AI Systems Asked to Generate Deleted Photos Reveal Training Data Paradox
**UNKEPT (2026)** explores a fundamental contradiction in AI image generation.
Artist Kevin Abosch prompted three AI systems—ChatGPT 4o, FLUX.2 Flex, and Grok Imagine—to create "unkept photographs": accidental shots, blurry frames, images meant for deletion.
**The core insight**: Every "throwaway" image in training data was actually kept. Truly deleted photos leave no trace for AI to learn from.
Each system produced different interpretations:
- Body parts blocking the lens (fingers, hands, feet)
- Messy room details (tangled cables, baseboards, shoes)
- Life detritus (receipts, notes, bottles)
The work highlights an **unverifiable loop**: AI can only simulate discarded images from what survived. The "unkept" is built entirely from the kept.
This unverifiability isn't a flaw—it's the point. The project reveals how AI training data shapes what machines can imagine, even when asked to picture absence.