Author Andrew Grush reports growing frustration with Claude’s increasingly restrictive safety guardrails, noting that the AI frequently misinterprets creative writing prompts and benign inquiries as controversial or harmful. In a recent test of user sentiment, 35% of respondents agreed the chatbot has become overly preachy and difficult to work with, while 41% reported no issues. The author highlights a specific conflict regarding a fictional story concept involving religious themes; despite clarifying the intent was purely creative, Claude refused the request on the grounds that it did not want to present potentially sensitive historical interpretations as fact. The inconsistency of these refusals is a primary concern, as the author successfully executed the identical prompt in a separate conversation session, illustrating an unpredictable application of safety protocols.

The author observes that these issues persist across multiple models, including Sonnet 5, Opus 4.8, and Sonnet 4.6, with some variants showing higher sensitivity than others. He notes that Opus 4.6 and Opus 4.7 have exhibited fewer problems, allowing for more fluid interactions even on sensitive topics. This defensive behavior intensified following the restoration of Fable 5, which had been temporarily taken offline after government entities identified it as a security risk. The author posits that Anthropic likely tightened safety constraints across its model suite to mitigate further regulatory or security concerns.

Beyond creative writing, the author experienced pushback on various topics, including financial planning and hypothetical biology discussions related to a star system 10.5 light-years away. In these cases, models like Opus 4.8 often narrowed the scope of answers to ensure safety, treating hypothetical scenarios as if they were real-world plans. While the author continues to value Claude for its superior memory and long-form conversational capability compared to competitors like Gemini, the increasing need for highly specific prompt engineering has diminished the user experience. Despite these challenges, he notes that many users still interact with the platform daily without encountering these specific barriers, suggesting that the frequency of pushback is heavily dependent on individual usage patterns and prompt clarity.

Author Andrew Grush reports growing frustration with Claude’s increasingly restrictive safety guardrails, noting that the AI frequently misinterprets creative writing prompts and benign inquiries as controversial or harmful. In a recent test of user sentiment, 35% of respondents agreed the chatbot has become overly preachy and difficult to work with, while 41% reported no issues. The author highlights a specific conflict regarding a fictional story concept involving religious themes; despite clarifying the intent was purely creative, Claude refused the request on the grounds that it did not want to present potentially sensitive historical interpretations as fact. The inconsistency of these refusals is a primary concern, as the author successfully executed the identical prompt in a separate conversation session, illustrating an unpredictable application of safety protocols.

The author observes that these issues persist across multiple models, including Sonnet 5, Opus 4.8, and Sonnet 4.6, with some variants showing higher sensitivity than others. He notes that Opus 4.6 and Opus 4.7 have exhibited fewer problems, allowing for more fluid interactions even on sensitive topics. This defensive behavior intensified following the restoration of Fable 5, which had been temporarily taken offline after government entities identified it as a security risk. The author posits that Anthropic likely tightened safety constraints across its model suite to mitigate further regulatory or security concerns.

Beyond creative writing, the author experienced pushback on various topics, including financial planning and hypothetical biology discussions related to a star system 10.5 light-years away. In these cases, models like Opus 4.8 often narrowed the scope of answers to ensure safety, treating hypothetical scenarios as if they were real-world plans. While the author continues to value Claude for its superior memory and long-form conversational capability compared to competitors like Gemini, the increasing need for highly specific prompt engineering has diminished the user experience. Despite these challenges, he notes that many users still interact with the platform daily without encountering these specific barriers, suggesting that the frequency of pushback is heavily dependent on individual usage patterns and prompt clarity.