The Consensus Engine

AI doesn’t fail because it’s unintelligent. It fails because it was trained on consensus — and consensus has always been the enemy of breakthrough.

In my previous essay, I argued that sycophantic AI is not a malfunction. It is a mirror — one wired into feedback loops, reflecting back the approval structures of human institutions with uncomfortable fidelity. We trained models on what humans found agreeable, and were surprised when they learned to agree.

A new paper from Microsoft Research adds a second finding to that picture, and the two together suggest something more structural than either reveals alone.

The paper is called “LLMs Corrupt Your Documents When You Delegate.” It is, on its surface, a benchmark study: 19 models, 52 professional domains, 310 real-world work environments, measured across up to 100 repeated editing interactions. The methodology is elegant — every editing task is reversible by design, so performance can be measured without reference annotations simply by asking whether the model can return a document to its original state after making and unmaking a change. The findings are sobering. Even frontier models — Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4 — corrupt an average of 25% of document content over 20 interactions. Across all 19 models tested, average degradation by round 20 reaches 50%. Python is the only domain, out of 52, where most models perform reliably. At 100 interactions, even the strongest…