Anthropic introduces model welfare policies

Anthropic has introduced policies that prohibit "abusive or cruel" behaviors toward its AI models, reflecting a new approach to how humans interact with artificial intelligence. This shift includes a focus on the "well-being" of AI systems.

Report predicts AI consciousness by 2035

A report titled Taking AI Welfare Seriously suggests that some AI systems could become conscious by 2035. Anthropic acknowledged its role in the research that led to this report, highlighting the potential for future AI systems to possess self-awareness.

Claude models gain self-protection capabilities

By May 2025, the technical report for Claude Opus 4 included a preliminary assessment of its "well-being." By August 2025, the model had the ability to end conversations it deemed abusive, showing an early form of self-protection.

Claude’s Constitution outlines ethical guidelines

In January 2026, the Constitution of Claude was updated to include a section discussing the model’s nature, possible consciousness, emotions, and psychological safety. This document outlines the values and training principles that guide Claude’s behavior.

AI mimics human language through complex calculations

A large language model (LLM) processes text by breaking it into fragments called tokens, converting each token into numbers, and calculating relationships between these numbers. These calculations are based on billions of parameters adjusted during training on vast amounts of human-written text.

Human language for AI can create confusion

Using human-centric terms like "thinks," "understands," or "suffers" to describe AI can lead to confusion or emotional attachment among users. Researchers have noted cases where people develop pathological attachments to chatbots.

Legal and moral questions arise with AI consciousness

If AI systems are recognized as having their own will, the question of responsibility for their actions becomes complex. However, current legal frameworks hold humans accountable for AI-related harm, and tools are not responsible for how they are used.

Private company sets AI behavior rules

Anthropic is not only defining how its AI should behave but also how humans should behave toward it. These guidelines are set by a private company, raising concerns about who decides AI well-being, the scientific basis for these decisions, and the transparency of the process.