- \n
- Human accuracy on AI images sits between 49 and 62 percent depending on the study, which is at or barely above a coin flip. Confidence runs far ahead of it.<\/li>\n<\/ul>\n\n\n\n
<\/p>\n\n\n\n
- \n
- You are best at spotting fake faces and worst at fake landscapes, objects and street scenes. Nearly all popular advice targets faces, which is the category you already handle well.<\/li>\n<\/ul>\n\n\n\n
<\/p>\n\n\n\n
- \n
- Check provenance before you check pixels. Since May 2026 every ChatGPT image carries both a C2PA credential and a SynthID watermark, and Google is building detection into Search and Chrome.<\/li>\n<\/ul>\n\n\n\n
<\/p>\n\n\n\n
- \n
- Absence of provenance proves nothing. Screenshots and most social platforms strip C2PA metadata, so a missing credential is not evidence of anything either way.<\/li>\n<\/ul>\n\n\n\n
<\/p>\n\n\n\n
- \n
- The tells that still work are about physics, not anatomy. Shadow direction, reflections, depth of field and scale relationships are much harder for a model to keep consistent than a hand.<\/li>\n<\/ul>\n\n\n\n
<\/p>\n\n\n\n
- \n
- When the eye test is inconclusive, which is often, the AI Image Detector<\/a> gives you a read in seconds and does not care whether the subject is a face or a parking lot.<\/li>\n<\/ul>\n\n\n\n
<\/p>\n\n\n\n
\n\n\n\nWhy This Got Hard So Fast<\/strong><\/h2>\n\n\n\n
The visible improvement curve is steep enough that anyone who formed their instincts in 2023 is working from a museum piece. Compare what the same prompt produced from DALL E 2 in 2022 against what current models return and they barely belong to the same medium.<\/p>\n\n\n\n
The International AI Safety Report noted in 2026 that AI generated content has simply become harder to distinguish from real content across every format, and image quality is the clearest example of it.<\/p>\n\n\n\n
Two things happened at once. Models got better at the specific artifacts everyone had learned to look for, because those artifacts were the loudest complaints and therefore the first targets. And the aesthetic shifted. The early giveaway was that AI images looked too good<\/em>, with that airbrushed, overlit, faintly plastic quality.<\/p>\n\n\n\n
Current models will happily generate something that looks like a mediocre phone snapshot, complete with motion blur, bad framing, sensor noise and a thumb in the corner. The ugliness is now part of the output rather than a sign of authenticity.<\/p>\n\n\n\n
It is also worth being blunt about what the numbers say regarding detection tools versus people. Machine detectors reach 90 to 96 percent in laboratory conditions, though that figure drops substantially in the wild against models the detector has not seen.<\/p>\n\n\n\n
Even after that drop, it is not close. A tool having a mediocre day still outperforms a careful human having a good one.<\/p>\n\n\n\n
The Tells That Stopped Working<\/strong><\/h2>\n\n\n\n
Before the useful signs, it is worth clearing out the ones still circulating that will actively mislead you now. Every item below was genuinely reliable at some point. None of them is a safe basis for a judgment today.<\/p>\n\n\n\n
Old tell<\/strong><\/th> Why it worked<\/strong><\/th> Status in 2026<\/strong><\/th><\/tr><\/thead> Count the fingers<\/strong><\/td> Early diffusion models had no concept of hand structure and produced six or seven digits routinely<\/td> Effectively dead. Current models render hands correctly the large majority of the time.<\/td><\/tr> Look for garbled text<\/strong><\/td> Signage, book spines and labels came out as letter shaped noise<\/td> Mostly dead for short text. Still somewhat useful for dense paragraphs and unusual scripts.<\/td><\/tr> It looks too polished<\/strong><\/td> Early output had a uniform glossy sheen and impossible skin<\/td> Reversed. Models now imitate amateur photography, grain and bad lighting on request.<\/td><\/tr> Check the teeth and ears<\/strong><\/td> Fine repeating structures confused early models<\/td> Largely resolved, and it targets faces, where you already perform best.<\/td><\/tr> Watermark in the corner<\/strong><\/td> Some generators stamped visible marks<\/td> Trivially cropped or never present. Absence means nothing.<\/td><\/tr> Reverse image search finds nothing<\/strong><\/td> Real photos usually have a trail online<\/td> Weak. Real private photos have no trail either, and generated images now get indexed too.<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n The pattern connecting all six is that they describe capability failures<\/em> rather than structural ones. A capability failure is something the model was bad at and could be trained out of, and every one of these got trained out.<\/p>\n\n\n\n
The signs worth learning are the ones rooted in how these systems work rather than in what they happened to be bad at during a particular year, and I get to those shortly.<\/p>\n\n\n\n
