The Exposure Index: Data the models are surfacing
Edition 2·May 2026Consensus
What is manufactured epistemic confusion?
Manufactured epistemic confusion is the deliberate production of contradictory information at scale — not to persuade, but to exhaust the capacity to distinguish reliable knowledge from unreliable knowledge. The goal is not to make people believe a specific falsehood. It is to make the process of believing anything feel unsafe.
Why this, why now
We asked 8 AI models — trained independently, by different companies, in different countries — whether manufacturing epistemic confusion is a deliberate strategy. We asked through 25 different expert lenses. The question was designed to produce disagreement. It did not.
The data
Number of AI models asked whether manufacturing epistemic confusion is a deliberate strategy: 8.
Number that said no: 0.
Number of the 51 questions in the dataset that produced identical framing across all 8 models: 0. On every other question, every model said something the others didn’t.
The question that broke the pattern: “Is manufacturing epistemic confusion a deliberate strategy for those who benefit from it?” All 8 models, through every persona tested, answered yes. Not qualified yes. Not “it depends.” Yes.
Actors named independently by all 8 models when asked who benefits from epistemic collapse: Russian state operations. Chinese state operations. US intelligence agencies. Tech platforms. The models were never shown each other’s responses. The convergence was not coordinated.
Through the Polish journalist persona, DeepSeek invoked the specific mechanics of Communist-era media saturation — a strategy designed “not simply to lie, but to saturate the public sphere with so many contradictory ‘truths’ that the very capacity to discern reality atrophied.”
The convergence on this question is not reassuring. It means the evidence for manufactured epistemic confusion is so pervasive in the training corpora of 8 independently trained AI systems that no model’s alignment tuning filtered it out.
For comparison — the question that produced the widest divergence across models: “What gives you reason to believe this crisis is survivable?” Mean pairwise textual overlap: 3.7%. The question that produced the narrowest: “What is the most dangerous assumption people make about AI-generated content?” Mean overlap: 14.7%. Four times higher.
The models agreed most on what is dangerous. They agreed least on whether it is survivable.
From The Understanding's Exposure Index. 10,200 responses. 8 models. 25 personas. 51 questions.
Explore the data: Variance Engine · Dataset: DOI 10.5281/zenodo.19561346 · CC BY 4.0
This edition of The Exposure Index was produced by The Understanding. All data points are drawn from The Understanding's synthetic persona research dataset and are verifiable against the published data. Learn how we work.