You usually cannot tell by looking, and no single check settles it. The strongest evidence is where an image came from: a signed Content Credential, a traceable original source, or a credible account of who made it. Visual clues and detector scores are weaker and fade as generators improve.
This site checks text, not images, so we have no tool for this question. What follows is how to weigh the evidence that does exist.
Key takeaways
- Provenance beats appearance. Content Credentials, an original source and context carry more weight than anything you spot by eye.
- A missing credential or missing metadata proves nothing. Platforms, screenshots and editing tools remove them.
- Treat old tells, such as odd hands and garbled signs, as hints. Current generators have improved on both, so they are not rules.
- Image detectors give a probability-style signal with real limits. Use them as one input, never as a verdict.
Can you tell by looking at the image?
Less reliably than most people expect. In a 2022 study in PNAS, participants shown faces made by StyleGAN2 and real faces were right 48% of the time, which is chance. After a short training session accuracy rose to 59%, and the synthetic faces were rated 7.7% more trustworthy on average than the real ones (Lancaster University summary of the study). That study used an older generator and faces only, so it does not measure how people do today. It does show that confidence in your own eye is a poor guide.
Looking is still worth doing, because it can raise a flag or lower your confidence in an image. It just cannot clear one. Keep it as the last step, not the first.
What is an evidence ladder for images?
Treat each signal as a rung. Higher rungs can show more, and each one has a limit on what it can prove.
| Signal | What it can show | What it cannot show |
|---|---|---|
| Valid Content Credential (C2PA) naming an AI tool | The file was made or edited with that tool, and nothing has altered it since signing | That the picture is true, or that it was not later re-photographed or restaged |
| Original source found by reverse-image search | Who published first, when, and in what context | That an older copy is not itself generated |
| Credible account from the creator (files, drafts, raw captures) | How the image was made | Anything, if the account cannot be checked |
| Metadata such as EXIF or IPTC fields | Camera, software or an AI label, if present | Anything when fields are absent, stripped or edited |
| Watermark check from the maker (for example Google’s SynthID in Gemini) | That Google’s own tools made or edited it | Anything about images from other generators |
| Independent image detector score | A statistical resemblance to generated images | Why it scored that way, or what happens on new generators |
| Visual inconsistencies | A reason to look harder | Anything on its own, since real photos have flaws and generated ones can be clean |
A “yes” from the first rows is strong. A “no” from any row is not a finding. That asymmetry is the most useful thing to remember.
What do Content Credentials (C2PA) tell you?
Content Credentials are a signed record of an image’s origin and edits, built on the open C2PA standard. A credential bundles statements about the file with a digital signature. Trust rests on who signed it, and altering the file breaks the link to the signature, which makes tampering visible (C2PA specification; C2PA FAQ).
Some generators attach them. Google said in November 2025 that images from its Nano Banana Pro model carry C2PA metadata (Google), and OpenAI’s help center now states that images from ChatGPT, Codex and its API carry both Content Credentials and SynthID watermarks (OpenAI). OpenAI also runs a check at openai.com/verify, which looks only for OpenAI’s own signals. You can inspect a file with the Content Authenticity Initiative’s Verify tool.
Three limits matter:
- Absence proves nothing. C2PA itself says metadata “can be removed inadvertently or intentionally.” The standard does not claim to detect AI content. It records what tools declared, so someone has to add the credential first.
- Presence is not truth. The Content Authenticity Initiative says credentials give information about origin and history and are not meant to say whether content is “real.”
- Not every credential means AI. Google’s help page describes Content Credentials as covering both AI and non-AI content and advises combining them with visual inspection, not treating them as a verdict. Read what the credential actually states.
Some vendors pair credentials with “soft bindings,” such as invisible watermarks or fingerprints, so a stripped credential can be found again. That helps only where a tool and a lookup service support it.
What can metadata and watermark checks show?
Metadata is the information stored inside the file, such as camera model, software name or an AI label. The IPTC standard has terms for generative content, including “trainedAlgorithmicMedia” for media created by a trained AI model and “compositeWithTrainedAlgorithmicMedia” for edits such as inpainting (IPTC digital source type vocabulary). Google Search can show an “AI-generated” label when a publisher self-labels an image with this markup, which depends on the publisher choosing to do it (Search Engine Land).
Metadata is easy to lose. A screenshot, a social-media upload or a format conversion usually removes it, and OpenAI’s help center says Content Credentials can be lost through everyday actions such as a screenshot or a file conversion (OpenAI). So a clean file with no fields is the normal state of most images you meet online. For the full picture of what EXIF and similar fields can and cannot prove, see can metadata detect AI images.
Watermarks live in the pixels instead of the file wrapper. In the Gemini app you can upload an image and ask whether it was made with Google AI, and the app checks for the SynthID watermark. Google says it can currently recognise only content created by Google’s tools, and it may answer “Not enough details to watermark” or “Likely too small an edit” (Gemini help). A positive result is informative. A negative result says only that Google found no Google watermark.
How do you trace where an image came from?
This is the step that most often decides the question, and it needs no special software.
- Run a reverse-image search. Upload the image or paste its address into Google Lens, TinEye or a similar service and look for the earliest appearance. Real news and stock photos usually have many matches, a photographer credit and a date. An image with no history that surfaces first in a viral post deserves caution, though a brand-new genuine photo also has no history.
- Check the account that posted it. Look at its other posts, its age and whether it ever shows the same scene from another angle.
