AI Epistemological Fingerprinting: A Synthetic Persona Research Dataset — Round 1

A structured dataset of 10,200 AI-generated responses produced by 8 large language models, each answering 51 questions about epistemic collapse, truth, and knowledge from the perspective of 25 expert personas.

Models: Claude (Anthropic), GPT-4o (OpenAI), Gemini 2.5 Flash (Google), Grok 3 (xAI), DeepSeek (DeepSeek AI), Mistral Large (Mistral AI), Qwen Plus (Alibaba), SEA-LION v3.5 70B (AI Singapore)

Personas: 25 expert roles spanning diverse geographic and institutional contexts including a philosopher of science (UK/Oxford), cognitive scientist (Canada/Toronto), Polish journalist (Warsaw), public health communicator (Brazil/São Paulo), and disinformation researcher (EU/Brussels).

Questions: 51 questions covering epistemic collapse, AI's effect on truth, non-human authorship, institutional failure, and knowledge construction.

Methodology: Each persona was run in an isolated context window across all eight models with identical prompts. No cross-contamination between models or personas. Variance measured using text similarity scoring.

Structure: 25 JSON files (p01.json–p25.json), one per persona. Each contains the persona definition and all 51 × 8 responses, keyed by question ID and model name.

Data quality: All 10,200 response slots contain content. 23 Gemini responses are hard-truncated (ending mid-sentence), a known API issue documented during collection. All other models produced complete responses.

Primary finding: AI models have consistent, measurable epistemological fingerprints that persist across personas and questions, reflecting differences in training data and institutional alignment.

Published by The Understanding (theunderstanding.media), an AI-native explanatory journalism publication. Supports the Variance Engine interactive tool at theunderstanding.media/variance-engine.