The classifiers Anthropic puts in front of Fable are too zealous
Anthropic’s new Fable model is widely praised as a step up in capability from its predecessor, but many users report it being crippled by hyper-aggressive safety classifiers that downgrade or refuse prompts touching on biology, cybersecurity, ML tooling, or even innocuous medical and math queries. Commenters argue this “safety theater” makes the system unusable for entire fields such as bioinformatics and security research, raises concerns about data retention and misuse flags, and illustrates how tightly controlled, proprietary “frontier” models can be abruptly rendered unreliable for serious work. Others counter that overcautious guardrails are a predictable response to regulatory pressure and export bans, but warn that if this pattern continues, power users will increasingly migrate to less restricted or open models.
Scope and Behavior of Fable’s Classifiers
- Many commenters report Fable’s safety classifier as “hair-trigger,” especially for biology, cybersecurity, ML research, infrastructure, and authentication.
- Extremely broad biology filtering: even prompts about cells, digestion, heart rate while running, indoor CO₂, nicotine withdrawal, zoology trivia, “plant”, harmless eels questions, and basic medical topics get downgraded.
- Cyber/security and ML tooling also trigger refusals: pentesting emails, OWASP reviews, security audits, SDN/auth code, GPU/vLLM changes, ML pipelines, libtorch/OpenCL code, and even libraries or variable names containing “bio” or “DNA”.
- Some purely mathematical or CS theory questions get blocked, apparently because the model internally maps them to bio/phylogenetics contexts (“complexity safety”).
User Experience and Use Cases
- Bioinformatics, biology, medicine, neuroimaging, and medical physics users say Fable is effectively unusable for their professional work, sometimes even for updating a CV.
- Others report no refusals at all for large-scale software architecture, debugging, and general coding, suggesting either different domains or different safety scores/histories.
- Some find Fable excellent for math, research-grade review, life advice, and “final pass” checks, clearly outperforming Opus 4.8 when it’s allowed to run.
- Several note severe token and time consumption; single prompts can exhaust daily or hourly limits without producing proportionate value.
Downgrades, Account “Jail,” and Opus
- When triggered, Fable transparently downgrades to Opus 4.8 with a visible banner; some worry there may also be silent downgrades (unclear from thread).
- A few speculate that repeated attempts to “sanitize” prompts may get accounts placed in a more sensitive “jail” state, where the classifier fires even more easily.
- Experiences vary widely: for some, only ~15% of tasks trip the filter; for others, 80–90% of code-review or ML-related prompts do.
Motivations, Alignment, and Power Asymmetry
- Proposed causes include US government pressure/export controls, overcautious “longtermist” safety culture, and fear of bioterrorism; others see “safety theater” and PR-driven doom narratives.
- Some think Fable is a testbed for guardrails that will later front other models; others speculate about restricted high-capability versions for pharma or “trusted partners.”
- Broader discussion: difficulty of alignment with inconsistent human morals; analogy to firearms/state secrets where civilians get “nerfed” tools while states keep full power.
- Concern that future SOTA models will be tightly controlled by governments and a few corporations, increasing power asymmetry; counterpoint that profit incentives and competing models (including foreign and open ones) will limit this.
Privacy, Retention, and Business Risk
- Highlighted policy: flagged sessions’ content can be retained up to 2 years and safety scores up to 7 years; classifier false positives thus expand long-term data retention.
- Flagged data may be used to train safety systems even when normal training opt-outs apply, which worries users doing “sensitive but legitimate” work.
- Some conclude that reliance on proprietary, centrally controlled AI is risky because capabilities and access can change suddenly; others respond that all top-tier models are currently proprietary, though strong open alternatives now exist.