Satya Nadella warns companies using AI: Proprietary AI models are Trojan horses

What happened

Microsoft CEO Satya Nadella published a post on X on July 12, 2026 titled 'The Reverse Information Paradox,' naming a new risk: enterprises paying AI labs for inference unknowingly give the data footprint of their business processes to their vendor. TechCrunch picked up the story on July 13.

Context and impact

Nadella referenced Kenneth Arrow's information paradox from economics and inverted it: the AI lab gains the value of the customer's information with every inference request. He also criticized restrictive distillation terms, arguing that if AI labs scrape the internet to train models, enterprises should have the right to 'distill' models — using model outputs to train their own, cheaper model. Both Anthropic and OpenAI prohibit distillation training in their Terms of Service.

Details

• Concept: 'Reverse Information Paradox' — a new form of proprietary know-how leakage
• Criticism: Anthropic and OpenAI prohibit distillation training in Terms of Service
• Nadella's proposal: distillation rights are a legitimate counterbalance for training on public data
• Warning: every prompt sent to GPT or Claude is a potential data gift to AI labs
• Context: Microsoft is developing its own MAI models as alternatives to external AI services