How to tell if an image or video is AI-generated in 2026 and why the old tricks stopped working
The “count the fingers” era of spotting fakes is over. Here’s an honest look at what still helps you verify images and video, what no longer works, and where the limits are.
Start with the hard truth
For years the advice for spotting fakes was to look for telltale glitches: extra fingers, mangled teeth, unnatural blinking. That guidance is largely obsolete. Modern generators have fixed most of the obvious artifacts, and research suggests people are no longer reliable at catching still-image fakes by eye. A University of Florida study published in early 2026 found that automated detectors were highly accurate — up to about 97% — at identifying AI-generated still faces, while human participants performed no better than chance on the same images.
Interestingly, the same study found humans still did better than the algorithms at spotting fake videos, correctly judging real from fake roughly two-thirds of the time, because moving footage gives the brain more cues to work with. The takeaway isn’t that detection is hopeless — it’s that you should lean on verification habits and tools rather than a checklist of visual “tells.”
What still works: verify the source, not the pixels
The most durable techniques have little to do with squinting at an image. Reverse image search remains valuable: uploading a picture can show where else it appears online, which helps reveal whether it has been reused, altered or taken out of context. If an image surfaces only on unrelated or low-credibility pages, treat it with caution. For any consequential claim, trace the content back to a primary source — the person, organization or outlet supposedly behind it — before believing or sharing it.
Check for content credentials (and know their limits)
A growing number of cameras and AI tools attach Content Credentials — a cryptographically signed record of provenance based on the C2PA standard — that can show how a file was created and edited. Verification is rolling out across major platforms and tools, and camera makers including Canon have begun building C2PA workflows for newsrooms. But credentials are a provenance signal, not proof of truth: metadata is often stripped when files are uploaded, screenshotted or re-encoded, and the absence of credentials does not prove an image is fake. Independent researchers have also shown valid manifests can be forged or attached to AI images, so treat credentials as one input, not a verdict.
For live video and calls: go off-script
Real-time voice and video deepfakes are now a documented fraud vector — in one widely cited 2024 case, attackers used a deepfake video call to impersonate executives and trick an employee into transferring about $25 million. Security practitioners recommend a low-tech defense: ask an unscripted question that a real person could answer instantly but a scripted impersonator would stumble on, and if anything feels off, hang up and call back on a known number. For families, agreeing on a private “safe word” to confirm identity during an unexpected emergency call is a simple, effective check against voice cloning.
A realistic mindset
Detection is an arms race, and the people generating fakes currently hold a speed advantage. No single tool or sign is definitive, and detector accuracy on brand-new generators tends to dip until they catch up. The practical goal is not perfect identification but informed judgment: combine source verification, reverse search, provenance checks and plain skepticism, and raise your guard in proportion to the stakes of the claim.
General information, not security or legal advice. Detection methods and tools change quickly; verify current options before relying on any single one. Compiled from sources as of June 30, 2026.
Sources
- University of Florida News — Machines spot deepfake pictures better than humans (Feb 2026) — https://news.ufl.edu/2026/02/deepfake-detection/
- MIT Media Lab — Detect DeepFakes project overview — https://www.media.mit.edu/projects/detect-fakes/overview/
- TechTarget — How to detect deepfakes manually and using AI — https://www.techtarget.com/searchsecurity/tip/How-to-detect-deepfakes-manually-and-using-AI
- GIJN — Reporter’s Guide to Detecting AI-Generated Content — https://gijn.org/resource/guide-detecting-ai-generated-content/
- C2PA — Content Credentials technical specification — https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html
- Canon Global — C2PA-compliant Authenticity Imaging System (May 2026) — https://global.canon/en/news/2026/20260511.html
