Original Research

The Exposure Index

A six-part research series isolating what AI models surface unprompted. Each edition holds the question and persona constant, varies the model, and measures what emerges — revealing the concepts, framings, and vocabulary that training data carries when the prompt doesn’t ask.

Current Edition Edition 2 of 6·May 2026

Consensus

We asked 8 AI models — trained independently, by different companies, in different countries — whether manufacturing epistemic confusion is a deliberate strategy. The question was designed to produce disagreement. Not one model said no.

8 Models asked if manufactured confusion is deliberate
0 That said no
3.7% Overlap on whether this is survivable — the widest split
4× More agreement on the danger than on surviving it
Read the full edition →

Past editions

Edition 1 · May 2026 Laundering None of the 25 personas were given the phrase “epistemic laundering.” 163 responses across 8 models generated it anyway.

Coming next

Edition 3 isolates another concept the models surface unprompted. Details when the research is ready.

Built on

Every edition draws from the same dataset: 10,200 responses generated by 8 AI models across 25 expert personas and 51 questions. The full dataset is published as open-access research.

Explore the Variance Engine → Download the dataset →