Our Mission

Realdle was created to help people develop their perception skills in an era where artificial intelligence can generate photorealistic images that rival human photography. As generative AI models advance rapidly, the line between reality and simulation becomes increasingly blurred. We believe that by creating an interactive, daily challenge, we can help our community learn to identify the subtle anomalies and "tells" that still exist in AI-generated media.

Our goal is not to diminish the incredible technological achievements of AI researchers, but rather to foster digital literacy. By comparing a real photograph with an AI-generated counterfeit side-by-side, users can train their eyes to notice inconsistencies in lighting, geometry, and textures.

How the Game Works

Every day, Realdle presents players with a new challenge: two images side-by-side. One is a genuine, human-captured photograph, while the other is an AI-generated imitation.

To create these puzzles, we take the original photograph and pass it through an advanced image-to-text model. This AI describes the scene in painstaking detail. We then take that text description and feed it into a state-of-the-art text-to-image generator. The result is a brand-new image that closely matches the content and composition of the original, but is entirely synthesized by a machine.

Your task is to analyze both images, spot the AI-generated fake, and click on the image you believe is real.

Tips for Spotting AI

While AI image generators are incredibly sophisticated, they often struggle with specific details. When analyzing our puzzles, look out for these common anomalies:

• Text and Typography: Generative models often produce gibberish or strangely shaped letters instead of coherent text on signs, clothing, or backgrounds. This is especially true in the case of long text or with many instances of text (like in a list or a grid).
• Hands and Fingers: AI frequently struggles with the complex geometry of human hands, sometimes generating too many fingers or awkward joints. Do note that models have improved drastically on this failing since the inception of generative imagery and this is no longer an absolute tell.
• Symmetry and Geometry: Look closely at architectural details like windows, tiles, or patterns. AI might make lines asymmetrical or misaligned.
• Background Details: While the main subject might look perfect, background elements are often blended together or nonsensical upon closer inspection.
• Reflections and Shadows: AI can sometimes produce light sources or reflections that defy physics or do not match the environment.
• Too Perfect: AI can sometimes produce images that are too perfect with a fairytale-like quality lacking in blemishes or imperfections.

Community and Heatmaps

After making your choice, we allow users to click on the exact location of the "giveaway" anomaly in the AI image. This data contributes to our community Heatmaps, revealing the most common visual artifacts that gave the fake away. Over time, these heatmaps provide fascinating insights into how both humans perceive images and where AI models consistently fail.

AI Model Performance & Rankings

Realdle isn't just a game; it's a living benchmark for generative AI. We track how often players are successfully tricked by different AI models, creating a real-time leaderboard of model "deceptiveness."

Our Model Rankings page provides transparent metrics on the "Trick Rate" of various state-of-the-art models. By aggregating thousands of human guesses, we can scientifically measure which models are currently leading in visual realism and which ones still exhibit identifiable "tells." This data helps researchers and the public understand the trajectory of synthetic media quality across different architectures.

Check out the latest leaders on our Model Rankings page to see which AI is currently the most difficult to spot.