AI's Persistent Problem with Drawing Human Fingers
A Vox video investigates why AI image generators struggle to accurately render human hands and fingers. Host Phil Edwards consults an artist, a robotics student, and a generative art teacher who has taught since 2018. They identify three main causes: a scarcity of high-quality training images of hands online, the AI's lack of understanding of how fingers move and oppose each other, and the absence of error detection in AI systems. A potential solution, partially implemented in ChatGPT 5.0, involves training the machine to recognize erroneous outputs through user feedback and expanded datasets. The video highlights a telltale flaw that still distinguishes AI-generated images from real ones.
Key facts
- AI image generators cannot correctly reproduce the number and positioning of human fingers.
- Vox published a video investigating this issue, hosted by Phil Edwards.
- An artist explained that drawing hands begins with observation and understanding their function.
- A robotics student and a generative art teacher (teaching since 2018) were also consulted.
- Three reasons were identified: lack of quality examples, lack of understanding of finger movement, and no error margin or bias in AI.
- ChatGPT 5.0 attempts to solve the problem by training AI to recognize errors via user feedback.
- The article was published on Artribune in April 2023.
- The flaw currently allows viewers to identify AI-generated images.
Entities
Artists
- Phil Edwards
Institutions
- Vox
- Artribune