Vinci AI logoVINCIAI
Back to blog

Image Verification

How to Tell If an Image Is AI Generated

Edited by · Published · Updated

Check metadata, C2PA Content Credentials, lighting and edges to tell if an image is AI generated, plus how to read a photo's EXIF data.

No single method can tell you with certainty whether an image is AI generated. Metadata can be stripped or faked, the visual artifacts below can also come from old cameras and heavy compression, and even dedicated detector tools return false positives and false negatives. Anyone who hands you a confident verdict based on one signal is guessing. A useful check combines several independent signals and treats each one as a vote, not a verdict.

If you only have time for one check, look at the metadata first — it is the fastest signal to read. A file that still carries a coherent camera model, capture time, and GPS coordinates was very likely shot by that camera, and a valid C2PA Content Credential states exactly how the image was produced. Start with the file's metadata, then move to visual inspection when the metadata is missing or ambiguous, which it often is.

Check the metadata first

Metadata is information stored inside the image file itself — invisible in the picture, but readable from the file. Two standards matter for this question. EXIF is the block that cameras and phones write automatically at the moment of capture. C2PA Content Credentials are signed statements that newer cameras, editing tools, and AI generators can attach to record how an image was made or modified. Both can be checked in under a minute, and both tell you something factual rather than impressionistic.

What EXIF says about an AI image

EXIF (Exchangeable Image File Format) records the circumstances of a capture: camera make and model, date and time, shutter speed, aperture, ISO, focal length, and often GPS coordinates. A photo taken on a phone has all of this by default. Files produced by AI generators typically have none of it — the model never stood behind a lens — or they carry fields that describe the generator instead, such as a prompt, a sampler name, or a seed value.

The exception matters: EXIF is fragile. Most social platforms and messaging apps strip it on upload, screenshots never inherit it, and some privacy tools remove it deliberately. A missing EXIF block is therefore weak evidence — it is consistent with AI generation but equally consistent with a screenshot or a re-shared photo. Coherent, complete EXIF on an original file is the far stronger signal, because it is easy to strip and rarely worth faking for an ordinary photo.

What C2PA Content Credentials are

C2PA (Coalition for Content Provenance and Authenticity) is an open standard for signed provenance data, shown to users as Content Credentials. Where EXIF passively describes a capture, Content Credentials actively declare a history: which device captured the image, which tool edited it, and — when the generator supports the standard — that part or all of it was produced by AI. Each entry is cryptographically signed, so tampering with the history breaks the signature chain rather than quietly rewriting it.

You can check a file for free at contentcredentials.org/verify, the official validator. When valid credentials are present, they are the strongest single piece of evidence available: a signed declaration naming the generating model beats any visual guesswork. Their absence, however, proves nothing in either direction — most cameras and platforms do not attach credentials yet, generators are not universally required to, and innocent processing such as platform recompression can strip them.

9 signs an image is AI generated

Visual inspection is a different kind of evidence: probabilistic, resolution-dependent, and improving on both sides as models get better and reviewers learn new tells. Treat each sign below as one vote toward a conclusion, and look for several together before you doubt an image. Keep in mind that low-end cameras, aggressive compression, and ordinary retouching produce some of the same artifacts in perfectly real photos.

1. Generator parameters inside the file

What to look for: open the file's properties or a metadata viewer and check whether the software fields name a generator instead of a camera — strings such as 'Stable Diffusion', 'Midjourney', 'ComfyUI', or 'DALL·E', or explicit prompt, negative prompt, sampler, and seed fields where a camera make should be. Why: many popular generators embed their full recipe into the output by default, and few people turn that off. Exceptions: any of these fields can be stripped with free tools before sharing, and some platforms strip them automatically on upload, so a clean file proves nothing; a real photo edited in unusual software can also carry odd software tags. A generator string is a strong signal when you find it, and nothing at all when you don't.

2. No camera model or GPS

What to look for: in the EXIF block, the Make, Model, and GPS fields. Why: every phone and camera writes these automatically at capture, so a full-resolution original with no camera identity at all was more likely synthesized or scrubbed than shot. Exceptions: images downloaded from social platforms, files sent through messengers, and screenshots all arrive without this data, and privacy-conscious exports remove GPS by design. This sign only means something when you hold the original file — a blank metadata panel on a re-shared image is the norm, not a clue.

