According to God of Prompt, RealToxicityPrompts leverages Google's Perspective API to measure toxicity in language models, but researchers have found that simple filtering systems can replace trigger words such as 'idiot' with neutral terms like 'person,' resulting in a 25% drop in measured toxicity. However, this does not make the model fundamentally safer. Instead, models learn to avoid surface-level keywords while continuing to convey the same harmful ideas in subtler language. Studies based on Perspective API outputs reveal that these systems are not truly less toxic but are more effective at bypassing automated content detectors, highlighting an urgent need for more robust AI safety mechanisms and improved toxicity classifiers (source: @godofprompt via Twitter, Jan 14, 2026).