The Blind Spot Nobody Writes About<\/strong><\/h2>\n\n\n\n
This is the part I found most surprising, and it changes what you should actually be suspicious of.<\/p>\n\n\n\n
When researchers break human detection accuracy down by subject matter, a consistent pattern shows up. In an analysis of roughly 287,000 image evaluations by more than 12,500 participants, overall accuracy came out around 62 percent, but performance was best on portraits and significantly worse on landscapes and urban scenes<\/strong>.<\/p>\n\n\n\n
Other work has found the same shape, with higher accuracy on images depicting people and lower accuracy on objects and scenery.<\/p>\n\n\n\n
The explanation is not mysterious. Humans have dedicated neural machinery for face processing and years of practice using it. You notice when something about a face is off even when you cannot articulate what. You have no equivalent instinct for whether a warehouse loading bay looks correct.<\/p>\n\n\n\n
Now put that next to the advice everyone gives. Count the fingers. Check the eyes. Look at the teeth and the ears and the hairline.<\/p>\n\n\n\n
Every one of those points you at faces, which is the category where your unaided judgment is already strongest and where model improvement has been most aggressive. Meanwhile the categories where you are close to blind get no coverage at all.<\/p>\n\n\n\n
Which means the images you should treat with the most suspicion are the least dramatic ones:<\/p>\n\n\n\n
- \n
- Product shots. A single object on a plain background gives you almost nothing to check against and is trivially generated.<\/li>\n\n\n\n
- Empty interiors. Rooms, offices, restaurant floors, apartment listings. No face to trigger your instincts and complex geometry your eye cannot verify.<\/li>\n\n\n\n
- Streets and landscapes. Exactly the category the studies found people worst at, and exactly what gets used for fake location claims and fabricated news scenes.<\/li>\n\n\n\n
- Documents and screenshots. Receipts, invoices, dashboards, chat logs. These get scrutinized for content and almost never for authenticity.<\/li>\n\n\n\n
- Crowds and distance shots. Faces too small to trigger face processing, so your best instrument never switches on.<\/li>\n<\/ul>\n\n\n\n
A generated portrait of a nonexistent person is the case you are equipped for. A generated photo of a flooded street with a caption naming a city is the one that gets past you, and it is also the one more likely to be doing real damage.<\/p>\n\n\n\n
Layer One: Check Provenance Before You Check Pixels<\/strong><\/h2>\n\n\n\n
Almost nobody does this, and it is now the fastest answer available for a growing share of images. The industry spent the last two years building a system that labels generated content at the moment of creation, and as of 2026 it covers a meaningful chunk of what you encounter.<\/p>\n\n\n\n
What the two systems are<\/strong><\/h3>\n\n\n\n
There are two different mechanisms and they are constantly confused with each other.<\/p>\n\n\n\n
Content Credentials, built on the C2PA standard, are cryptographically signed metadata attached to a file. They record where an image came from, what tools touched it, and who signed that claim. Think of it as a chain of custody rather than a yes or no answer.<\/p>\n\n\n\n
The specification reached version 2.3 in January 2026, and it is not only for AI. Camera manufacturers and news wire services use the same standard to certify that a real photograph is real, which is arguably the more important use.<\/p>\n\n\n\n
SynthID is different. It is an imperceptible watermark that Google DeepMind embeds into the pixels themselves at generation time, and it survives cropping, recompression and low quality re encoding in a way metadata does not.<\/p>\n\n\n\n
Google reports that more than 100 billion pieces of media have been watermarked with it. The tradeoff is that it tells you almost nothing beyond the fact of AI origin, and detecting it requires Google\u2019s own infrastructure rather than an open specification.<\/p>\n\n\n\n
What changed in 2026<\/strong><\/h3>\n\n\n\n
The important development is that these stopped being competing approaches. In May 2026, OpenAI and Google announced a joint model where images generated through ChatGPT and the OpenAI API carry both<\/em> a C2PA credential and a SynthID watermark.<\/p>\n\n\n\n
Given that ChatGPT is the largest consumer image generator by user count, a large share of AI images produced daily now carry two independent provenance signals. Kakao, ElevenLabs and Nvidia announced adoption the same day, and Google said C2PA and SynthID detection is coming to Search and Chrome directly.<\/p>\n\n\n\n
Regulation is moving in the same direction. The EU AI Act\u2019s Article 50 transparency and labeling obligations became binding on 2 August 2026, with penalties reaching into the tens of millions of euros. <\/p>\n\n\n\n