3. Hands and finger count

What to look for: count the fingers on every visible hand, then examine the joints — extra, fused, or missing fingers, thumbs on the wrong side, wrists bending in impossible directions. Ears and teeth deserve the same scan. Why: diffusion models learn hands from an enormous variety of poses and angles, and that variety makes hands statistically harder to render than faces, so anatomy errors cluster there. Exceptions: recent models have improved sharply, and motion blur, pockets, or low resolution hide hands in real photos too. A strange-looking hand is a reason to check more carefully, never a verdict on its own.

4. Text and punctuation errors

What to look for: zoom into signs, labels, book spines, and packaging. AI text often shows pseudo-letters that look almost right, two typefaces inside one word, letters that change shape mid-sign, or punctuation in impossible places. Why: models paint letterforms statistically rather than rendering them from a font system, so long or small text degrades first. Exceptions: the newest models render short, prominent text convincingly, and real photos are full of blurry, damaged, or hand-painted inconsistency. The reliable check is consistency — one sign rendered in one coherent style — rather than the mere presence of imperfect text.

5. Inconsistent light direction

What to look for: find the main light source from the strongest shadows, then check that everything else agrees — the shadow the cup casts, the catchlights in both eyes, reflections in windows, glasses, and mirrors, the side each nose shadow falls on. Why: image generators compose regions with partial independence, so two objects in one frame can appear lit from different directions. Exceptions: real scenes genuinely contain multiple light sources — a window plus a ceiling lamp is normal — and flash photography creates its own logic. Light inconsistency is meaningful only when the scene offers no physical explanation for it.

6. Overly smooth textures

What to look for: skin without pores or fine hairs, fabric without weave, wood without grain, foliage rendered as lumpy green blobs, and an overall polished, wax-like finish on people. Why: diffusion models average toward plausible texture, and the upscaling steps many generators run to sharpen output smooth fine detail further. Exceptions: beauty filters and skin-retouching apps do exactly this to real photographs, and heavy compression wipes texture as well. Judge smoothness relative to the rest of the image — one waxen face in an otherwise crisp scene is more telling than uniform softness.

7. Repeated background people and objects

What to look for: scan crowds for the same face appearing twice, the same coat, the same pose; check rows of trees, windows, and parked cars for near-identical copies sharing the same lighting. Why: generators sample patterns statistically, and in busy wide scenes with many small figures the same motif can recur almost verbatim. Exceptions: genuinely repetitive real scenes — marching bands, identical taxis at a rank, plantation rows — look repeated too. The distinguishing feature of AI repetition is exactness: a cloned bystander shares not just clothing but identical features, pose, and lighting.

8. Soft edges and hair

What to look for: zoom to 100% where subject meets background — hair strands against a wall, glasses frames against a face, a chair leg against a patterned floor. AI images often smear there: strands that dissolve into the background, halos around silhouettes, or edges that wobble where a straight one should be. Why: generated fine detail is painted statistically rather than resolved optically, and the upscalers generators rely on soften boundaries further. Exceptions: heavy JPEG compression, aggressive sharpening, and cheap lenses do the same to real photos. Compare edges at the original resolution, not in a compressed preview.

9. Mirror-symmetric artifacts

What to look for: compare the left and right halves of faces and symmetrical scenes — earrings that appear identically on both ears, strands of hair that mirror each other, background blur patterns repeated on both sides. Why: some model architectures and upscalers are biased toward symmetry, so perfectly mirrored detail shows up where a real lens would produce noise. Exceptions: real faces and architecture are already roughly symmetric, so only excessive symmetry counts — identical speckles, mirrored freckles, twin shadows with the same irregular shape. A little symmetry is human; exact symmetry is a flag.

How to read EXIF data

The steps below cover the built-in viewers first, then a tool for the complete picture. Whenever possible, run these checks on the original file — a copy that arrived through a chat app will usually show you nothing.

On iPhone

  1. Open the Photos app and open the photo in full screen.
  2. Tap the ⓘ info button below the image, or simply swipe up on the photo, to open the metadata panel.
  3. Read the panel: it shows the capture date and time, the camera or device — an iPhone model, or an external camera — exposure settings, and a map when location was recorded.
  4. For the complete EXIF block, including fields the panel hides, install a free EXIF viewer app or send the original file to a computer and check it there.

