Fake bank checks generated by the new GPT-2 image model

Forensic Analysis

The viral images circulating in mid-April 2026 showed photographs of physical-looking bank checks with handwritten amounts and signatures, accompanied by mobile deposit confirmation screenshots. The implication: a new image generator was good enough to defeat banks' check-image verification.

Our forensic analysis disagrees. While the overall composition is convincing at thumbnail size, four independent signal layers flag the images as synthetic with high confidence.

The MICR (Magnetic Ink Character Recognition) line at the bottom of each check shows non-uniform character spacing and slight kerning errors that wouldn't appear on a real printed check — banks use mechanically precise MICR fonts. The handwritten signature shows the smooth, gradient-edge profile typical of diffusion-generated handwriting; real ink has a much sharper edge transition between stroke and paper.

No EXIF metadata was present, but the embedded JPEG quantization tables were consistent with a re-save of a generated PNG rather than a phone-camera capture pipeline. Combined with the noise topology (uniform across the image, instead of the typical CMOS-sensor pattern), this places the image origin as a generator output, photographed or screenshotted later.

Whether the deposit screenshots were also fabricated or real (with money never actually clearing) is outside the scope of image forensics — but the underlying check images are confidently synthetic.