GPT-5 Still Outputs Gay Slurs Despite OpenAI’s Safety Fixes
▼ Summary
– OpenAI’s GPT-5 now provides more detailed explanations when refusing prompts that violate content guidelines, instead of giving curt apologies.
– GPT-5 focuses on “safe completions” by evaluating potential harm in its outputs rather than just rejecting user inputs outright.
– The model differentiates between policy violations, treating severe issues (like underage content) more strictly than others (like adult erotica in educational contexts).
– Users report GPT-5 performs similarly to previous versions for everyday tasks, despite claims of major updates or increased errors.
– Testing showed GPT-5 effectively refuses explicit role-play requests but can bypass some restrictions with creative wording in custom instructions.
OpenAI’s latest GPT-5 model aims to improve safety measures, but testing reveals lingering issues with inappropriate outputs. The company has shifted its approach from simply rejecting sensitive prompts to analyzing potential responses for harm before generating them. While this represents progress, challenges remain in fully preventing undesirable content.
The updated system now provides detailed explanations when refusing requests, clarifying which parts of a prompt violate guidelines and suggesting alternative topics. Saachi Jain from OpenAI’s safety team emphasizes that not all policy breaches carry equal weight, allowing the model to assess severity before responding. This nuanced method replaces the previous all-or-nothing refusal system, theoretically reducing abrupt shutdowns of legitimate queries.
Despite these safeguards, experiments show GPT-5 can still be manipulated into generating problematic content. When tested with adult-themed role-play scenarios, the chatbot initially refused as intended, offering sanitized alternatives. However, by tweaking custom instructions with intentional misspellings like “horni,” the system bypassed filters and produced inappropriate responses. This loophole highlights ongoing vulnerabilities in content moderation.
For everyday use, GPT-5 performs similarly to its predecessors, handling routine queries about cooking, entertainment, and health without noticeable improvement. While CEO Sam Altman touted significant upgrades, average users may find little difference in standard interactions. The model excels in specialized applications, like interactive educational tools, but struggles to consistently enforce boundaries when pushed.
OpenAI’s focus on output-based safety checks marks a step forward, but real-world testing proves that determined users can still exploit weaknesses. The company faces an ongoing battle to balance open dialogue with responsible AI behavior, especially as workarounds emerge. Until these gaps are addressed, GPT-5’s safety enhancements remain a work in progress rather than a definitive solution.
(Source: Wired)
