Anthropic blocks users for anything touching training larger models
Anthropic’s Claude models are increasingly refusing or throttling queries related to training or improving other AI systems, including relatively benign ML and reinforcement learning work. Users argue this reflects aggressive safety and anti-competition policies, hypocrisy around data use and copyright, and a dependence risk when a dominant vendor can unilaterally cut off important workflows. Many describe shifting to OpenAI, Chinese and open-weight models, or local setups as fallbacks, and see stronger openness, portability and decentralization as critical to avoiding future lock-in.
Scope of Claude Restrictions
- Multiple commenters say Anthropic has been restricting help with training larger or competing models for months.
- The cited case appears to be a safeguard-triggered refusal within a session, not a full account ban.
- One factor discussed: user is based in Hong Kong, where Claude is geo-blocked; using VPN plus doing ML training is seen as a likely trigger. Others in US/EU report similar blocks, so it’s not purely geolocation.
User Experiences and Friction
- Some users report smooth help with small-scale ML (e.g., BERT, custom attention, RL for games) and infrastructure (custom training/inference stacks, LoRA adapters).
- Others hit frequent refusals on:
- Model training, data labeling, distillation, and “training a competitor”.
- Low-level systems work (Linux kernel, drivers, hypervisors, firmware).
- Cybersecurity reviews, exploit development, or even security-focused PR review.
- Browser automation, logins, handling credentials, and certain refactors.
- Safeguards are described as session-contagious, opaque, and sometimes silently degrading to older models.
IP, Copyright, and “Hypocrisy” Debate
- Heated debate over Anthropic’s use of web-scraped and pirated datasets (e.g., Books3/The Pile) vs. its efforts to block model distillation or competitive training.
- Some argue this is large‑scale IP theft and ethically incompatible with using Claude; others see web scraping as not clearly an IP violation or excuse some behavior in noncommercial research contexts.
- A lawsuit and large settlement over training on pirated books is cited inside the thread, though details are acknowledged as model‑researched, not independently verified.
Competition, Lock‑in, and Governance
- Many see the restrictions as anti‑competitive, building vendor lock‑in rather than “safety”.
- Warnings that Anthropic is willing to sacrifice entire user segments when interests diverge; users are urged to maintain portable workflows and multiple harnesses.
- Discussion of PBC structure vs. de facto corporate behavior, IPO dynamics, and relationships with defense/surveillance contractors fuels distrust.
Open / Local Models and Safety Philosophy
- Strong push toward open‑weight and local models (Qwen, DeepSeek, Kimi, GLM, Mistral, etc.) as strategic hedge against future lock‑downs, including potential hardware‑level controls.
- Some argue frontier closed models (especially Claude) are still far ahead for complex, long‑horizon coding and reasoning; others claim open or Chinese models are already “good enough”.
- Philosophical split: per‑token/preventive safeguards vs. post‑hoc enforcement (law enforcement after wrongdoing), with many criticizing Anthropic’s aggressive classifiers as self‑defeating and unreliable.