Is this image AI?
Free online AI image detector: check if an image is AI generated or a real photo in seconds. Six transparent steps, each showing what it checked and what it found. No login required.
Drop, click, or paste an image
JPG · PNG · WEBP · ≤ 20MB · No login · Ctrl+V supported
Upload the original file — platform downloads are compressed and stripped of evidence.
Why this AI image detector is different
The evidence, shown
Most detectors only give a score. We show the evidence: signed strong proof (C2PA, watermarks) always comes first, and every verdict lists its known failure modes.
Generated vs edited
A real photo touched up in Photoshop is not an AI image. C2PA credentials tell “AI-generated” and “contains AI editing” apart, a distinction most free checkers miss.
Free and private
No account needed: a few free checks every day, images auto-deleted after 24 hours. Curious how it works? Read the full detection methodology.
What the six detection steps check
Every upload runs through the same evidence pipeline, ordered from strongest to weakest. Strong evidence short-circuits the run: once a signed marker is found, the remaining steps are marked as skipped and the verdict is issued immediately.
Step 1
C2PA digital signature verification
Some cameras and every major AI tool that joined the Content Credentials coalition (OpenAI, Adobe, Google and others) embed a cryptographically signed provenance manifest into the file itself. We read that manifest, verify the signature, and check whether the signer is on a trust list. A valid signature from a known AI tool is the closest thing to proof this field has: it names the exact tool that made the image. We also tell apart "AI-generated" from "AI-edited", because a real photo with generative fill is a different claim entirely.
Step 2
Invisible watermark decoding
The original open-source Stable Diffusion pipeline embeds an invisible DWT-DCT watermark in every image it saves. We decode that watermark directly from the pixels and compare it bit by bit against the known payload. A decode that matches is strong evidence, and it survives metadata stripping because it lives in the image data itself. Note the honest limit: the most popular local UIs (A1111, Forge, ComfyUI) do not embed this watermark, so a negative result here tells you nothing.
Step 3
EXIF metadata scan
Generated files often carry tell-tale software fields, like "Made with Google AI" or a DALL·E marker left behind by the export path. We read the full EXIF block and match it against a rule table of known generator fingerprints. Metadata hits are strong evidence, though not cryptographic proof: fields can be forged or removed with one command.
Step 4
PNG text chunks
Local generation workflows (Stable Diffusion WebUI, Forge, ComfyUI, NovelAI) write the full generation recipe into PNG text chunks: the prompt, the sampler, the seed, even the entire workflow JSON. If a file still carries those chunks, we can often name the exact pipeline and show you the parameters. This is the highest-hit-rate signal for locally generated images that were saved, not screenshotted.
Step 5
IPTC declarations
The press and stock industry standardized a machine-readable declaration for synthetic media: DigitalSourceType and trainedAlgorithmicMedia. Newer Midjourney exports, OpenAI images and newsroom tooling write these fields. When present, they are a deliberate, standardized "this is algorithmic media" statement, which is why we treat them as strong evidence.
Step 6
Forensic signal analysis
Frequency-domain analysis, noise-residual statistics and compression traces round out the report. These signals never decide a verdict on their own. Instead they explain context: whether the file was heavily recompressed, whether sensor noise that a real photo would carry is missing, and how much confidence any classifier-based judgment deserves on this particular file.
When the six steps find nothing, signed-in users can run one extra check: a vision-model classifier that inspects the image for generation artifacts and returns a probability with its reasoning. It is a second opinion, quota-limited to three per day, and it abstains by design when its score is not decisive.
Is this image AI? How to read the verdict
The report ends in one of three states, and the difference matters more than any score. Tier A, provenance found: the file carries a verifiable marker, a signature, a watermark, or generator metadata, and the report names it. This is the high-confidence case. Tier B, classifier lean: no markers survived, but the vision model leans one way with a probability and its stated reasoning. Treat it as an informed opinion, not proof. Tier C, inconclusive: no usable evidence was found. This is not a verdict that the image is real; it is a statement that the file can no longer be judged responsibly, and the report says exactly that.
When the verdict is Tier C, the page offers two responsible next steps: cross-check the file with the official verifiers (Content Credentials Verify, SynthID Detector, OpenAI Verify), all linked directly in the report, or publish it to the voting square where the community judges it blind.
When people ask: is this image AI?
The question comes up wherever images carry stakes. Before you buy a second-hand item listed with glossy photos, check whether the pictures were generated. When a dramatic image goes viral during a disaster, run it here before sharing. Profile pictures on dating apps and marketplaces, portfolio pieces that look too polished, contest entries where AI submissions are banned, product photos from a new supplier, anything a stranger asks you to trust at face value: all of these are one upload away from an evidence-based answer.
And when the machine is unsure, other people are not. Publish the image to the voting square and the community judges it blind, one vote per person, with the machine verdict revealed only after voting. Your own judgment gets graded too: streaks, agreement rates and a weekly leaderboard keep score.
How to check if an image is AI generated
Pipeline · 6 steps · 1 verdict
- Upload the original file: platform downloads are compressed and stripped of metadata.
- Watch the six detection steps run, each one showing what it checked and what it found.
- Read the three-state verdict: provenance found, classifier lean, or inconclusive (which never means “real”).
- Want a second opinion? Cross-check with official tools (Content Credentials Verify, SynthID Detector), linked right in the report.
01 c2pa · 02 watermark · 03 exif · 04 png · 05 iptc · 06 forensic
Questions people ask before checking
Can you really tell if an image is AI generated?
Sometimes with near certainty, sometimes not at all, and we tell you which situation you are in. When a file carries signed provenance or a decodable watermark, the answer is high-confidence and evidence-backed. When all signals have been stripped away, no detector on earth can give you a trustworthy answer from pixels alone, and we say so instead of guessing.
Is this AI image detector really free?
Yes. Detection runs without an account, and you get free checks every day. Signing in only adds features: publishing images to the voting square, voting on other people's images, and a limited number of vision-model second opinions per day.
Why does the result say "inconclusive"?
Because the honest answer is unknown. Screenshots, messaging apps and social platforms strip metadata and re-compress pixels, destroying most evidence. An inconclusive verdict never means the image is real; it means the file no longer carries enough signals for a responsible judgment. In that case you can publish it to our voting square and let the community weigh in.
What happens to the images I upload?
Private detection images are stored in a private bucket and automatically deleted after 24 hours; the database keeps only a content hash and the verdict. Images you deliberately publish to the voting square stay public until you withdraw them. The full retention rules are on the privacy page.
How accurate are AI image detectors in general?
Worse than most marketing pages admit. An independent 2025 benchmark put 17 detectors at roughly 58% zero-shot accuracy, and a 2026 audit found commercial tools wrongly flagging real news photos about 13% of the time. That is exactly why this site shows its evidence, lists known failure modes next to every verdict, and refuses to output a fake "100%".
What this tool will never claim
No detector, including this one, can certify that an image is real. Absence of evidence is not evidence of authenticity: screenshots and re-encoding erase every active signal. Detection results are probabilistic and must never be the sole basis for a legal, insurance, academic or account-ban decision. What you get here is the full evidence trail, an honest confidence level, and the known failure modes printed next to the verdict. The complete method, with sources and benchmarks, is on the methodology page.