you find one of these, you have your answer. When you do not, you have learned nothing. A simple background gives the model nothing to fail on.

AI Image Metadata Verification: What the File Could Tell You

When the visuals are not enough, the file itself can be evidence. Checking an image's EXIF data means looking at the format's hidden fields for tags that reveal the generator. The most useful fields: the 'generator' or 'software' tag, which may contain the literal string 'Midjourney', 'DALL-E', 'Stable Diffusion', or 'Adobe Firefly'; the EXIF 'UserComment' field, which some tools use to write 'AI-generated'; and, for PNG files, the tEXt chunk named 'parameters', a signature of Stable Diffusion outputs that contains the prompt and settings. Adobe Firefly embeds C2PA content credentials by default since its 2023 launch. Open AI added C2PA metadata to DALL-E 3 outputs in 2024. If you have the original file and the right tool, the provenance can be unambiguous.

Here is the limit. Never rely on metadata alone. Social media platforms strip EXIF data as a matter of course. What's App, Telegram, X, Instragram, and Facebook all re-encode images. That removes most metadata in the process. The image you receive on a forwarded message is almost certainly a re-encoded version. Its hidden fields are gone. The absence of metadata proves nothing. Absence is the normal state of any image that has moved through a messaging app. Even when metadata is present, it can be forged. A determined actor can copy an innocent photo's EXIF data onto a synthetic image, or strip all metadata from a real photo and re-save it. The C2PA standard is the most robust answer. It uses cryptographic signatures that are tamper-evident. Its adoption is still limited to a few generation platforms and almost no social media sites. The practical move: download the original file, check the metadata fields listed above, and treat a hit as strong evidence. Treat a miss as no evidence at all.

Reverse Image Search and Lateral Reading: The Only Method That Scales

The visual checks and the metadata checks are the start. They are not the end. The most reliable way to tell if an image is synthetic is to stop looking at the image and start looking at where it came from. Media literacy professionals call this lateral reading. Instead of evaluating the image in front of you, open a new tab and check the source. The first move is a reverse image search using Google Lens, Tineye, Yandex Images, or Bing Visual Search. These tools find where the image has appeared before online. They are better than they used to be. Google Lens now indexes some Midjourney and Adobe Stock AI collections. A search may return the original prompt or a page that labels the image as synthetic. If the reverse image search turns up nothing, that is a warning sign on its own. Real photos of newsworthy events tend to be picked up by news agencies. A synthetic image often exists only in the one place you found it.

A reverse image search is not a verdict. It is a starting point. The next step is to search for the claim the image is accompanying. If the image purports to show a Singapore scene, search for the scene plus the word 'fact-check' or 'true' in the local context. In Singapore, that means checking Factually, the government's fact-checking portal that publishes clarifications on false claims circulating locally, or Black Dot Research, an independent Singapore-based fact-checking and media literacy organisation founded in 2022. These groups publish verdicts with visible evidence. A claim they have not touched is a claim you should not share without more work. The failure mode to avoid is tool over-trust. Uploading an image to a single AI-detection tool and treating the result as definitive is a mistake. Independent testing shows these tools have accuracy rates around 60 to 90 percent depending on the image and the dataset. False positive rates reach 5 to 10 percent for even the best, like Hive Moderation. That is high enough that a real photo of a textured surface or a busy street scene can be flagged as synthetic. The tools are a lead, never a verdict. The verdict comes from the lateral reading.

Singapore-Specific Resources and the Legal Backdrop

If you are reading this in Singapore, or verifying an image that claims to show something local, the resources are specific and better than most countries can offer. The Media Literacy Council, a Singapore charity that runs public education campaigns, promotes a framework called S.U.R.E.: Source, Understand, Research, Evaluate. It is embedded in the school curriculum and public libraries. It works for images as well as text. The Singapore Police Force and the Ministry of Health do not run a hotline for 'is this image real', but they do publish advisories. A claim that cites a law enforcement action can be checked against their official website. On the legal side, the relevant framework is POFMA, the Protection from Online Falsehoods and Manipulation Act. It has been operational since October 2019 and allows the government to issue correction directions on false claims. For deepfakes specifically, the Penal Code was amended in 2019 to add Section 377BG, which criminalises the distribution of intimate images without consent, including computer-generated ones. POFMA's Section 7A covers inauthentic digital content that depicts a person saying or doing something they did not. Penalties reach up to SGD 100,000 or 10 years in prison.