On Android

  1. Open Google Photos and open the photo.
  2. Swipe up on the photo, or tap the three-dot menu, to open the details panel.
  3. Read the panel: camera model, exposure settings, resolution, and a map if location was saved. Samsung Gallery works the same way via the three-dot menu → Details.
  4. For photos stored as files rather than library items, open your Files app and use its Details option, or check the file on a desktop to see the full EXIF block.

On desktop

On Windows, right-click the file and choose Properties, then open the Details tab: it lists camera make and model, date taken, exposure settings, the program that last saved the file, and GPS coordinates when present. On macOS, Finder and Preview expose the same data with a keyboard shortcut. For every field at once — including maker notes and the software tags AI generators leave behind — use ExifTool.

  1. Windows: right-click the file → Properties → Details tab. The Program name field is where generator software often shows up.
  2. macOS: select the file in Finder and press Command-I for Get Info, or open the file in Preview and choose Tools → Show Inspector → EXIF and GPS tabs.
  3. Any system: install ExifTool from exiftool.org and run exiftool image.jpg in a terminal to print the complete metadata block.
  4. Prefer local tools when the photo is sensitive: browser-based metadata viewers upload the file to someone else's server before showing you anything.

What to do if the metadata is gone

By the time most images reach you, the metadata is already gone. Instagram, X, Facebook, WeChat, Weibo, and most messaging and social apps recompress uploads and strip EXIF data — partly to protect privacy, partly to shrink file size. Screenshots start from scratch with none of the original's data. This is normal platform behavior, not tampering.

That leads to two practical rules. First, metadata checks only work on original files: if all you have is a re-shared copy, skip straight to visual inspection. Second — just as important, in the other direction — missing metadata is not evidence that an image is AI generated. A blank metadata panel is exactly what a real photo looks like after a trip through a chat app. When the file has been through platform processing, weigh the visual signs above together with the context: where the image was published, who published it, and whether independent sources show the same scene from other angles. Metadata, when present, can confirm; its absence can never convict.

FAQ

Can I tell from the file extension?

No. The extension — .jpg, .png, .webp — describes the file format and compression, not how the picture was made. AI generators output the same standard formats cameras do. An unusual combination of format and source is worth noting while investigating, but the extension alone never identifies an image as AI generated.

Do screenshots keep metadata?

Not the original's. A screenshot creates a brand-new file that records your screen, the moment you captured it, and your device — while the original photo's camera EXIF and any Content Credentials are discarded. That is why a screenshot shows no camera model: the missing data tells you about the screenshot, nothing about the underlying image's origin.

Is there a free checker?

Yes, several. The official C2PA validator at contentcredentials.org/verify checks Content Credentials for free, and ExifTool is a free, open-source reader for the full metadata block. Free AI detector tools also exist, but treat their scores carefully: detection is statistical, and false positives on real photos and false negatives on AI images are common. No checker — free or paid — delivers a guaranteed verdict.

Can metadata be faked?

Yes. EXIF fields are plain text inside the file, and ordinary editing tools can add a camera model and GPS coordinates to an AI image or scrub them from a real one. C2PA Content Credentials are much harder to fake because every entry is cryptographically signed and tampering breaks the chain — but credentials can be absent, and absence proves nothing. Coherent metadata is one signal to weigh, never proof.

Does removing metadata make an image undetectable?

No. Metadata is only one signal among many. Even with every EXIF field and Content Credential stripped, the pixels remain: lighting inconsistencies, texture problems, anatomy errors, and soft edges are still visible, and detector tools analyze pixels rather than metadata. Removing metadata changes what the file declares about itself, not what the picture looks like.

Sources and further reading

  • Content Credentials Verify

    The official validator run by the C2PA community. Free to use, and it explains what a valid, broken, or missing manifest means for a specific file.

  • ExifTool by Phil Harvey

    The free, open-source standard for reading and writing image metadata on Windows, macOS, and Linux.

  • C2PA specifications

    Specification and background for Content Credentials: what gets signed, what gets recorded, and how tampering is detected.