AI image detector
(Midjourney, DALL·E, Flux)

Upload the image to ScanTrace. In under 15 seconds you receive one of three verdicts — REAL, AI_GENERATED or INDETERMINATE — together with a confidence score between 0 and 1 and a narrative report explaining the forensic signals. No account is required for a single test analysis.

Three independent forensic layers run in parallel:

1. Pixel-level statistical analysis. Generative models leave fingerprints in the frequency domain — DCT coefficients, noise residues and PRNU patterns differ from those produced by physical camera sensors. SightEngine's genai model has been trained on millions of labeled examples from every public generator.

2. EXIF metadata inspection. Real camera files carry hundreds of EXIF tags: camera make, lens, focal length, ISO, shutter speed, GPS, color profile and a sensor serial number. AI-generated images either have no EXIF or contain tell-tale "Software: Midjourney" entries.

3. Multimodal contextual reasoning. Anatomical inconsistencies (extra fingers, malformed ears, impossible reflections, distorted text) are flagged by an LLM that receives only the numeric signals from layers 1 and 2 — never the raw pixels — and writes a human-readable explanation.

According to Europol's Innovation Lab, more than 90% of online content could be synthetically generated by 2026. The European Broadcasting Union reports that 77% of newsrooms have already published an AI-generated image by mistake at least once. Adobe Firefly alone crossed 6 billion generated images in 2024.

Always check before publishing any image you did not personally capture, especially when sourced from social media, instant messaging, breaking-news feeds, eyewitness submissions or material from conflict zones. The 15-second turnaround makes verification compatible with editorial deadlines.

How accurate is the AI image detector?

ScanTrace reaches above 96% accuracy on high-resolution images with sufficient forensic signals. When signals are weak or contradictory the tool returns an INDETERMINATE verdict instead of guessing — false positives damage trust more than missed detections.

Which generative models can it identify?

The detector is trained against the dominant 2024–2026 models: Midjourney v5/v6/v7, DALL·E 2 and 3, Stable Diffusion 1.5/XL/3, Flux.1 (Pro/Dev/Schnell), Adobe Firefly, Google Imagen and Ideogram. The fingerprint database is updated as new public models ship.

Is it really free?

Yes. The free plan includes 10 image analyses per month with no credit card. Paid plans (Pro 400/month, Newsroom 1500/month) unlock PDF certificates, video analysis and team workspaces.

Does it work on screenshots and compressed images?

Compression and re-uploads degrade the forensic signal. ScanTrace still works on JPEG screenshots but the lower the resolution and the more re-encodings, the more likely the verdict will be INDETERMINATE.

Will it detect images from a brand-new model I have not heard of?

Often yes, because most generative models share common statistical fingerprints (frequency-domain anomalies, lack of optical noise, missing EXIF). Brand-new models may slip through until the database is updated — when in doubt the verdict is INDETERMINATE rather than REAL.

Do you store the images I upload?

Anonymous free analyses are processed in memory and discarded immediately. Logged-in users can opt to keep an analysis history in their dashboard, stored in a private encrypted Supabase bucket with signed URLs that expire in one hour.