The practical takeaway is not that you should read the statutes. The ecosystem has checks in place. Use them. When a forwarded message contains an image and a claim, the first question is not 'is this image real' but 'who published this claim and can I verify it independently?' If the claim names a government body, check that body's official site. If it names a private company, check the company. If it cites a news report, find that report in a second source. The Singapore-specific failure mode to avoid is the recency bias that makes a fresh-looking screenshot of a current event feel automatically true. The screenshot can be fabricated. The event can be real but misattributed. The image can be old and recycled. Black Dot Research and Factually exist specifically to catch these cases. Their verdicts are free and searchable. There is no excuse for being the person who forwards the fake when the check takes ninety seconds.

The Limits of Every Method and the Honest Caveat

The uncomfortable truth: every method on this page has a shelf life, and the shelf is shorter than it has ever been. The hands check is weakening with each model release. The text check fails on short words and non-Latin scripts. The reflection check requires a scene with reflections. The metadata check fails on any image that has been re-encoded, which is most of them. The lateral reading check fails the moment a deepfake is so new that no fact-checker has published a verdict on it yet. None of this means the checks are useless. It means they are probabilistic. Combine at least two before you feel confident. An image that fails the hands check and the text check gives you a high-confidence answer. An image that passes all the visual checks but where lateral reading finds no trace of the scene or the source gives you a low-confidence answer. Treat that image as unverified rather than as real.

The deepest limit is one the tools cannot fix: the human brain's confidence in its own pattern recognition. The AI realism trap is real. It gets worse with every generation of models, because the models get better at the specific things humans notice. The honest caveat is that you will eventually be wrong. The cost of being wrong is not to you. It is to the people who see your shared image and believe it. The practical rule, the one this entire guide reduces to: do not share an image you have not verified. If you cannot verify it, say so. The moment you type 'not sure if this is real' before the forward, you have done the work. The moment you run the image through a single detector and trust it, you have not. The tools are getting better. The skill is not a technology; it is a habit. Build the habit. The question of whether an image is synthetic becomes a question you can answer more often than not. That is the best anyone can honestly offer.

Frequently Asked Questions

Q: Can I reliably spot a synthetic image with the naked eye?
A: Not reliably, and less every year. Your eyes catch common tells like hands and text. Newer models fix those. A visual pass is a lead, not a verdict.

Q: Is there a single AI detection tool that works every time?
A: No. Independent testing shows accuracy rates between 60 and 90 percent and false positive rates of 5 to 10 percent. A single tool's output is never definitive.

Q: If the metadata is gone, does that mean it is not synthetic?
A: No. Social media strips EXIF data on upload. The absence of metadata proves nothing. It is the default state for any shared image.

Q: What is C2PA and why should I care?
A: C2PA is a cryptographic content credential standard adopted by Adobe, Microsoft, Open AI, and Google. It is the most tamper-evident option. Its limited adoption means you will rarely encounter it.

Q: How do I check if an image is real when the text inside it is in a language I cannot read?
A: You cannot eyeball it. Use a reverse image search to find the original and check whether a credible source, in Singapore, Factually or Black Dot Research, has published a verdict.

Q: What is the difference between a deepfake and a cheapfake?
A: A deepfake is AI-generated synthetic media. A cheapfake is a real image or video edited or re-captioned to deceive. The difference is the method, not the intent.

Q: Is it illegal to create or share a synthetic image in Singapore?
A: Not on its own. Distribution can violate POFMA for false claims, the Penal Code for intimate images, or the Elections Act for election-related content. The image is not the crime. The harm it causes is.

Hive Moderation, one of the most widely used commercial detectors, had a false positive rate high enough that a real photo of a textured surface can be flagged as synthetic. The tools are a lead, never a verdict.