How to spot AI-generated images (2026 edition)
AI image generators get better every quarter. Here are the tells that still work — and the ones that no longer do.
By February 2026, frontier image generators (Midjourney v7, OpenAI gpt-image-1, Imagen 4, Stable Diffusion 4) produce photo-realistic output that defeats most casual visual inspection. The old “count the fingers” trick is largely obsolete. Here’s what still works.
Reliable tells (2026)
Text and signage. AI generators still struggle with consistent readable text in images. A storefront sign that looks like real English at first glance but dissolves into garbled letters on close inspection is a strong AI signal.
Reflections. Reflective surfaces (mirrors, water, sunglasses, polished tables) should reflect what’s in front of them. AI generators often produce reflections that don’t match the scene’s geometry.
Hands holding specific objects. Generic hands are now well-rendered. Hands holding *specific* objects (pens, phones, glasses with stems) often produce subtle anomalies — fingers passing through the object, missing thumbs, wrong angles.
Repeated patterns. AI generators tend to produce slight inconsistencies in repetitive elements: bricks of unequal height in a wall, books with similar-but-different spines, crowd-scene faces that share suspicious similarities.
Symmetric body parts (eyes, ears, eyebrows). Real human eyes are slightly asymmetric. AI-generated faces sometimes produce eyes that are *too* perfectly mirrored.
Image metadata. Right-click → properties (or use `exiftool`). Real cameras embed model/lens metadata. AI-generated images often have either no metadata or generator-specific signatures (Midjourney, etc.).
Tells that NO LONGER work in 2026
- Counting fingers (generators fixed this in 2024) - Looking for blurry edges (resolution is no longer a tell) - Looking for unnatural lighting (generators are now lighting-consistent) - Checking eyelashes / individual hairs (no longer reliable)
The C2PA / Content Provenance standard
The C2PA standard, now widely adopted by major camera makers and AI companies, signs images at the moment of capture or generation. Software that supports C2PA (browser extensions, social platforms) can show you the provenance chain. Absence of provenance doesn’t prove fakery, but its presence is a strong authenticity signal.
When in doubt
Default to skepticism on emotionally-charged images that lack a clear news-outlet source. If a powerful image is being shared without a byline, source link, or contextual reporting, treat it as suspect until verified.
Related: Deepfake · Content provenance · How to verify a viral video.