When your staff use a tool like ChatGPT or Claude for work, they are not just using it. They are feeding it. Every question asked, every document uploaded, every correction made when the AI gets it wrong: all of that teaches the AI more about your business, your clients, your pricing logic, and your processes.
This week, Microsoft CEO Satya Nadella made that concern explicit in a widely circulated post. He called it the "Reverse Information Paradox." The idea is simple: you pay for an AI tool with money, but you also pay for it with your company's knowledge, and you may not realise the second payment is happening.
Nadella's complaint goes further than data privacy. He is pointing at something structural. AI providers train their models on the internet's public data, which they defend as fair use. But when those same providers restrict others from doing anything similar with their models, while also holding the right to learn from every customer interaction, the arrangement only benefits one side.
His specific target, though unnamed, is fairly clear. Anthropic recently required users of its most powerful model to accept a 30-day data retention clause. That means your interactions are stored and potentially used for training. Separately, Anthropic lobbied the US government to restrict a practice called distillation, where one AI learns from another's outputs. Nadella's point: you cannot argue that the world's data is fair game for your training, then also argue that your own outputs are off-limits to everyone else.
Here is why this matters beyond the politics of Silicon Valley. If AI providers learn from their most active enterprise customers, they accumulate a detailed picture of how entire industries operate. Over time, the AI itself becomes the expert on your sector. The knowledge that makes your firm different from competitors could gradually be absorbed into a general-purpose tool available to everyone, including those competitors.
The shift already underway in large companies is telling. A survey across more than 10,000 organisations found that 76% of companies using AI are now choosing open source models, often alongside proprietary ones. Larger companies, in particular, are moving toward open source models installed on their own systems, where data never leaves their building. The reasoning is practical: an open source model running on your own infrastructure can handle roughly 90% of what the big commercial tools can do, at a fraction of the cost, with full control over your data.
The nuance worth knowing: running your own AI is not free. You still need hardware, technical staff to manage it, and ongoing maintenance. Open source restructures costs rather than eliminating them. But for organisations in regulated industries like insurance, healthcare, or finance, where data sovereignty is already a daily concern, the calculation is often worth it.
Nadella's practical advice is to build what he calls a "proprietary learning environment," a setup where your AI interactions stay inside your own infrastructure and the knowledge your staff generates while using AI belongs to your firm, not to the tool provider. He also recommends building a way to switch between different AI providers easily, so you are never entirely dependent on one.
The irony of all this is obvious: Nadella runs a company that has invested heavily in both OpenAI and Anthropic. Microsoft holds roughly 27% of OpenAI and recently agreed to invest up to $5 billion in Anthropic. He is, in a sense, warning customers about the risks posed by companies he owns a stake in. The more charitable reading is that he sees the long game: if enterprises lose trust in AI providers, they move to self-hosted open source models, and that infrastructure still runs on Microsoft's cloud. Either way, Azure wins.
For any organisation using AI tools today, the question worth asking internally is simple: do your current AI contracts let the provider learn from your data? If the answer is yes, or if you are not sure, that is worth finding out before your next contract renewal.