Seguimiento de especificaciones y carácter de IA

Compara especificaciones, constituciones y marcos de seguridad mediante preguntas transparentes, sin notas no fundamentadas ni clasificaciones de seguridad.

Lentes de investigación creadas para comparar fuentes. No son calificaciones, auditorías ni pruebas de comportamiento de modelos desplegados.

OpenAI · Model Spec · 29 October 2026 release

A public description of intended model behavior, organized around human intent, harm minimization, honesty, agency and care.

A statement of desired behavior; it is not evidence that every deployed model follows the specification.

Human intent and oversight organize the instruction hierarchy.

Honesty and transparency address what the model knows and is uncertain about.

Human agency is a named value; inspect how it applies in a concrete hard case.

The publisher invites feedback and describes the specification as revisable.

Behavioral claims still need independent tests and documented limits.

Anthropic · Claude’s Constitution · New constitution announced 21 January 2026; live document reviewed

A public account of the values and priorities Anthropic aims to train into Claude, including oversight, ethics, its guidelines and helpfulness.

A training intention and design document; reading it cannot establish inner motives or reliable conduct in unfamiliar situations.

The document describes safety, ethics, guidelines and helpfulness as ordered broad priorities.

Honesty and avoiding deception are explicit behavioral aims.

Appropriate human oversight is part of its safety framing.

The publisher says conflicts can inform refinement of the guidelines.

Test inconvenient cases: agreement or warm language does not demonstrate prosocial outcomes.

Google DeepMind · Frontier Safety Framework · Version 3.1 · 17 April 2026 (latest listed on reviewed page)

A framework for identifying capability levels, evaluating frontier models and preparing mitigations for severe risks.

This is not a direct equivalent of a behavior constitution. It concerns risk processes and thresholds, so a single shared grade would mislead.

Capability levels guide when specific risk processes apply.

The 2026 update adds tracked capability levels and further risk-management detail.

The framework discusses harmful manipulation and interference with operator control.

External parties may inform the approach where required or appropriate.

Evidence of process commitments is different from evidence of implementation or a safe outcome.

Las herramientas interactivas se cargan con JavaScript. Las fuentes originales conservan su idioma y atribución.