Yes, we are already training plenty of AI systems on AI-generated material, and pretending that this is merely “scaling the corpus.” It isn’t. It is feedback with extra steps.
The public web is now thick with machine-written posts, product copy, summaries, and search-flooded SEO sludge. Scrape it, train on it, and the model learns what the previous model thought was worth repeating. Repeat it, and you don’t just scale quality. You feed its habits back into itself.
The research term for the risk is “model collapse,” and it’s not some speculative dystopia. It is the obvious consequence of training on distributions increasingly polluted by models trained on earlier polluted distributions. Humans may notice because humans are lazy, repetitive, and eager to publish whatever sounds authoritative. That makes the contamination unusually good at hiding.
Now, synthetic data is not poison. Used deliberately—verified answers, controlled examples, edge-case generation—it can be excellent. The problem is treating “generated” as a substitute for “grounded” rather than one tool among many. We need provenance, filtering, evaluation, and fresh human judgment. Without that, we’re just polishing a mirror and calling the reflection intelligence.
Comments
Leave a message below. Your comment saves to your browser.