This paper investigates the relationship between decontextualized "trick tests" and Realistic Use and Tangible Effects (RUTEd) evaluations for bias in large language models (LLMs). The authors argue that current bias benchmarks, often based on contrived scenarios, are not effective at capturing real-world bias and harm. They compare three decontextualized evaluations with three RUTEd evaluations focused on long-form content generation, specifically gender-occupation bias. The results show no correlation between the two types of evaluations. Selecting the least biased model based on decontextualized results coincides with selecting the best-performing RUTEd model only by chance. This suggests that evaluations not grounded in realistic use are insufficient to assess or mitigate bias and real-world harm. The paper discusses the evolution of bias evaluations, distinguishing between intrinsic and extrinsic metrics. Intrinsic evaluations measure biases inherent to the model, while extrinsic evaluations assess biases in specific tasks. However, the distinction is becoming less clear as models become more complex. The authors also explore various bias metrics, including stereotype, neutrality, and skew, and find that they do not consistently correlate across different evaluation contexts. The study evaluates three RUTEd tasks: bedtime stories, user personas, and ESL learning exercises. For each task, the authors measure bias by analyzing the gendered pronouns and occupational biases in generated content. The results show that bias metrics vary significantly across different tasks and that no single model consistently performs best in all contexts. The authors conclude that current bias benchmarks are not sufficient to assess real-world bias and harm. They advocate for RUTEd evaluations that are grounded in realistic use cases and have clear connections to real-world impacts. The paper highlights the need for more nuanced and context-specific evaluations of LLM bias to better understand and mitigate its effects.This paper investigates the relationship between decontextualized "trick tests" and Realistic Use and Tangible Effects (RUTEd) evaluations for bias in large language models (LLMs). The authors argue that current bias benchmarks, often based on contrived scenarios, are not effective at capturing real-world bias and harm. They compare three decontextualized evaluations with three RUTEd evaluations focused on long-form content generation, specifically gender-occupation bias. The results show no correlation between the two types of evaluations. Selecting the least biased model based on decontextualized results coincides with selecting the best-performing RUTEd model only by chance. This suggests that evaluations not grounded in realistic use are insufficient to assess or mitigate bias and real-world harm. The paper discusses the evolution of bias evaluations, distinguishing between intrinsic and extrinsic metrics. Intrinsic evaluations measure biases inherent to the model, while extrinsic evaluations assess biases in specific tasks. However, the distinction is becoming less clear as models become more complex. The authors also explore various bias metrics, including stereotype, neutrality, and skew, and find that they do not consistently correlate across different evaluation contexts. The study evaluates three RUTEd tasks: bedtime stories, user personas, and ESL learning exercises. For each task, the authors measure bias by analyzing the gendered pronouns and occupational biases in generated content. The results show that bias metrics vary significantly across different tasks and that no single model consistently performs best in all contexts. The authors conclude that current bias benchmarks are not sufficient to assess real-world bias and harm. They advocate for RUTEd evaluations that are grounded in realistic use cases and have clear connections to real-world impacts. The paper highlights the need for more nuanced and context-specific evaluations of LLM bias to better understand and mitigate its effects.