To get much out of AI, a company has to give it context: internal documents, research, code, and the reasoning of its own experts. The more it shares, the more useful the AI becomes, and the more of its knowledge leaves its direct control. In 2026, Microsoft CEO Satya Nadella called this the "Reverse Information Paradox", building on an idea the economist Kenneth Arrow described in 1962.
In this study I look at how the paradox shows up when companies adopt AI, and why it's a real strategic risk. I go through two case studies: the controversy around AI and the Navier–Stokes problem in mathematics, and The New York Times' lawsuit against OpenAI and Microsoft. Then I ask whether open-weight models, which a company can run on infrastructure it controls, can help.
What I found
- The gap between the best open-weight and closed models fell from about 8% to 1.7% between January 2024 and February 2025, according to Stanford's 2025 AI Index. That makes open-weight models a credible option for a lot of professional work.
- Running a model in-house keeps prompts, documents and corrections inside the company. It also lets the company adapt the model to its own data and makes costs more predictable.
- The trade-off is responsibility: security, compliance, support and upgrades all fall on the company.
- Most companies will probably end up using both: open-weight models for sensitive work, and closed models where they're clearly better. The question for each task is how sensitive, complex, costly and regulated it is.
Trying it myself
I've started running Qwen3 8B, an open-weight model, locally on my MacBook Pro with Ollama. I connected it to n8n to build a workflow that analyzes emails, picks out important tasks and extracts deadlines, all on a model I control.