Can you tell a real photo from an AI one?

Why This Is Harder Than It Looks

Public deepfake detection studies keep landing on the same uncomfortable number: people identify synthetic media correctly about 55 percent of the time. That is a coin flip with extra steps. The reason is not that people are careless. It is that the tells moved. Generated media stopped failing in obvious ways — six-fingered hands and melted text — and started failing in places nobody thinks to check, like whether both eyes reflect the same light source or whether the bokeh follows real lens optics.

Each question in this test names the specific tell after you answer, along with the share of people who get that image wrong. The point is not the score. It is learning which category of detail your eye currently skips.

What Is the Real or AI Challenge?

The Real or AI Challenge presents a curated selection of images and asks a single question: was this authentically captured, or generated or manipulated by artificial intelligence? It runs free in this browser, and the same challenge library is built into the HumanMeter app for iOS and Android.

The difficulty ramps as you progress. Early images carry recognisable AI artifacts such as distorted anatomy, warped text, or unnaturally smooth skin. Later ones mirror the output of current generative systems, where the only remaining giveaways are optical: light direction, sensor grain, depth-of-field behaviour, and frame-to-frame consistency. Both beginners and experienced viewers find a wall somewhere in the set.

Every image carries a verified answer and a documented tell, so a wrong guess teaches you something specific rather than leaving you to wonder. Over time this builds a practiced intuition for synthetic media that does not depend on having a detection tool in your hand.

Why Train Your AI Detection Skills?

The landscape of digital media has shifted dramatically. AI-generated video content has become remarkably convincing, and the tools to produce it are more accessible than ever. Deepfakes that once required specialized hardware and expertise can now be created with consumer-grade software in minutes. This means the volume of synthetic content circulating on social media is growing at an unprecedented rate.

Misinformation campaigns increasingly rely on fabricated video to lend credibility to false narratives. A convincing deepfake of a public figure can spread across millions of feeds before fact-checkers have a chance to intervene. Political manipulation, financial fraud, and personal harassment are all fueled by synthetic media that looks and sounds authentic to untrained viewers.

Automated detection tools like HumanMeter provide a critical layer of defense, but no single technology catches every piece of manipulated content. Building your own detection skills creates a second line of defense. When you understand what to look for -- the subtle inconsistencies in lighting, the telltale shimmer around edges, the unnatural cadence of generated speech -- you become far harder to deceive. Training your eye is not a replacement for technology; it is a complement that makes both human judgment and AI analysis stronger together.

What You'll Learn

Each challenge round is built to develop specific observation skills. Over time, you will become proficient at identifying the following indicators of AI-generated content:

• Spotting unnatural motion: AI-generated videos frequently exhibit movement that looks slightly off. Limbs may bend at impossible angles, walking cadence may appear too smooth or too rigid, and facial expressions may shift in ways that do not match natural human behavior.
• Identifying lighting inconsistencies: Generative models often struggle to maintain consistent light sources across a scene. Shadows may fall in conflicting directions, reflections may not align with their surroundings, and skin tones may shift unnaturally between frames.
• Recognizing morphing artifacts: When AI blends or transitions between generated frames, morphing artifacts can appear. These include warping around the edges of faces, distortion in hair and clothing, and objects that briefly change shape or dissolve before reforming.
• Detecting temporal glitches: Authentic video maintains consistent detail across consecutive frames. AI-generated content may introduce flickering textures, disappearing accessories, or background elements that shift position without explanation between frames.
• Understanding AI generation tells: Each generation model leaves distinct fingerprints. These can include overly symmetrical facial features, unnaturally perfect skin texture, repeating patterns in backgrounds, and characteristic distortions in hands, teeth, and fine details like jewelry or text.

Frequently Asked Questions