To check if a photo or video is real, first look for the original with a reverse image search, then check its provenance with Content Credentials and SynthID, its metadata, the place and the time. AI detectors come last, because they only give leads. Much misleading material is not faked at all: it is real photos and clips taken out of context.
Last checked:
Go in order: the first steps are also the cheapest, and often the first two are enough.
Upload the photo or paste its link into Google Images (Lens), TinEye, Bing and Yandex: they have different indexes, so try several. Look for the oldest appearance and compare the descriptions. If the same photo appears with different stories, or long before the event it is said to show, the problem is the context, not the pixels. On TinEye, the "oldest" sort shows the first time TinEye found the image, not its first appearance online, and the engine finds exact or altered copies, not similar images. On Google, "About this image" shows when similar versions were seen, but it is not available in every region.
Where to find them: Google Images · About this image · TinEye · Bing Images · Yandex Images
Install the InVID-WeVerify plugin, in Chrome or Firefox, paste the clip's link and take the keyframes. Search them as in step 1. Also look for the full recording: a cut can change the meaning without altering a single pixel. The plugin also has a magnifier, metadata and analysis filters; the advanced tools, including the deepfake ones, are reserved for journalists, fact-checkers and researchers.
Where to find them: InVID-WeVerify · InVID-WeVerify (Firefox)
Upload the file closest to the original to Content Credentials Verify. If it has a C2PA signature, you can see who created it and what was changed. Also try SynthID Detector, which looks for the invisible watermark added by Google and the companies that have adopted the technology. A positive result is a strong lead. Its absence proves nothing: screenshots lose the data, metadata can be removed, and SynthID says of itself that it is not a general AI detector.
Where to find them: Content Credentials Verify · SynthID Detector
ExifTool, a command-line program, shows the camera, the date and the editing software, if they are still in the file; many social networks strip them on upload, so their absence does not mean manipulation. Forensically and FotoForensics offer analyses such as ELA, which highlights areas with different compression. Neither gives a verdict: FotoForensics says it shows raw data, and Forensically compares itself to a magnifying glass. Mind what you upload: on Forensically images stay on your computer, on FotoForensics they stay on the server for at least 3 months, and the site says openly that it offers no privacy.
Where to find them: ExifTool · Forensically · FotoForensics
Look for buildings, road signs or terrain you can find again in Google Earth and Street View. The time slider in Google Earth shows imagery from past years, but the dates can be inaccurate. With SunCalc you can compare the shadows in the image with the sun's position at the claimed place and date. Then check whether other independent sources captured the same event.
Where to find them: Google Earth · SunCalc
Hive Detect, Illuminarty, AI or Not and Sightengine estimate whether material is generated. Do not stop at their result. An article published by the Reuters Institute in 2024 describes how, after cropping and scaling, a detector ended up suggesting that generated material was real. The Deepfake-Eval-2024 study, a preprint, shows that open-source detectors perform much worse on real material from 2024 than on academic datasets, and that commercial ones, though better, do not yet reach the accuracy of forensic analysts. Accuracy figures on vendors' websites come from the vendors.
Where to find them: Hive Detect · Illuminarty · AI or Not · Sightengine · Reuters Institute (2024) · Deepfake-Eval-2024
Wrong hands, meaningless text, inconsistent shadows or a mismatch between lips and sound can give away generated material. Their absence does not prove it is real, and the models keep improving. Use them as a reason to keep checking, not as a result.
An authentic photo can support a false claim: it may be old, from another country or wrongly described. That is why the first step is the reverse search, not the detector. Apart from whether the material is authentic, also check what is being claimed about it.
If the material matters and you cannot reach a conclusion, first search for it in Google Fact Check Explorer, then send it to a newsroom: AFP has a form for photos and clips, and our list of tools shows the fact-checker for your country.
Where to find them: Google Fact Check Explorer · AFP · Which fact-checking tools exist in Europe
Justificat.ro is one option, not a replacement for the steps above, and we make it, so we are not neutral. You upload the photo or paste the clip's link: the platform transcribes the clip, checks the claims against sources on the web and reports signs of manipulation or AI generation separately. The authenticity result is a lead, not proof, and Justificat.ro does not currently read Content Credentials, so for C2PA use step 3.
Where to find them: How Justificat.ro works
We read each tool's page and documentation and listed only what they say. We did not test them against each other and we do not give scores. We removed tools recommended in older guides that no longer work, such as Amnesty's YouTube DataViewer and Jeffrey's Image Metadata Viewer. If something is wrong or out of date, tell us.
Where to find them: Write to us
To check if a photo is real, first run a reverse image search in Google Images or TinEye and look for the oldest appearance. Then check its provenance in Content Credentials Verify, its metadata, the place and the time. Use an AI detector only at the end, as a lead rather than as proof.
To check if a video is a deepfake, extract keyframes with InVID-WeVerify, search them in reverse and find the full recording. Look at lip sync and the edges of the face, then try SynthID Detector and an AI detector. No single test is certain on its own; what matters is whether several point the same way.
No, no website can tell for sure if a photo is AI-generated. Detector results can change after cropping or scaling, and SynthID recognises only the watermarks of certain companies. A positive result is a lead, and a negative one does not prove the photo is real. That is why the reverse search and the context matter.
No, missing Content Credentials do not mean a photo is fake. The project's documentation says this explicitly: screenshots lose this data and metadata can be removed. Their presence is a signal about origin, but it does not prove that what the image shows is true.