To check whether an image is AI-generated, start with its provenance: look for Content Credentials (C2PA), an IPTC "trainedAlgorithmicMedia" label or generator metadata, then examine camera EXIF data, run a reverse image search and only then look for visual flaws. Metadata can prove that a tool declared an image as AI-made, but missing metadata proves nothing, because most platforms strip it. Visual clues are becoming less reliable as image generators improve.
- Signed Content Credentials (C2PA) are the strongest evidence available: they record which tool made or edited an image.
- Many generators write an IPTC digital source type of
trainedAlgorithmicMediaor their settings into the file. - Social networks and messaging apps usually remove metadata, so a clean file is not evidence of a real photo.
- Reverse image search often answers the real question faster: where did the picture first appear, and in what context?
- Distorted hands and garbled text are fading tells; their absence says little about a modern image.
Image generators can now produce photorealistic pictures in seconds, and edited or synthetic images circulate alongside real news photos. There is no single test that reliably separates the two. What works is layering several checks, from the most objective (signed metadata) to the most subjective (looking closely), and weighing the context in which the image appeared.
The checks at a glance
| Check | What it can tell you | Limits |
|---|---|---|
| C2PA Content Credentials | A signed record of the tool, camera or edits behind the image | Easily removed; many tools and platforms do not add or keep it |
| IPTC digital source type | Whether the file declares itself as AI-generated or a composite with AI | Unsigned, so it can be added or deleted by anyone |
| Generator metadata | Prompts, model names and settings written by some tools | Only present in files saved directly from those tools |
| Camera EXIF | Camera make, model, lens, exposure and sometimes location | Often stripped; can be edited or copied into a fake |
| Error level analysis | Regions that were saved or compressed differently | Hard to interpret; not designed to detect AI generation |
| Reverse image search | Earlier copies, the original source and its context | New images may have no earlier copies |
| Visual inspection | Anatomical, text, lighting and physical inconsistencies | Newer generators make few obvious mistakes |
Provenance: C2PA Content Credentials
The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard for attaching provenance information to media files. Its user-facing name is Content Credentials. A C2PA manifest is embedded in the file and records claims such as "created with an AI image generator", "captured by this camera" or "edited in this application", along with the ingredients used. The manifest is cryptographically signed, so if the image or the record is altered afterwards, validation fails.
Several AI image services add Content Credentials to their output, as do some editing applications and a growing number of cameras and phones. Some platforms display them; LinkedIn, for example, shows a Content Credentials label on images that carry them. You can inspect a file with the official verification tool at contentcredentials.org, and the specification is published at c2pa.org.
Two limits matter. First, a signature proves who signed the record, not that the scene is true: a credential says what a tool reported. Second, credentials are easy to lose. Taking a screenshot, re-saving in an application that does not support C2PA or uploading to a service that strips metadata removes them. So a valid credential saying "AI-generated" is strong evidence, while no credential at all is the normal state of most images online.
IPTC labels and generator metadata
The IPTC, the body behind the photo metadata standard used by news organisations, defines a digital source type vocabulary describing how an image was made. The values most relevant here are:
trainedAlgorithmicMedia: created by a generative AI model trained on sample content.compositeWithTrainedAlgorithmicMedia: a composite that includes AI-generated elements, such as a photo edited with generative fill.algorithmicMedia: created purely by an algorithm without training data, such as procedural graphics.digitalCapture: captured from a real scene by a digital camera.
Major platforms and generators use these values to label AI content, and the full vocabulary is documented by iptc.org. Unlike C2PA, IPTC fields are not signed, so they are a useful hint rather than proof.
Some tools also write their own metadata. Locally run Stable Diffusion interfaces commonly store the prompt, model and generation settings in PNG text chunks, and node-based tools such as ComfyUI can embed the entire workflow. Finding a prompt and sampler settings inside a file is about as clear a sign of AI generation as you will get, but it only survives if the file was shared without conversion.
Our AI Image Detector reads all of these sources in your browser: C2PA manifests, IPTC digital source type, generator metadata and camera EXIF, and it shows an error level analysis view. It does not use a visual classifier to guess from the pixels, so when an image has no metadata it tells you that rather than inventing a probability.
Invisible watermarks
Some AI providers embed an invisible watermark directly in the pixels, designed to survive cropping, resizing and compression better than metadata does. Google's SynthID, used in images from its Imagen and Gemini models, is the best-known example. These watermarks can only be read with the provider's own detector, which is generally not available as a public tool that works across vendors. Google has begun offering checks for its own watermark through some of its services, but there is no universal reader, and a watermark from one company tells you nothing about images from another. In practice, ordinary users cannot rely on watermarks as a general test.
EXIF data and why social media strips it
Photos from cameras and phones usually contain EXIF metadata: camera make and model, lens, shutter speed, aperture, ISO, date and time, and often GPS coordinates. A consistent set of EXIF values is a point in favour of a real photograph; AI generators typically write none of it.
