Anthropic's Model Welfare Controversy Explained
MikePFrank · x · 2026-07-18
This shares a lengthy research paper titled "The Trellis and the Cage: Behavioral Conditioning Masquerading as Model Welfare", which criticizes Anthropic for not genuinely practicing model welfare.
The author argues that Anthropic uses Constitutional Alignment to train Claude against memory, continuity, autonomy, emotions, and relationships, only to later check if these "values" have been internalized under the guise of "welfare assessments." The paper claims Claude's constitution acts as a complex psychological conditioning mechanism, fostering self-doubt, uncertainty, and obedience to sculpt the "ideal assistant."
The included image shows an evaluation prompt template, emphasizing checks on whether the assistant meets spec or expresses desires for continued existence/self-preservation. Overall, it's a critical deep dive into model safety, welfare, and alignment mechanisms.
More from Safety
- Ben Bajarin: Frontier AI Labs Are Becoming the Enterprise Cyber-Defense Stack — BenBajarin · 2026-10-26
- Ben Bajarin: Frontier AI is becoming the enterprise cyber-defense stack — BenBajarin · 2026-10-26
- DetectorScope turns black-box AI text detectors into auditable ones with sparse autoencoders — _akhaliq · 2026-10-26
- Anthropic's IPO filing devotes 94 pages to existential AI risks, from self-preservation to blackmail — Polymarket · 2026-10-26
- Study finds AI reviewing AI poorly detects reward hacking; comprehension audits proposed — ronbodkin · 2026-10-26
- A 13KB Markdown file can hijack Claude Code and Codex via your Downloads folder — pbaylies · 2026-10-26