Is this image AI-generated?

Drop a JPEG, PNG, or WebP. We check C2PA / Content Credentials first, then run four forensic transforms in your browser — ELA, noise residual, FFT spectrum, and channel decomposition — and show you exactly what each one reveals.

Honest about the limits: a May 2026 benchmark of 291 generators put the best open detector at 75% mean accuracy — with Flux Dev at 21%, Midjourney v7 at 24%, and Imagen 4 at 19% on average across tested methods. That's why we show every signal instead of a single confidence score you can't audit.

What you'll see

Anatomy of a forensic result

Every check renders to its own panel. You can audit each signal independently — and that's the point. Combined evidence beats any single classifier.

  • Error Level Analysis

    ELA

    Re-saves the image at known JPEG quality and measures compression delta. Splices and edits glow.

  • FFT Spectrum

    FFT

    Fourier transform reveals periodic artifacts left by GANs and diffusion samplers — invisible to the eye, loud in frequency.

  • RGB

    Channel Decomposition

    RGB / YCbCr

    Splits the image into color channels. AI generators leak color statistics that don't match real cameras.

  • Noise Residual

    PRNU

    Real cameras leave a unique sensor-noise fingerprint. AI images either lack one or have a learned, repetitive one.

  • Saliency Heatmap

    Grad-CAM

    Where is the detector model paying attention? Heatmap shows the regions driving the verdict.

  • cr.signed · OpenAI · 2026-04-30

    C2PA Provenance

    Cryptographic

    Reads tamper-evident signatures from Adobe, OpenAI, Sony, BBC content. When present, this beats every other signal.

Provenance, in four states

C2PA / Content Credentials are a real signal in May 2026 — Leica M11-P / SL3-S, Sony Alpha 1 II / 1 / 9 III / 7S III / 7 IV, Canon R1 / R5 II, Pixel 10 native, plus OpenAI, Adobe Firefly, Microsoft, Google Gemini / Veo / Lyria. We show the right answer in four states: verified, present but untrusted, present but broken, or no credentials found. A green check isn't the same as a fully revocation-checked signature — and we won't pretend it is.

Hybrid forensics, not one classifier

The frontier (Flux, SD3.5, Midjourney v7, Imagen 4, GPT Image) broke single-model detectors. The current research consensus is hybrid: a semantic-aware learned detector (CO-SPY, AIDE) + a diffusion-reconstruction expert (DRCT, DIRE) + classical overlays for explainability. We ship the overlays today; the learned ensemble is next.

Want the science?

Read The Tells of an AI Image for a forensic walkthrough, or our State of AI Image Detection for the May 2026 benchmark numbers and method comparison.