A little clarity for every image

AI Photo Checker

Check a photo for AI source signals with free metadata and OpenAI provenance checks. Review the evidence and choose optional paid AI and deepfake analysis.

Start with free image source checks, then choose whether to add a paid model analysis. Source evidence and a model prediction answer different questions.

Check evidence before interpreting a score

  1. Choose the original image. A screenshot or re-export may have different metadata and pixels.
  2. Review the five local source checks and the free OpenAI source check when applicable. Metadata-only mode does not send the image to a detection provider.
  3. Read execution status and evidence details. Choose a paid analysis only if you want a model score and accept the displayed scope, credit cost and data transfer.

Five local evidence checks

The five groups are Content Credentials, IPTC/XMP source fields, AIGC labels, generation parameters and ordinary EXIF. Generation parameters include both Automatic1111 and ComfyUI records. A record may describe the file’s claimed origin without proving the identity of the writer. Ordinary camera EXIF is context, not proof that no AI was involved.

Found evidence opens by default. Expand the other checks to distinguish no matching field, unsupported data, a failed check and a check that did not run. GPS and prompt details remain hidden unless you choose to reveal or export them.

Compare the three kinds of AI image checks

Start with the closest available original. These checks answer different questions; a model score does not replace a credential validation.

CheckWhat it reportsProcessing and price
Local source evidenceC2PA, source labels, generation parameters and ordinary EXIF; validation status stays separateIn this browser; free
OpenAI provenanceSupported C2PA and SynthID outcomes for OpenAI-generated contentImage sent to OpenAI; free to you
Hive model analysisAI-generation and applicable face-swap scores; predictions, not proofImage sent to Hive after explicit confirmation; paid when available

Real example reports: read the evidence, not a verdict

These are recorded integration examples from September 15, 2026, not results for your selected image or an accuracy benchmark. Original and transformed files can receive different responses.

Recorded exampleObserved responseInterpretation
Newly generated OpenAI original PNGOpenAI: detected SynthID and trusted C2PAA positive supported-source result for this particular file
Sample 06: NASA astronaut; sample 07: JPEG Q50 derivativeHive AI score: 0.531209 on the original; 0.001501 on the derivativeCompression changed the score; neither score proves authorship
Sample 04: CAI Firefly tabby catOpenAI: Adobe issuer, valid C2PA, OpenAI outcome not_detected; local validation valid but untrustedAn Adobe credential is not evidence of OpenAI generation; validity and trust are separate

Read the probability with its status

A percentage appears only for a completed model response that contains a usable score. Not calculated means no score was produced; Unavailable means the analysis could not produce a usable result. Neither means zero probability.

Hive scores remain separate from provenance evidence. A low score is not a certificate that a photograph is real, and a high score is not conclusive proof against a person. Compression, editing and unfamiliar image types can affect predictions. Use original sources and context for important decisions.

Frequently asked questions

Is the free check a universal AI detector?

No. It reads supported provenance records and OpenAI signals. Missing signals do not rule out AI generation.

Will the free check spend my credits?

It does not automatically start paid model analysis. Paid work requires a separate confirmation of scope and cost.

How accurate is the AI photo detector?

No representative accuracy benchmark is claimed here. Hive scores can change with image type and compression. Use source records and publication context alongside the result.

Does deleting metadata make an image undetectable?

No. A pixel-based detector or supported watermark check can work independently of ordinary metadata.

Can I check a screenshot?

Yes, if it is a supported static image, but its metadata describes the screenshot and may omit the original provenance. A missing signal does not establish a human origin.

Which files and source fields are supported?

Use static JPG, PNG or WebP up to 20,000,000 bytes, 20 megapixels and 8192 pixels per side. Local checks inspect supported C2PA, IPTC/XMP, AIGC and generation-workflow records; unsupported or unreadable fields are reported separately.