AI-generated images and videos are flooding social feeds, and synthetic art is now appearing in galleries and auctions. The dramatic image below — the Hollywood sign apparently engulfed in flames — spread widely as real news during the 2025 California wildfires. It is entirely AI-generated. Spotting synthetic media is no longer a niche skill for fact-checkers; it is basic digital literacy. Here is how to do it: the visual tells that still work, the free verification tools worth using, and what to do when something looks suspicious.
The Visual Tells That Still Work in 2026
Newer models have fixed the obvious giveaways — basic finger counts, clean faces — so the old checklist is unreliable on its own. Open the image at full size and scan the edges and background, where generators cut corners:
- Text is the weakest link. License plates, shop signs, book spines, and logos often render as garbled lettering that looks like language from a distance but is gibberish up close. Background text breaks down first.
- Hands in complex poses. Gripping, pointing, or intertwined fingers still produce fused knuckles, extra digits, or impossible bends.
- Impossible physics. Shadows pointing in conflicting directions, reflections in mirrors or water that do not match the scene, jewelry or straps merging into skin.
- Backgrounds that melt. A sharp subject with dissolving signage or crowds behind them, foliage tiling like wallpaper at 100% zoom, architecture with impossible geometry.
- Objects that could not function. Scissors hinged in the wrong place, cables running to nowhere, a chair missing a leg behind a table.
- Over-smooth skin — with caution. Waxy, poreless skin is a classic tell, but phone retouching and beauty filters create the same look, so this alone proves nothing.
Researchers at Northwestern's Kellogg School grouped these into five families — anatomical implausibilities, stylistic artifacts, functional implausibilities, violations of physics, and sociocultural implausibilities. The framing pushes your eye toward the edges of the frame, where the evidence usually lives.
Free Tools That Check What Your Eyes Cannot
Visual inspection alone was never reliable, and in 2026 it is even less so without tools. A 2026 benchmark of 23 detection tools found roughly 75% accuracy against 2020–2021 generators but only about 18–30% against today's commercial models like Flux and recent Midjourney versions; in one large Microsoft study, human judges managed only about 62%. Stack independent signals and never trust a single score.
- Reverse image search first. Google Lens or TinEye reveal the earliest appearance. A dramatic image with no earlier footprint deserves suspicion.
- Check provenance. The free C2PA Content Credentials Verify tool reads signed metadata and edit history in your browser. OpenAI attaches C2PA to its image outputs, Google embeds invisible SynthID watermarks, Adobe Firefly signs its output, and Leica, Nikon, Sony, and Canon now ship C2PA-capable cameras. Treat metadata as a one-way signal: presence is close to conclusive, absence proves nothing.
- Pixel-based detectors as backup. Hive Moderation offers free detection for images and video; Illuminarty adds a heatmap of suspicious regions; aiimagedetector.com and Sightengine give solid free single checks. Run two or three and look for agreement.
- Edited real photo? FotoForensics' free error-level analysis spots compression inconsistencies from manipulation.
Spotting Fake Videos and Cloned Voices
Provenance first, then your eyes and ears. Watch for lip movements drifting out of sync, blurring teeth, unnatural blinking, or a face that wobbles when the head turns. Cloned voices can sound right but carry flat delivery, odd pacing, or missing breaths. The Content Credentials tool also accepts MP4, MOV, MP3, and WAV files, and the Gemini app can check clips for SynthID.
The strongest defense is procedural: if a "relative" or "boss" calls asking for money or access codes, hang up and call back on a number you already have stored. An agreed family code word defeats voice clones completely — the $25 million deepfake video-call fraud in Hong Kong proved no detector substitutes for this habit.
What to Do When You Suspect a Fake
- Do not share it first. Sharing is how a fake gets its distribution; a pause costs nothing.
- Check the source, not just the pixels. Who posted it first? Do reliable outlets show the same event from other angles? A dramatic image existing on only one account deserves suspicion.
- Stack your evidence. Visual check plus reverse search plus provenance plus two or more detectors — and document the process for consequential decisions.
- Never accuse on a single score. A detector saying "human" on a modern generation is often simply wrong; flagging a real photo as AI can harm innocent people.
The takeaway: no single trick catches everything in 2026 — generators improved faster than detectors. Your real defense is a layered habit: zoom in, reverse-search, check provenance, cross-check tools, verify the source.
Sources: geotag.world, Eyesift, SecurityElites, TechZei.