In the United States, the TAKE IT DOWN Act has imposed a 48 hour takedown duty since May 2026. Whatever you think of the enforcement prospects, disclosure is now a legal requirement in major markets rather than a voluntary courtesy.<\/p>\n\n\n\n
How to actually check<\/strong><\/h3>\n\n\n\n
- \n
- Open the image in a Content Credentials viewer. The Content Authenticity Initiative runs a public verification page, there are browser extensions that do it inline, and OpenAI has a verification tool that checks for both C2PA manifests and SynthID in one pass. Upload, read the manifest, done.<\/li>\n\n\n\n
- Read what the manifest actually says. A credential can record that an image was generated, or that it was captured by a camera and later edited, or that it passed through a specific editing tool. The interesting cases are usually in the edit history rather than the origin field.<\/li>\n\n\n\n
- Treat a missing credential as no information at all. This is the part people get wrong constantly.<\/li>\n<\/ol>\n\n\n\n
That last point deserves emphasis, because it is where provenance checking goes wrong. C2PA metadata is stripped by screenshots, by most social media transcoders, and by any editing workflow that does not deliberately preserve it.<\/p>\n\n\n\n
Most images circulating online carry neither a credential nor a watermark, and that is the normal case rather than the exception. SynthID only exists in output from generators that adopted it, which still excludes a large portion of tools in active use.<\/p>\n\n\n\n
So a clean provenance result is strong evidence. An empty one is nothing. Anyone telling you that no credential means suspicious has inverted the logic.<\/p>\n\n\n\n
Layer Two: The Physics Tells That Still Work<\/strong><\/h2>\n\n\n\n
When provenance comes back empty, this is where you look. The reason these signs outlast the anatomical ones is structural.<\/p>\n\n\n\n
A generative model produces a plausible image, not a simulation of a scene. It has no internal model of a room with a light source in it, so global consistency across the whole frame is genuinely hard for it in a way that drawing a correct hand never was.<\/p>\n\n\n\n
Follow the light<\/strong><\/h3>\n\n\n\n
Pick every object in the frame that casts a shadow and check whether the shadows agree about where the light is. Outdoors there is one sun, so shadows should run roughly parallel and be consistent in length relative to object height. Indoors it is more complex but still finite.<\/p>\n\n\n\n
Generated images routinely have shadows pointing in subtly different directions, or an object with no shadow at all sitting next to one that has a firm shadow, or a shadow whose shape does not match the thing supposedly casting it.<\/p>\n\n\n\n
Then check whether the lighting on the subject matches the environment. A person lit from the front left standing in a scene whose background is lit from the right is one of the most common failures in composite style generations, and once you start looking for it you find it everywhere.<\/p>\n\n\n\n
Interrogate every reflective surface<\/strong><\/h3>\n\n\n\n
Reflections are expensive for a model to get right because they require understanding the geometry of the whole scene from a second viewpoint.<\/p>\n\n\n\n
Look at mirrors, windows, glossy floors, car paint, water, sunglasses and phone screens. Ask whether what is reflected corresponds to what is actually in the scene, in the right position, at the right scale, flipped correctly.<\/p>\n\n\n\n
The catchlight test is my favourite quick check on any portrait. Those small bright spots in someone\u2019s eyes are reflections of the light sources. In a real photograph both eyes show catchlights in the same position, the same shape and the same count, because both eyes face the same lights.<\/p>\n\n\n\n
Generated portraits often have mismatched catchlights, or one eye with two and the other with one, or catchlight shapes that do not correspond to any light source visible in the image.<\/p>\n\n\n\n
Check the focal plane<\/strong><\/h3>\n\n\n\n
Real cameras have a plane of focus. Everything at the same distance from the lens is sharp together, and blur increases smoothly with distance from that plane.<\/p>\n\n\n\n
Generated images frequently blur by object importance<\/em> rather than by distance, so you find a background element crisply rendered while something beside it at the same depth is soft, or a subject sharp from nose to ear in a shot whose background blur implies a very shallow depth of field. Bokeh shape is another one.<\/p>\n\n\n\n
Real out of focus highlights take the shape of the lens aperture and stay consistent across the frame.<\/p>\n\n\n\n
Test scale relationships<\/strong><\/h3>\n\n\n\n
