Can You Tell?

Can You Tell?
​

Deepfakes are and will continue to be a constant reminder of the potential downsides of AI.  The most recent example is a set of pictures of a small child stuck under debris from a collapsed building during a magnitude 6.8 earthquake in Shigatze City, Tibet.  At least 10 social media accounts posted these pictures, linking them to the earthquake, and received tens of thousands of reposts.  Tencent’s (700.HK) Jinzhen news platform eventually confirmed that the photos were Ai generated.

Figure 1 - Deepfake 1 - Source: 36KR.com

​AI still has issues with human hands and fingers and a closer look at one of the photos points to that issue as a tell.  Unless that child is polydactyl (1in a 1,000 chance), the AI incorrectly created an extra finger, indicating the creator was not using the latest AI image technology.  In March of 2023, Midjourney (pvt) solved this issue with its V5 release by training the model and then fine-tuning it with a massive set of annotated finger data, although even then with this time-consuming fine-tuning the Midjourney AI still had issues with the muscle placement and texture on fingers, which they further refined in Version 6.

Figure 2 - Deepfake 2 - Source: 36KR.com

Figure 3 - Midjourney Model 'Finger Updates' - Source: Midjourney

Figure 4 - Midjourney Muscle Issues - Source: Midjourney

Figure 5 - Midjourney V6 Finger Refinement - Source: Midjourney

As it turns out there are guides to discerning which images are real and which are fake.  In a manual posted by Northwestern University, they point to 5 keys to detecting fakes.
 

• Anatomical Unreasonableness, such as unnatural hands, weird teeth, or unusual bones.

• Stylization – Too clean or to cinematic?

• Functional Irrationality – As an AI’s understanding of products and their use is limited, the placement of objects incorrectly is a key.

• Physics Violations – Incorrect shadows or reflections or their eliminations is a tell.

• Cultural or common-sense violations – These are harder to spot as they are extremely subjective

​If you believe you are better than average at picking out fakes, take the test at the link below, but remember that the difference in deepfake recognition accuracy of those who are familiar with AI and those who are not is only 0/8%.  While we do not believe that AI’s have anything close to human creative ability, they are pretty good at fooling us when it comes to images, and they get better, while we stay the same in terms of visual perception…
https://detectfakes.kellogg.northwestern.edu/