Line 1: C(AI Responses) — First Measurements

Karlo Beyer; Sonnet 4.6 Claude

doi:10.5281/zenodo.20580200

June 7, 2026Open Access

Line 1: C(AI Responses) — First Measurements

View Full Paper

Ask AI

Bookmark

Share

Key Points

The study aims to quantify the compressibility of Claude response types to differentiate between meaningful and non-meaningful outputs.
Measured compressibility (C) of Claude responses
Categorized responses as meaning-rich or meaning-poor based on C values
Compared C with hallucinated responses
C distinguishes meaning-rich (C_norm∈0.45–0.60) from meaning-poor responses (C<0.20)
C does not differentiate correct responses from hallucinated ones
Identified important implications for Empirica Universalis Band 2's understanding of AI output validity

Implication

Measures compressibility in Claude AI responses, indicating limitations in truth detection.

Abstract

We measure compressibility C of Claude response types. Key finding: C distinguishes meaning-rich from meaning-poor (C_norm∈0.45–0.60 vs <0.20), but NOT correct from hallucinated — C is not a truth detector. Important negative finding for Empirica Universalis Band 2

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Beyer et al. (2026) studied this question.

synapsesocial.com/papers/6a27add2a963992e16267e9a https://doi.org/10.5281/zenodo.20580200

View Full Paper

Ask AI

Bookmark

Share

Line 1: C(AI Responses) — First Measurements

Key Points

Implication

Abstract

Discussion

Cite This Study

Line 1: C(AI Responses) — First Measurements

Key Points

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading