How image checks work
Read the evidence, understand the handling, and keep uncertainty visible.
Step 1
Read source clues
The browser initially inspects up to the first 64 KB for source text, generation parameters and camera metadata. Fields can be edited, removed or located beyond that slice. Dimensions are only context.
Step 2
Use a header fallback when needed
If the local pass cannot decide, the first 64 KB can be sent to our edge function. This is file content and may include metadata or pixels. The function processes it in memory.
Step 3
Try an optional model check
When the metadata is inconclusive and a provider is available, the browser redraws a JPEG with a longest edge of at most 1,536 pixels. That metadata-free copy is sent to the backend and configured detection provider. Limits or failures can prevent a model check.
Step 4
Explain what remains unknown
A source declaration is not verified provenance. We do not validate C2PA signatures, asset bindings or certificate trust, and have no dedicated SynthID decoder. Missing metadata does not prove a camera origin. A general model cannot authenticate an identity or exact generator version.
Scores and source clues are different
A completed pixel-model score at or below 0.35 is displayed as likely camera-origin; at or above 0.70 as likely AI-generated. The middle range is inconclusive. These are display thresholds, not a validated probability calibration. Metadata-only readings describe editable source clues and are not model probabilities.
Conflicting metadata, damaged files and unavailable model checks are different states. A missing or failed check must not be read as a verified real image.
What we have evaluated
We do not publish a validated overall accuracy figure. An internal partial Hive engineering run requested 144 images, scored 98 and left 46 unscored. It was limited by service availability and a specific image-preparation pipeline. It is not a representative benchmark, does not establish per-generator accuracy and is not evidence of zero false positives.
Examples demonstrate input and result handling. Controlled metadata examples are not independent camera or generator benchmarks. Research must identify actual sample counts, sources, transformations and limitations separately.
Image handling
Source clues are checked in your browser first. When needed, file-header data and a smaller image copy may be sent to our backend and detection provider. Model copies have embedded metadata removed. Image URLs are retrieved by our server; see the privacy policy for processing and provider retention.
The backend may cache a provider score keyed by the reduced image hash for up to 30 days. This is separate from storing an original image. Detection providers have their own processing and retention policies.
Read the privacy policyMaintenance and corrections
Maintained by Check AI Free. Report reproducible errors or unclear descriptions through our contact page. The public MCP integration accesses the same service; it is not an independent accuracy endorsement.