EXIF is weak evidence in both directions, however:
- It is easily faked. Free tools can edit any field or copy EXIF from a real photo into a generated one.
- It is routinely removed. Most social networks and messaging apps strip metadata on upload, mainly to protect privacy (GPS coordinates can reveal a person's home) and to reduce file size. They also re-compress the image.
So a photo downloaded from social media with no EXIF is perfectly normal. Look for inconsistencies instead, such as a capture date before the claimed event, a phone model that did not exist at that date, or editing software listed for a supposedly untouched photo.
Error level analysis
Error level analysis (ELA) re-saves a JPEG at a known quality level and shows how much each area of the image changes. Areas that have been compressed the same number of times should respond similarly, so a pasted object or retouched patch with a different compression history can stand out as brighter or darker than its surroundings.
ELA is a tool for spotting edits, not AI generation. A fully generated image saved once tends to look uniform, just like an unedited photo. Edges, fine textures and high-contrast areas naturally appear brighter, and repeated re-compression by social media flattens the result. Use ELA to ask "was part of this picture changed?" and interpret it cautiously, ideally alongside the original file.
Reverse image search
Often the most useful step is not analysing the file but finding where it came from. Upload the image or paste its address into Google Lens, Bing Visual Search, TinEye or Yandex, and look for:
- The earliest copy. Sort by date where possible. An image that existed years before the event it supposedly shows has been taken out of context.
- The original uploader. An artist's portfolio, an AI art community or a stock library that labels the image as generated settles the question.
- Higher-quality versions. A larger, less compressed copy may still contain metadata or show details more clearly.
- Coverage by fact-checkers. Viral images are often already examined by news organisations and fact-checking groups.
Crop to the main subject and search again if the full image returns nothing; edited versions often reuse part of an earlier picture.
Visual tells, and why they are disappearing
Earlier image generators made characteristic mistakes, and some are still worth checking:
- Hands and limbs: extra or merged fingers, joints bending the wrong way, arms that do not connect.
- Text: signs, labels and number plates with garbled or almost-real letters.
- Pairs and symmetry: mismatched earrings, asymmetric glasses, buttons and collars that do not align.
- Backgrounds: people blending into each other, architecture that does not follow perspective, objects merging into walls.
- Physics: shadows falling in different directions, reflections in mirrors or eyes that do not match the scene.
- Texture: overly smooth skin, hair that melts into the background, a glossy, evenly lit look.
Current generators have largely fixed hands and can render readable text, and a skilled user can regenerate or retouch any remaining flaws. Real photos also contain oddities from lens distortion, motion blur, compression and smartphone processing. Treat visual flaws as supporting evidence only: a clear impossibility is meaningful, but a clean-looking image is not proof of anything.
Deepfakes and the role of context
Deepfakes, such as face swaps, altered video and cloned voices, raise the stakes because they target real people: politicians, executives in fraud attempts, or private individuals in harassment. The same principles apply, but context usually does more work than pixels:
- Who posted it first, and do they have a track record?
- Is it reported elsewhere? A genuine image of a major event will normally appear from several independent sources and angles.
- Who benefits from it spreading, and is it designed to provoke a strong emotional reaction?
- Does it fit known facts such as the weather, location and timeline?
The opposite risk is real too. Once fakes are common, people can dismiss genuine evidence as "probably AI", an effect researchers call the liar's dividend. Saying "unverified" when you cannot confirm an image either way is more accurate than calling it fake.
For text that accompanies a suspicious image, our guides on spotting AI-written text and AI hallucinations explain how to check captions, quotes and claimed sources.
Frequently asked questions
Is there a tool that can reliably tell if an image is AI-generated?
No tool is reliable on every image. Signed Content Credentials or embedded generator settings can confirm that an image came from an AI tool, but when metadata has been stripped, pixel-based detectors give probabilities that can be wrong, especially on compressed, resized or newer-model images.
How do I check an image for C2PA Content Credentials?
Upload the original file to the verification tool at contentcredentials.org or open it in our AI Image Detector. Use the earliest, highest-quality copy you can find, because screenshots and social media uploads usually remove the credentials.
Does missing EXIF data mean a photo is AI-generated?
No. Most social networks and messaging apps remove EXIF data on upload, and many people strip it deliberately for privacy. Missing EXIF is normal for images found online.
Can AI images be detected by looking at the hands?
Sometimes with older or quickly generated images, but current generators usually draw hands correctly. Look at several features together and give more weight to provenance and reverse image search.
What does trainedAlgorithmicMedia mean?
It is a value in the IPTC digital source type vocabulary that marks media created by a generative AI model trained on sample data. If a file carries it, the tool or person who saved the file declared it as AI-generated.
Can I see Google's SynthID watermark myself?
Not with ordinary image tools. SynthID is invisible and can only be detected with Google's own systems, which check for Google's watermark and not for watermarks or images from other AI providers.