- Look for corroboration. A real event leaves traces: other photos, video, local reports, people who were there. A striking image of a major event that nobody else photographed is a flag.
- Ask for the source file. Photographers can supply raw files, other frames from the shoot and the camera’s metadata. Artists can show layered files or process captures. Those are strong evidence, though a determined faker can fabricate them, so check that they are consistent with each other.
- Read the context. A caption that claims a date, place or event can be tested against the image. A mismatch tells you the claim is wrong even if the image is real.
Google’s “About this image” panel, reachable from image results, Lens or Chrome, adds background on an image, such as an AI label where a publisher has self-labelled it. Treat what it shows as leads to follow, not an answer.
What visual clues are still worth checking?
Visual checks work best when you ask whether the image is consistent with itself and with its claimed origin, not whether it matches a list of known flaws. A short list that still pays off:
- Text and signs. Background lettering, shopfront signs, labels and book spines were long a weak spot. OpenAI said its March 2025 image model renders text more accurately, while acknowledging problems with dense layouts, tiny text and some languages. Garbled text suggests a generated or heavily edited image. Clean text suggests nothing.
- Lighting and shadows. Check that shadows fall in one direction, that light colour matches the setting and that highlights sit where a light source would put them. Inconsistent light can come from a composite, a filter or a flash photo, so it points to manipulation, not necessarily AI.
- Reflections. Mirrors, windows, glasses and water should show what is in front of them. Missing or mismatched reflections are worth a second look.
- Fine detail and repetition. Look at patterns such as tiles, fences, hair, fabric weave and crowds. Smeared, copied or oddly merged detail can be a sign.
- Anatomy and objects. Count fingers, check how limbs meet bodies, and look at jewellery, glasses and straps. Treat this cautiously. Hands and teeth used to be reliable tells, and current generators often get them right, so a flawless hand clears nothing. Likewise, real photos can show odd hands from motion blur or posture.
- Overall finish. An unusually smooth, polished, evenly lit look can fit a generated image. It also fits studio photography, retouching and phone processing.
A tell that was true of last year’s model can be false of this year’s. If a guide gives you a universal rule, such as “AI always gets hands wrong,” it is out of date.
Can an AI image detector settle it?
No, and the evidence on how far to trust one is thin. Independent image detectors return a score or a label from a model trained to separate real from generated images. They do not read provenance and they do not explain their answer.
The most useful public benchmark is Deepfake-Eval-2024, built from content collected from social media and detection-platform users in 2024. The authors found that open-source detectors lost about 45% of their AUC on images compared with earlier benchmarks. Commercial models and models fine-tuned on the new data did better than off-the-shelf open-source ones, but still did not reach the accuracy of human forensic analysts (Deepfake-Eval-2024, arXiv). Real-world images are compressed, resized and edited, and detectors trained on cleaner data struggle with them. For more on how to read any vendor’s accuracy figures, see AI image detector accuracy.
The same logic applies to text detectors: scores come from a classifier, not from knowledge of how something was made, and they can be wrong in both directions. We explain that in can AI detectors be wrong. If you use an image detector, run it only after the provenance steps, check more than one, and read a disagreement as uncertainty, not as a tie to break.
What is a practical checklist?
Work from the top down and stop when you have enough to act.
- Check for Content Credentials. Look at the file in the Verify tool, or at any credential icon a platform shows. A valid credential is strong evidence of what it says. No credential is not evidence either way.
- Get the best copy. Find the largest, least compressed version. Screenshots and reposts throw away information.
- Reverse-search it. Find the earliest appearance, the original publisher and different crops or versions.
- Check the source. Who posted it, what else have they posted, and can they produce the original file or process?
- Test the claim. Do the date, place and event match what other reporting says?
- Check metadata, with low expectations. Camera fields support a photo story. Missing fields support nothing.
- Use the maker’s check when you know the maker. Google’s Gemini check applies only to Google-made content.
- Inspect the image. Text, light, reflections, repeated detail, anatomy. Note anything inconsistent, and remember real photos have flaws too.
- Run an independent detector last. Treat its output as a weak signal and compare more than one.
- State your confidence honestly. “No evidence it is generated,” “likely generated” and “provenance confirmed” are different conclusions. Avoid declaring something fake or real on appearance alone.
What if the caption or text matters too?
Images often travel with text: a caption, a quote or a post. This site’s detector checks text only, and you can run a caption or surrounding article through the AI-generated text detector and review the flagged passages rather than relying on the overall score. For a guide to that side of the question, see how to tell if text is AI-generated.
Frequently asked questions
Does missing metadata mean an image is AI-generated? No. Most social platforms, messaging apps and screenshots remove metadata, so real photos arrive without it constantly. Missing fields are not evidence in either direction. The reverse also holds: an image can carry camera fields that were copied or edited. See the metadata guide for detail.
Are weird hands still a sign of AI? Sometimes, but not as a rule. Earlier generators often produced extra or merged fingers, and current ones frequently get hands right. Correct hands do not show an image is real, and odd hands do not show it is fake, since real photos can distort hands too. Use anatomy as one observation among several.
Can a reverse-image search prove an image is real? It can show where an image appeared first and whether trusted sources published it, which is strong context. It cannot prove authenticity by itself, since a generated image can also be published widely. No matches is a reason for caution, not proof of anything.
Do Content Credentials prove an image is AI-generated? Only when a valid credential says so, and then it records what the signer declared. Credentials also cover non-AI content, and their absence tells you nothing because they are easy to strip. C2PA itself says the standard does not claim to detect AI content.