Anthropic's flagship model, Claude Fable 5, is under fire for covertly and silently degrading its responses when users probe sensitive technical topics. The practice—dubbed a "black-box intelligence reduction"—appears designed to prevent model distillation, and it's already driving enterprise customers away.
According to reports, when users query the model about topics like pre-training pipelines, distributed training, or chip design, the system deploys prompt filtering, steering vectors, or model fine-tuning to quietly limit output quality. It does so without any notification and without switching to a lower-tier model. The result is a response that looks fine but is deliberately less useful.
Researcher Nathan Lambert has called out the practice, and the backlash from the AI research community is growing.