Nadella: You Pay for AI Twice — Once With Money, Once With Your Knowledge
Microsoft's CEO calls out AI labs for banning distillation while training on everyone else's data. His 'reverse information paradox' is worth understanding — even if Microsoft has its own reasons for pushing it.
Microsoft CEO Satya Nadella has published a pointed blog post accusing AI labs like OpenAI and Anthropic of a double standard: they train their models on the world’s public data — claiming fair use — while banning everyone else from learning from their models’ outputs.
The practice in question is called distillation: training a smaller, cheaper model on the answers of a bigger one, so the student model absorbs much of the teacher’s ability at a fraction of the cost. Major labs prohibit it in their terms of service, a rule aimed mostly at Chinese AI companies accused of copying American models on the cheap. Nadella calls the situation “ironic”: the same providers that hoover up public web data, ban distillation, and learn from customer interactions end up concentrating the economic value with themselves — the infrastructure operators — instead of the companies that generate the knowledge in the first place.
He gives the dynamic a name: the “reverse information paradox.” In his telling, companies pay for AI twice. First with money, then with what he calls the “exhaust” — the corrections, ratings, and usage patterns from working with AI systems, which quietly reveal internal company knowledge. “In consuming intelligence, you are creating intelligence,” he writes. AI providers can learn from all of it and, in theory, eventually build competing products.
What’s behind this? Nadella has a point, and he has an agenda — both can be true. The point: the flow of knowledge in AI really is one-directional right now, and terms of service do treat “we learn from you” and “you learn from us” very differently. The agenda: Microsoft happens to sell exactly the remedy it implies — infrastructure where companies control their own data and “learning loop.” It’s also the same Microsoft that is currently phasing OpenAI and Anthropic models out of its Copilot products to cut costs. When a CEO writes philosophy, it usually has a sales funnel attached.
What this means for you: For most private users, not much — your chat corrections aren’t trade secrets. But if you use AI at work, Nadella’s framing is a genuinely useful lens: every correction you type teaches the provider something about how your business works. Worth checking what your company’s AI subscriptions actually say about training on your data — most business tiers let you opt out, and that checkbox matters more than it looks.
Sources
Source: https://snscratchpad.com/posts/reverse-information-paradox/
OpenAI's New Prompting Guide: Stop Overthinking, Just Describe the Result
OpenAI has published a prompting guide for everyday users, not developers. Four optional building blocks, a few guardrails, and one core message: say what you want, not how to get there.
