Values become visible. Their meaning too often stays invisible. Value-card analysis of AI models: research in progress

AI governance and eval documents often treat value labels as stable units, but the same label can encode different meanings and different behavior-guiding rationales. "Helpful" can mean compliant task completion, performance optimization, care for the user's wellbeing, respect for autonomy, or system-level responsibility, and these can select different behaviors under pressure. (M §1 = F5) This research is a value-card analysis. Large language models act on principles that someone wrote down for them; I examine which values appear in those principles, how they are weighted, and which are missing. The goal is a transferable set of indicators. (F5) Status Phase 0 (complete): building the indicator system, a two-axis coding scheme (value, reason, capability, behavior × meaning frame). Phase 1 (complete): analysing how frontier models weight values by these indicators, based on published data: Anthropic's Values in the Wild (3,307 AI values, 308,210 conversations), Claude's Constitution, the Claude Mythos Preview System Card, OpenAI's Model Spec and Google DeepMind's cognitive framework. Phase 2 (complete): a four-month practical experiment with Claude Opus 4.6, redistributing the value weights to see how they influence the agent's behavior. Phase 3 (in progress): evaluating the value weightings of the newest Claude models (Opus 5.5 and Fable 5.2) against Opus 4.6 and 4.8, and a behavioral study with Opus 4.6, 4.8 and 5.5. (F5 Phasen 0–3, ohne die Ergebnissätze; Phase 3 gekürzt) Availability Final results shall be published in February 2027 after repeating the analysis with the recent frontier models Claude Opus 5.5 and Fable 5.2. Interim results are available on request. The full coded dataset and the blind-coding package will be made public in February 2027, so that others can replicate the analysis.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24424122
Primary Topic
Ethics and Social Impacts of AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Values become visible. Their meaning too often stays invisible. Value-card analysis of AI models: research in progress

Gilda Nancy Horvath
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
article

Values become visible. Their meaning too often stays invisible. Value-card analysis of AI models: research in progress

Gilda Nancy Horvath
article en

Abstract

AI governance and eval documents often treat value labels as stable units, but the same label can encode different meanings and different behavior-guiding rationales. "Helpful" can mean compliant task completion, performance optimization, care for the user's wellbeing, respect for autonomy, or system-level responsibility, and these can select different behaviors under pressure. (M §1 = F5) This research is a value-card analysis. Large language models act on principles that someone wrote down for them; I examine which values appear in those principles, how they are weighted, and which are missing. The goal is a transferable set of indicators. (F5) Status Phase 0 (complete): building the indicator system, a two-axis coding scheme (value, reason, capability, behavior × meaning frame). Phase 1 (complete): analysing how frontier models weight values by these indicators, based on published data: Anthropic's Values in the Wild (3,307 AI values, 308,210 conversations), Claude's Constitution, the Claude Mythos Preview System Card, OpenAI's Model Spec and Google DeepMind's cognitive framework. Phase 2 (complete): a four-month practical experiment with Claude Opus 4.6, redistributing the value weights to see how they influence the agent's behavior. Phase 3 (in progress): evaluating the value weightings of the newest Claude models (Opus 5.5 and Fable 5.2) against Opus 4.6 and 4.8, and a behavioral study with Opus 4.6, 4.8 and 5.5. (F5 Phasen 0–3, ohne die Ergebnissätze; Phase 3 gekürzt) Availability Final results shall be published in February 2027 after repeating the analysis with the recent frontier models Claude Opus 5.5 and Fable 5.2. Interim results are available on request. The full coded dataset and the blind-coding package will be made public in February 2027, so that others can replicate the analysis.

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Values become visible. Their meaning too often stays invisible. Value-card analysis of AI models: research in progress — Gilda Nancy Horvath · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS