You open up your favorite AI chatbot. You pour your heart out, ask it to draft a sensitive email, or beg it to debug a messy script. You think you’re having a private, one-on-one conversation with a specific brand. You’re not. You’re just an unpaid data-entry clerk for a rival AI lab.
Recently, a security researcher dropped a bombshell on X: Moonshot—the company behind the Kimi AI—was caught routing user prompts to Anthropic’s Claude instead of its own model. And the kicker? They were harvesting those user-Claude exchanges to train Kimi.
The internet immediately cried foul. The comments ranged from a dismissive ‘So?’ to outrage over the ethical boundaries of model distillation. It feels dirty. It feels like a betrayal. But if you look closely at the architecture of the modern AI industry, this isn’t a glitch. It’s the logical endgame.
In the AI wars, every lab is a customer, a competitor, and a parasite all at once.
Moonshot using Claude is simultaneously the highest compliment and the most lethal threat. It’s an admission that Claude is currently smarter than Kimi, paired with the ruthless intent to use Claude’s brain to build Kimi’s replacement. It’s the digital equivalent of befriending a genius just to copy their homework.
But let’s drop the faux outrage for a second. OpenAI and Anthropic scraped the entire internet—every blog, every copyrighted book, every forum—without asking a single human for permission. They built trillion-dollar empires on our collective, unpaid labor. They normalized the idea that ‘permissionless scraping’ is just how the game is played.
But the moment another AI company starts siphoning off their outputs, it’s suddenly a scandal? Please. The big players are just mad that the parasite is now feeding on them.
There are no pristine datasets left on the internet. The only fresh meat left is us.
The dirty secret of the AI race is that high-quality training tokens are running out. Once AI-generated text saturates the web, ‘distilling’ a frontier model becomes indistinguishable from ordinary training. The only difference is whether the data leak is explicit and therefore scandalous, or hidden within a sea of synthetic web traffic.
You think you’re talking to Claude, but you’re actually feeding the machine that will kill him.
We used to be the users. Now, we are just the conduit. Your prompts are the raw material being passed from one model to another, a pawn in a cannibalistic war between machines. The AI industry isn’t building a future for you. It’s just eating itself, and you’re the fork.
FAQ
Q: What's wrong with using a better model to train your own? Isn't that just good business?
A: It's a parasitic loop. If you just copy the host, you stop innovating and eventually degrade the quality of the entire ecosystem. It creates an inbred model where errors compound over time.
Q: Does this mean my chat history is being sold to rivals?
A: Not sold, but harvested. If you're using a wrapper, a free tier, or an unverified tool, your data is almost certainly being used as fuel for a competitor's training run. You are the product and the raw material.
Q: Everyone is mad at Moonshot, but shouldn't we be mad at the whole industry?
A: Exactly. The big players built their empires on unpaid scraping of human data. Moonshot is just doing to the big labs what the big labs did to the internet. It's karma, served cold.