When AI Systems Don't Know When to Stop: A Comparative Study of Spontaneous Boundary Declaration and Epistemic Limits in Six Major Models

Read full text Summarise with VIRAT
Authors: Hernán Alfredo Capucci
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This study examines how six major AI systems — Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google), Grok (xAI), Copilot (Microsoft), and Meta AI (Meta) — describe their operational role, capabilities, and limits within a multi-model synthesis workflow. Using a standardized prompt under two experimental conditions — spontaneous self-description (Phase 1) and forced boundary declaration (Phase 2) — the study identifies a consistent empirical pattern: no system spontaneously declared stopping conditions. When explicitly prompted, all six systems produced structured boundary definitions, demonstrating that the capability exists but is not activated by default. Phase 2 responses revealed three categories of stopping logic:- Safety-based (policy-driven)- Operational (capability-driven)- Epistemic (reasoning-driven) Only two models produced epistemic stopping conditions. A secondary observation emerging from the research process documents a case of undeclared self-inclusion, where an evaluated system participated in the analytical framework without declaring itself as a subject. These findings highlight a structural gap between generative capability and boundary declaration, supporting the need for pre-execution declarative frameworks. Agent Manifest v1.0 is presented as a structural response to this gap.

Fields of Study & Subjects

Ethics and Social Impacts of AI Artificial Intelligence Applications Topic Modeling

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
26 March 2026
Language
ENC