Although photoshopped and falsified photographs have existed for as long as the Internet, anyone can now easily experiment with them thanks to recent advancements in machine learning. With AI image generators such as Midjourney, Microsoft Designer, and DALL-E 3, anyone can quickly create a scenario that looks realistic. It’s simple to understand how these kinds of visuals may easily mislead and manipulate entire communities. Thankfully, there is still hope because there are a few techniques we can use to identify AI-generated photographs.
How to spot an AI-generated image
AI image generators blend advanced technologies like GPT-4 and diffusion models to create realistic images from minimal input. Despite their impressive capabilities, these AI-generated images often display subtle cues revealing their artificial nature. Here are some key indicators to watch for.
1. Anatomical Features
AI, trained on vast datasets, can recreate environments by stitching images together, but struggles with fine details like body parts. For example, in an image of two people sitting cross-legged, one person might appear to have an extra limb due to the AI’s limitations in less common poses. Similarly, AI often merges fingers or omits joints in rudimentary image generation, as seen in an image of Pope Francis wearing a puffer jacket where his fingers are fused into a single mass despite appearing to hold an object.
2. Unnatural Hair
After verifying anatomical accuracy, scrutinize finer details like fur and hair, which pose challenges for AI in terms of texture and light interaction. Look for inconsistencies, particularly in how hair interacts with objects. Meta’s Imagine AI typically represents hair more realistically compared to DALL-E 3, which often presents hair as a smooth mass lacking natural flow and flyaway strands.
3. Garbled or glitched text
Modern AI struggles to convincingly mimic text, especially handwriting, and often fails to generate coherent sentences. Look for blurry or pixelated text, nonsensical characters, and assess whether the text fits the image’s context. In an example by Midjourney, AI attempts to replicate a watermark but produces a poor result.
4. Smooth or waxy skin and surfaces
While Midjourney’s latest version has improved significantly, it still struggles with surface texture, especially in replicating organic materials like skin. This often leads to nearly photorealistic images with obvious flaws like unnaturally smooth faces or waxy skin, akin to early smartphone cameras with aggressive beauty mode settings. When examining images with humans, watch for overly perfect skin lacking blemishes or pores, revealing their AI-generated nature upon closer inspection.
5. Out of focus backgrounds
Depth of field, crucial in photography for focusing attention, can seem artificial if not handled well. AI-generated images may exhibit overly blurred backgrounds lacking a natural gradient. In professional photography, distant objects blur progressively, but in AI images, this gradient may be absent. Uniform blur regardless of distance could be a sign of AI generation.
Conclusion
In conclusion, while AI image generators have made significant strides in replicating real-world scenes, they still grapple with various challenges, ranging from mimicking intricate details like text and surface texture to mastering depth of field. Despite their impressive capabilities, subtle cues often betray their artificial origins. As we continue to explore and scrutinize AI-generated images, understanding these limitations becomes crucial in distinguishing between genuine and AI-created content.