By 2026, image generators produce results clean enough that the obvious tells from two years ago (six-fingered hands, garbled text, waxy skin) have mostly been fixed. That doesn’t mean AI images are undetectable, it means the useful checks have shifted to things that are harder to fix because they’re baked into how these models actually work, not surface-level rendering glitches.
Checks that still work in 2026
1. Text in the background, not the foreground
Foreground text on signs and labels has improved dramatically since 2023, but text sitting in the background, on a distant storefront, a license plate, a book spine on a shelf, small print on packaging, is usually still where models fail. Diffusion models allocate less attention to background detail because it contributes less to the overall composition score during training. Zoom into any small, secondary text in an image before trusting it.
2. Reflections and shadows that don’t match their source
Check whether a shadow’s direction matches the apparent light source, and whether reflections in glass, water, or metal show the same scene that a real reflection would. AI models generate reflections as a plausible-looking pattern rather than a geometrically accurate mirror of the scene, so a reflection in a puddle or window that shows different objects, or objects at the wrong angle, than what’s actually in frame is still a strong signal.
3. Ears, teeth, and hands under stress poses
Hands in simple resting poses are now rendered correctly most of the time. Hands gripping an object at an unusual angle, hands overlapping other hands, or hands partially obscured by hair or clothing still produce extra or fused fingers more often than not. The same applies to ears (irregular cartilage folds are still a common failure point) and teeth in wide, natural smiles, where individual tooth boundaries sometimes blur together.
4. Repeating patterns that don’t actually repeat correctly
Brick walls, tiled floors, fabric weaves, and fence patterns should repeat with mathematical regularity. AI-generated versions of these patterns frequently drift, a brick course that’s straight on the left of the image curves by the right edge, or a tile grid that subtly changes size across the frame. This is one of the more reliable tells because it requires the model to maintain exact geometric consistency over a wide area, which diffusion architectures still struggle with.
5. Metadata and provenance signals
Real cameras embed EXIF data: camera make and model, lens information, GPS coordinates if location services were on, and a timestamp. Most AI image generators either strip this entirely or embed generic/placeholder metadata. Checking an image’s EXIF data with any standard metadata reader won’t prove an image is real, since metadata is trivially stripped or faked, but a complete absence of any camera metadata on an image claiming to be a real photograph is a signal worth noting alongside everything else.
A more forward-looking signal is C2PA content credentials, a standard backed by Adobe, Microsoft, Google, and camera manufacturers including Sony, Nikon, and Leica, which cryptographically signs an image at capture and records any edits made afterward. Adoption is still uneven, most images circulating online don’t carry a C2PA credential either way, but where one is present and verifies cleanly, it’s a stronger signal than any visual analysis. Where an image claims newsworthy authenticity and has zero provenance data of any kind, treat that as one more data point rather than proof.
6. Google’s SynthID and other embedded watermarks
Google’s Gemini and Imagen models embed SynthID, an invisible statistical watermark, into every image they generate. It survives cropping, compression, and most common edits, and Google provides a verification tool for checking whether an image carries a SynthID signature. OpenAI’s DALL-E and Sora models similarly embed C2PA metadata by default as of 2025. The practical catch is that watermark detection only works if you have access to the matching verification tool, and open-source or older models don’t embed anything at all, so a negative result from a SynthID check tells you nothing conclusive on its own.
7. Reverse image search for compositing tells
A reverse image search (Google Lens, TinEye, or Bing Visual Search) won’t detect AI generation directly, but it catches a related and increasingly common problem: real photos that have been partially AI-edited or composited with a generated element. If a search turns up the same base photo circulating without the element in question, that’s evidence of manipulation even if the underlying photo was real.
What no longer works reliably
Skin texture smoothness, once a dead giveaway, has been solved by most current-generation models through deliberate grain and pore simulation. Eye reflections (checking for symmetric, matching catchlights in both eyes) used to be a reliable tell but modern models now render asymmetric, scene-consistent reflections correctly in most cases. And “uncanny valley” facial expression stiffness, a common heuristic in 2023-2024, is no longer dependable since the latest generation of models handles subtle micro-expressions convincingly.
Using detection tools, and their limits
Automated detectors like Hive Moderation, Reality Defender, and Sensity AI score images for the statistical fingerprints diffusion models leave behind. These tools are genuinely useful for high-volume moderation but should be treated as one input among several rather than a verdict, because their accuracy varies significantly depending on which model generated the image and whether it’s been compressed or re-uploaded since generation, both of which can degrade detection accuracy substantially.
The realistic bar for 2026
No single check above is conclusive on its own. The combination that actually holds up is: zoom into background text and repeating patterns, check reflections and shadow direction, look at hands and ears under any kind of stress pose, pull whatever metadata is available, and run a reverse image search if the image is being used to support a factual claim. None of that requires special software, and together it catches the overwhelming majority of AI-generated images still circulating as authentic in 2026, even as the underlying models keep improving.
Related reading: how LLM API gateways handle routing and failover across providers, and the naming controversy behind Sakana AI’s Namazu model.