In 2022 and 2023, deepfake photos showed obvious visual artifacts: blurry ear edges, asymmetric faces, nonsensical text, and unnaturally smooth skin. By 2025 and 2026, these tells have largely been eliminated by diffusion model improvements. Expert human reviewers can now correctly identify AI-generated images only 30 to 40% of the time on blind test sets. This shift makes forensic analysis the primary, and often only, reliable method of detection.

Visual signs still worth checking

While visual detection is unreliable against modern AI, certain patterns remain more common in AI images than authentic photographs: hands and fingers with unusual proportions or extra digits; teeth that are unnaturally uniform or blend strangely at the edges; earrings, necklaces, or accessories that change shape or disappear at different viewing angles; background elements with subtle tiling patterns or perspective errors; and text in the image that looks like letters but is not actually readable.

Forensic signs in the image file

File-level forensic signs are more reliable than visual inspection. Missing EXIF on a JPEG file: a genuine photograph always has camera metadata. PNG format: most cameras save JPEG; AI generators commonly output PNG. Exact standard dimensions: 1024x1024, 896x1152, 1344x768 are AI generator output sizes, not camera sizes. AI software tags: metadata referencing Stable Diffusion, Midjourney, DALL-E, or similar platforms is definitive. Generation parameters in PNG metadata: AUTOMATIC1111 and ComfyUI embed full prompt text in PNG files by default.

Face-swap deepfake signs

Face-swap deepfakes (where a real person's face is composited onto existing content rather than fully generated) show different signs than pure AI generation. Look for: inconsistent lighting between the face and body; a slight halo or soft blur at the face boundary; mismatched skin tone or texture between face and neck; and temporal inconsistencies in video (blinking patterns, head movement that differs from body movement). EXIF analysis may be less useful for face-swaps applied to existing photos, since the base image may have authentic camera metadata.

Platform context as a signal

The context in which an image appears provides additional evidence. Profile photos on accounts created very recently with no history, or images shared specifically to coerce or embarrass someone, are higher suspicion. Images that cannot be found through reverse image search on any other platform (suggesting they were newly generated) are more suspicious than images with a traceable origin.

Frequently asked questions