As the amount of AI-generated content continues to spread and grow across the internet, so does our collective need to successfully suss out the “fakes from the facts.” This article is the first in a three part series that will provide some easy tips, tactics and tools that anyone can use to sharpen their eye for content generated by AI.
Today’s first installment looks at five ways to tell if an image is AI generated.
Texture and Resolution: AI-produced images may display inconsistencies across textures. Look for areas of unnatural smoothness or blurring where there should be continuous textural patterns.
This visual nit usually occurs due to AI’s processing limitations in that it has challenges uniformly rendering images with its current technology and programming. These nuanced discrepancies are usually evident in the background, midground and side frames of the image — not so much in the main areas of focus.
Anatomy Errors: While this particular area has seen rapid improvement, AI still makes mistakes regarding visual proportions of arms and legs, fine details around toes, fingers and teeth as well as precise location of facial features.
These mistakes can range from subtle to glaringly obvious but are telltale signs that the image has been edited or generated using AI tech. If there’s something about the people in the picture that doesn’t look quite right to your eye but you don’t know exactly what it is — it was probably made with AI.
Tool Recommendations: As the tide of deepfakes rises on the interwebs, there are several AI image detection tools available. Two that you can trust without any reservations are Sightengine and Hive. Both have free versions of the AI picture detection feature, as well as multiple sophisticated functions beyond just imagery, for a fee.
What’s really cool is that they both give you a percentage indicator as to which specific mainstream AI platform was used to produce the image such as Midjourney, Dall-E, Firefly…etc. In the case of the image below, Sightengine accurately nailed with 99% confidence that the purple dinosaur holding a bitcoin while saying “HODL on” was indeed produced using the Midjourney genAI image platform.
In a recent comparative study conducted by the University of Rochester and the University of Kentucky, Sightengine was determined to have the highest accuracy in the current marketplace when it came to AI detection.
However, it’s worth noting that no single tool is foolproof so using multiple tools is a best practice. The AI-generated dino image below was also tested on Hive, which accurately identified that the image was produced using artificial intelligence with more than a 99% confidence rating.
Shadows and Lighting: This is another area where the current generation of generative AI seems to struggle. It has difficulty accurately recreating the contours produced from real lighting and shadows. Even to a casual observer, these visual irregularities won’t align with a lifetime of contrast imagery and can be a clear tip-off that the image is AI fakery.
Specific things to look for might be shadows that are too dark or too light in relation to the ambient light in the image, or the shadows are cast in directions that don’t match the directional light sources within the picture.
AI Watermarking: Several of the AI tools available on the market automatically place a subtle watermark or nuanced pattern on the free versions of the images produced.
This serves as an incentive to upsell prospects to the paid version of the service, which produces images without the watermark. Regardless, AI watermarks are one of the easiest things to spy for when trying to identify whether the image is real or not.
As AI technology continues to evolve – rapidly – the tools for detection must continually keep pace because soon the human eye and intuition won’t be enough. As you evaluate the tools you want to consider using going forward, here are some things to keep in mind when it comes to AI detection:
- Understand potential inaccuracies and limitations in current detection methods.
- Always consider the source and context of the image in question.
- Be sure to use multiple tools for cross-verification purposes.
- Stay current with the latest advancements and updates within the AI detection space.
There is no doubt that as the fields of AI generation and detection continue to advance, new and improved tools will emerge on both sides of that ledger. We can only hope they remain in balance.
The next installment will explore tips, tactics and tools to identify written AI content.
