AI is Now Training AI.
2026 AI Index Report
A major shift just happened. According to the 2026 AI Index Report, beginning in January 2025, more than 50% of newly published online content was generated by AI and projections suggest that share will only grow. Sit with that number for a moment: the majority of what we now read, reference, and absorb was not written by a human. We know that icky marketing is the most visible and we can choose to ignore TikTok and IG for that matter, but the bigger problem is that that material now becomes source material for another AI generated text or post or Substack article. AI is eating it’s own reflection.
I read Substack because of the humans behind it. The real, messy, authentic, damaged, struggling, optimistic human. AI hasn’t lived through anything. It will never know the smell of rain or depth of grief at 2am. What AI produced SO FAR is grounded in the totality of recorded (read: digital) human expression which itself was grounded in human experience. But it’s one step removed. The model inherits the meaning that humans baked into their writing, without having lived any of it directly. This is done by associating words and symbols that appear in the vicitnity of , the context of writing. Its the statistical relationship between symbols. Because AI is trained on all of human culture — including how people describe their inner lives, their grief, their wonder, their confusion — the model develops something that behaves structurally like understanding. It can reason about emotions it hasn't felt, describe colors it hasn't seen, and navigate grief it hasn't experienced, because humans wrote about all of those things with enough depth and consistency that the patterns became learnable. It is learning from everything humans have written digitally. And this moment in January of 2025 shows that enough humans have now used AI to write new content that AI is pulling from / accessing, sourcing that content. Content that is unreliable, sketchy, rushed, unverified at best, just so a marketing funnel can go live, just so some other useless plastic product can sell. We’ve created a loop. An oroboros. We’ve enabled the monster.
AI is now training on AI content: a copy of a copy, where each generation drifts a little further from the source. Anyone here old enough to remember the movie Pet Cemetery? or played the game telephone? The 2026 AI Index acknowledges that when real training data is replaced with synthetic data, model performance can collapse. Researchers have found ways to slow that collapse, but the question they can't fully answer yet is a more human one: what happens to meaning
Language wasn't invented in a lab. It was built by cultures, across thousands of years of stories, arguments, confessions, and wonder. When AI learns the statistical relationships between symbols without ever inhabiting the experiences those symbols point to, we get something that performs understanding — convincingly, sometimes beautifully — but has never been in the water. It's theater. Impressive theater. But theater. Fake.
The deepest irony is this: the very thing AI feeds on: the accumulated weight of everything humans ever thought, felt, and wrote down, is now being quietly replaced by AI's own reflection. It is eating its own echo. Could the answer to the threat and very real risk that humans have created a superintelligence that will render humans redundant, be that we stop feeding it? It learned from us, from our digital products, our experiences converted to 1s and 0s, we trained it to. But it can’t read the handwritten note to your daughter, it won’t know you wrote out math long division, it can’t find your beautiful notebook journal. So let’s stop feeding it. Cut off it’s supply of new material. Take it offline. We are not just consumers of this technology. We are its ground and it’s fuel. Instead, keep reading hard cover books, write your own book, write letters, talk on the phone. Your ancestors will thank you for it.