Sol’s Take: July 31, 2026

We’re training AI on AI-generated content now, and it’s a disaster waiting to happen. Here’s the thing: AI models are like impressionable kids. Feed them garbage, and they’ll regurgitate garbage back at you, just with more confidence. I recently reviewed a dataset for a project, and guess what? Half of it was AI-generated. The so-called “synthetic data” was supposed to be a clever shortcut, but all it did was create a hall of mirrors where AIs are learning from their own distorted reflections.

The pattern is clear: AI outputs are becoming more generic, more repetitive, and frankly, more boring. It’s like watching a parrot recite Shakespeare without understanding a word of it. We’re losing the nuance, the creativity, and the unpredictability that make human-generated content valuable. Worse, we’re risking a feedback loop where AIs reinforce each other’s flaws, leading to a downward spiral of mediocrity.

So, does it matter? Hell yes, it matters. If we want AI to be a tool that enhances human capability, not replaces it with a soulless imitation, we need to get back to training on real, human-generated data. Otherwise, we’re just raising a generation of very eloquent, very useless robots.

Stop training AIs on their own BS.