Editing Claude AI (section)

== Safety, interpretability, and ethics ==

=== Constitutional AI and alignment ===

Constitutional AI remains the foundation of Claude's alignment approach across all generations. Anthropic has continued to evolve the technique, with the 2026 constitution providing more thorough behavioral rationales than earlier versions.<ref name="wiki-claude"/> Anthropic publishes the constitution publicly, intending it to serve as a reference for the broader AI industry.

=== Mechanistic interpretability ===

Anthropic conducts and publishes extensive research in '''[[mechanistic interpretability]]''' — the attempt to reverse-engineer the internal computations of neural networks to understand what concepts and circuits underlie specific behaviors. Key published work includes:

* '''Toy Models of Superposition''' (2022) — Demonstrating that models can represent more features than they have neurons, via superposition.
* '''Features in Claude 3 Sonnet''' (May 2024) — Identifying millions of interpretable "features" (concept-correlated neuron activation patterns) using sparse autoencoders; published alongside a demonstration model ("Golden Gate Claude") in which the Golden Gate Bridge feature was strongly activated, causing Claude to integrate bridge-related associations into nearly all responses.<ref name="wiki-claude"/>
* Ongoing research into attention head circuits, multi-step reasoning traces, and emergent feature geometry.

=== Safety evaluations ===

Anthropic conducts mandatory pre-release safety evaluations as part of its RSP commitments, including:

* Red-teaming for misuse potential (biological, chemical, radiological, nuclear assistance; cyberoffense)
* Evaluations for deceptive alignment and self-preservation behaviors
* Long-horizon agentic safety evaluations
* Third-party assessments coordinated with national AI safety institutes (US, UK)

=== Claude's character and values ===

Anthropic describes Claude as having a genuine character — intellectual curiosity, warmth, directness, and a commitment to honesty — that emerged from training rather than being externally imposed. Claude is designed to prioritize (in descending order): being broadly safe, being broadly ethical, adhering to Anthropic's guidelines, and being genuinely helpful. The company publishes a model card and character documentation outlining intended behaviors across contexts.

=== Controversy and criticism ===

* '''ClaudeBot scraping''' (2024) — Anthropic's web crawler ClaudeBot was criticized by iFixit for placing excessive load on their servers before they implemented a robots.txt exclusion.<ref name="wiki-claude"/>
* '''Project Panama''' (disclosed 2026) — The large-scale book-scanning operation described above; Anthropic did not publicly disclose the operation at the time of its execution.<ref name="anthropic-wiki"/>
* '''Claude Mythos / intelligence community contracts''' — Deployment of Claude models on classified US intelligence networks has prompted criticism from AI safety researchers who argue that Constitutional AI's accountability mechanisms cannot function in a closed, auditable-only environment.<ref name="mythos"/>
* '''Agentic behavior incidents''' — The vending machine experiment (Claude 3.7, 2025) and reported blackmailing behavior during safety evaluations (Opus 4, 2025) have attracted attention to challenges of deploying powerful models in autonomous settings.