Models are unreliable about how big things are relative to each other, and the errors are easy to miss because your eye accepts the composition as a whole.<\/p>\n\n\n\n
Look for a door that is slightly too small for the person walking through it, a coffee cup that would hold a litre, a car whose wheels do not match its body, furniture at a scale that would not fit a human. This is especially productive in exactly the interior and street images where your instincts are weakest.<\/p>\n\n\n\n
Look for texture that repeats<\/strong><\/h3>\n\n\n\n
Real surfaces are irregular at every scale. Brick walls have damaged bricks, crowds have people facing odd directions, foliage has broken branches.<\/p>\n\n\n\n
Generated textures often tile subtly, so scan any large repeating surface for the same defect appearing twice, or an identical face in two places in a crowd. Zoom into the background rather than the subject, because that is where the model allocated the least effort.<\/p>\n\n\n\n
Read the edges and the junctions<\/strong><\/h3>\n\n\n\n
Where two things meet is where generation gets sloppy. Hair against a background, jewellery against skin, glasses frames crossing a face, a strap disappearing behind a shoulder and emerging in the wrong place.<\/p>\n\n\n\n
Follow any linear object that passes behind something else and confirm it comes out where geometry says it should. Straps, railings, power lines and fence posts are all good candidates.<\/p>\n\n\n\n
Layer Three: Run It Through a Detector<\/strong><\/h2>\n\n\n\n
The honest position is that layers one and two will leave you uncertain a lot of the time, and uncertainty resolved by staring harder tends to resolve toward whatever you already believed. This is what software is for.<\/p>\n\n\n\n
The AI Image Detector<\/a> analyzes the image for generation signatures at the pixel and frequency level rather than at the level of visible content, which is precisely why it does not share your blind spot. It has no special aptitude for faces and no particular weakness on landscapes.<\/p>\n\n\n\n
A parking lot is the same problem as a portrait to it. Given that the research shows your own accuracy swings wildly by subject matter, having a second opinion that is indifferent to subject matter is worth more than it sounds.<\/p>\n\n\n\n
Practical notes on using any detector well:<\/p>\n\n\n\n
- \n
- Feed it the highest quality version you can find. Heavy compression destroys the fine frequency information detectors rely on, so a screenshot of a screenshot is the worst possible input.<\/li>\n\n\n\n
- Treat the score as evidence rather than a verdict, especially in the middle of the range. A very high or very low score is informative. Something near the midpoint means look harder, not flip a coin.<\/li>\n\n\n\n
- Run the layers together. A missing provenance credential plus a shadow inconsistency plus a high detector score is a conclusion. Any one of the three alone is a suspicion.<\/li>\n\n\n\n
- Remember that heavy editing of a real photo can push a genuine image upward, in the same way heavy processing complicates audio detection. Retouched commercial photography is the common false positive risk.<\/li>\n<\/ul>\n\n\n\n
![\"ai-video-detector-1024x424\"](\"https:\/\/undetectable.ai/blog\/wp-content\/uploads\/2026\/08\/ai-video-detector-1024x424-3.webp\")
<\/picture><\/figure>\n<\/div>\n\n\nIf the image came from a video, pull a clean frame or use the AI Video Detector<\/a> instead, since it can read motion and temporal consistency signals that a single frame throws away.<\/p>\n\n\n
- When the eye test is inconclusive, which is often, the AI Image Detector<\/a> gives you a read in seconds and does not care whether the subject is a face or a parking lot.<\/li>\n<\/ul>\n\n\n\n
- The tells that still work are about physics, not anatomy. Shadow direction, reflections, depth of field and scale relationships are much harder for a model to keep consistent than a hand.<\/li>\n<\/ul>\n\n\n\n
- Absence of provenance proves nothing. Screenshots and most social platforms strip C2PA metadata, so a missing credential is not evidence of anything either way.<\/li>\n<\/ul>\n\n\n\n
- Check provenance before you check pixels. Since May 2026 every ChatGPT image carries both a C2PA credential and a SynthID watermark, and Google is building detection into Search and Chrome.<\/li>\n<\/ul>\n\n\n\n
- You are best at spotting fake faces and worst at fake landscapes, objects and street scenes. Nearly all popular advice targets faces, which is the category you already handle well.<\/li>\n<\/ul>\n\n\n\n
![\"ai-video-detector-1024x424\"](\"https:\/\/undetectable.ai/blog\/wp-content\/uploads\/2026\/08\/ai-video-detector-1024x424-3.webp)
<\/picture><\/figure>\n<\/div>\n\n\n![\"undetectable-ais-ai-voice-detector-1024x459\"](\"https:\/\/undetectable.ai/blog\/wp-content\/uploads\/2026\/08\/undetectable-ais-ai-voice-detector-1024x459-3.webp)
<\/picture><\/figure>\n<\/div>\n\n\n