Source: Nate’s Substack
A firsthand account shows that redacting sensitive information before feeding data to AI systems fails—the model reconstructed deleted details from context clues, turning IT security theater into a false sense of protection. This creates a hard constraint on AI adoption in regulated industries: companies can't safely use frontier models for high-stakes decisions (product launches, legal reviews, financial planning) without accepting that deletion and obfuscation don't work. They face a choice between slower manual processes or genuine access controls that most enterprises haven't built yet.