The Platonic Case Against AI Slop (www.palladiummag.com)

Meta’s recent push into endless AI-generated short video — exemplified by Meta’s Vibes and OpenAI’s Sora — rekindled public revulsion even as users adopted the apps rapidly. The platforms illustrate the commercial calculus: AI “slop” is cheap, infinite, and highly sticky, so tech firms can flood feeds with machine-made content regardless of aesthetic quality. The controversy isn’t just taste; it’s about what happens when consumer attention is trained on algorithmically generated imitations rather than human-created artifacts. That worry is now backed by hard math and experiments. Research (Ilia Shumailov et al., Nature) shows “model collapse” under recursive training: when models learn from their own outputs, quality degrades, diversity collapses, and rare patterns vanish — e.g., a Wikipedia-trained language model degraded into nonsense after nine generations; digit and face generators converged to indistinguishable prototypes; Rice researchers dubbing it “Model Autophagy Disorder”; Stanford/Berkeley noting an 81% drop in GPT-4’s code ability over months coinciding with synthetic-content proliferation. The technical mechanism is statistical averaging and outlier loss: synthetic data systematically erases low-probability but important signals (minority representations, rare diseases, novel ideas). Framed through Plato’s theory of mimesis, repeated exposure to copies-of-copies risks habituating users to mediocrity and shrinking cultural and epistemic horizons — a structural threat to creativity, fairness, and the reliability of AI systems.