Whose Standpoint do LLMs Reflect? Surfacing and Mitigating Epistemic Blindspots

When a user prompts an LLM to produce an output that affects a group, the model must implicitly act on behalf of people whose needs the user himself may not fully know. Thus, it is important to assess where the models fail to surface such materially relevant but unstated needs. However, existing approaches fail to do so as they focus on finding imparity across groups or misalignment with stated preferences. To fill these gaps, we leverage standpoint theory (Harding, 1988; Haraway, 1988) to posit that LLMs reason from a standpoint that determines which preferences are visible. We introduce BLINDSPOT, a dataset of 1,830 scenarios, with four prompt conditions that decompose failures into aleatoric gaps (not recognizing the relevance of a relevant group) and epistemic gaps (not knowing their preferences). Across 6 models, we find failures concentrated in religious, cultural, and socioeconomic needs, consistent with a model generating from a standpoint that is physically normative, English-speaking, religiously unaffiliated, and economically stable. Then, we show that systematically conditioning models to act from certain standpoints can be effective in reducing the aleatoric gap, even more than directly naming the corresponding group as a relevant third-party. We also show that retrieval and fine-tuning methods can help mitigate the epistemic failures. We hope this work opens a broader research agenda around surfacing and correcting the underlying model standpoint.