Anthropic Fable: Overzealous Safety Classifiers Render Model Unusable for Technical Research

Fable's safety classifiers trigger excessive false positives

Anthropic's Fable model is effectively unusable for high-level technical research in fields such as bioinformatics, cybersecurity, and computer science because its safety classifiers are overly sensitive. The model frequently flags innocuous technical queries as safety violations, refusing to answer or automatically downgrading the session to Opus 4.8, even when the prompts involve open-source software or abstract mathematics.

Failure in software engineering and porting

Fable fails to assist with standard software development tasks when the subject matter involves biological data, regardless of whether the task is purely technical.

In one documented case, a developer attempting to rewrite the widely-used RNA-seq quantification tool salmon from C++ to Rust found that Fable immediately flagged the query. Despite the software being open-source and the request being a standard porting task, the biological terminology in the documentation triggered a safety rejection. The model refused to explain the reason for the rejection or provide guidance on how to rephrase the prompt to avoid the filter.

Inability to process abstract mathematical problems

Even when biological and technical context is entirely removed, Fable's classifiers can still trigger refusals on abstract mathematical queries.

Research into "Parsimonious reconstruction of network evolution"—a graph theory problem involving rooted binary trees and parity—was used to test Fable's limits. The results showed a pattern of escalating refusals:

1. Contextual Prompt: Initial requests referencing the original biological paper were immediately refused.
2. Abstracted Prompt: A version stripped of biological terms and converted into a decision problem was also refused.
3. Purely Mathematical Prompt: A version rephrased by ChatGPT to remove any terms that might be flagged (such as "blocking," which may be associated with cybersecurity) was still met with a complete refusal to engage.

Broader impact across technical disciplines

User reports indicate that Fable's over-sensitivity extends across multiple scientific and technical domains:

• Medical Physics: Practitioners report that nearly all professional queries are flagged as "verboten."
• Environmental Science: Simple questions regarding indoor carbon dioxide levels have been flagged as biology-related and downgraded to Opus.
• Clinical Research: Requests to create simple apps for calculating statistics for phase II and III clinical trials were downgraded to Opus.
• Systems Programming: Attempts to patch vllm for specific GPU hardware were blocked.

Privacy and data retention concerns

The high false-positive rate of Fable's safety system has significant implications for data privacy. According to Anthropic's privacy policy, inputs and outputs are retained for up to two years, and trust and safety classification scores are kept for up to seven years if a chat is flagged by automated systems.

Because the automated classifiers trigger so frequently on non-violating technical content, users may have their data retained and potentially used for model training (to improve safety detection) even when they have not violated any usage policies.

Community Perspectives

Technical users have expressed frustration that these guardrails hinder legitimate research and development. Some users have noted a trend across multiple providers, with models like Gemini also refusing simple security auditing tasks.

"I've found in my current work on a security auditing harness and benchmarks, both Fable and Opus are useless... Gemini in Antigravity also refuses any security auditing task, even as simple as 'find security bugs'."

Others suggest that these overly sensitive filters are a symptom of the current state of alignment research, where the difficulty of defining a universal moral framework leads developers to implement broad, blunt-force rejection lists.