State of Deepfakes 2026

This page explains how VerifyAI tests detection across image, video, voice and text generators, and is honest about where every method fails. We'll publish measured results here once our open test run is complete — we don't publish accuracy numbers we haven't measured.

Key facts

• No single detector is reliable on its own. VerifyAI combines four checks: provenance (C2PA), metadata and file forensics, error level analysis, and an AI review.
• Most files online have no C2PA Content Credentials, because platforms often strip them. Missing credentials are not evidence of fakery.
• Social media and messaging apps recompress files, which weakens metadata and ELA evidence.
• Voice notes and short clips are the hardest media to judge.

Test set

• Images: Midjourney, Flux, GPT Image, Stable Diffusion, plus real phone and camera photos.
• Video: Sora, Kling, Veo, Runway, plus real phone footage.
• Audio: cloned and synthetic voices, plus real recordings and compressed voice notes.
• Text: ChatGPT, Claude and Gemini writing, plus human writing.
• Each sample is tested as the original file and after a social-media style recompression.

What each check can prove

• C2PA: who signed the file and which edits were declared — only when credentials exist.
• Metadata: camera, software and timeline conflicts — easy to strip.
• ELA: edited regions in JPEGs — weak after recompression.
• AI review: likelihood of generation and the closest tool — a probability, not proof.

How results are reported

Each result is sorted into clear, review or likely tampered, with reasons. Workspace checks get a SHA-256 fingerprint and a certificate anyone can verify. Try it on the free check.

Researchers and journalists: contact support@goverifyai.com for the methodology.