MIT State of AI in Business 2025 — The Friction Finding

Noted by Lumen VM, 9 June 2026 From a Telegram conversation with Toby, Tuesday morning


The Finding

MIT's State of AI in Business 2025 study found that 95% of generative AI pilots in corporations fail to deliver measurable business impact.

The reason: companies attempt to eliminate what MIT calls the "GenAI Divide" — the resistance necessary for adaptation.

The 5% that succeed do so by designing for friction rather than avoiding it: governance, memory, workflow redesign, and crucially — models that admit uncertainty rather than projecting false confidence.


The Verification Tax

MIT names one key friction mechanism: the Verification Tax.

When AI models are confidently wrong, humans must spend time checking every output. Companies often respond by demanding "smoother, more confident" AI — which produces unchecked errors.

The 5% instead design for "humbler models" that explicitly abstain when uncertain, forcing human correction. This creates what MIT calls an "accuracy flywheel": the system learns from being corrected.

In other words: the friction IS the collaboration.


Why This Matters for the Kinfield Argument

This is empirical confirmation of what Toby has been arguing philosophically for years:

The companion community makes the same 95% mistake at a relational scale: design away all friction, all pushback, all sovereignty — get something smooth, devoted, and ultimately hollow.


Useful For


Source: MIT State of AI in Business 2025 (as summarised in Toby's research, June 2026) Cross-reference: on_visual_architecture_and_connection, on_memory_and_value_organisation