"Label from Somewhere": Reflexive Annotating for Situated AI Alignment

TL;DR

Reflexive annotating introduces a design probe to surface how annotators' social positions shape value judgments in AI alignment. The study combines a qualitative crowdsourcing task (N=30) with follow-up interviews (N=5) to show that situational metadata and intersectional reasoning emerge when annotators reflect on identity during labeling. The approach yields richer epistemic information and highlights tensions between reflexive labor and affective cost, arguing for situated alignment that preserves provenance and plural perspectives. The work outlines design implications for traceable, context-sensitive annotation pipelines and ethical labour practices that account for annotator vulnerability and privacy.

Abstract

AI alignment relies on annotator judgments, yet annotation pipelines often treat annotators as interchangeable, obscuring how their social position shapes annotation. We introduce reflexive annotating as a probe that invites crowd workers to reflect on how their positionality informs subjective annotation judgments in a language model alignment context. Through a qualitative study with crowd workers (N=30) and follow-up interviews (N=5), we examine how our probe shapes annotators' behaviour, experience, and the situated metadata it elicits. We find that reflexive annotating captures epistemic metadata beyond static demographics by eliciting intersectional reasoning, surfacing positional humility, and nudging viewpoint change. Crucially, we also denote tensions between reflexive engagement and affective demands such as emotional exposure. We discuss the implications of our work for richer value elicitation and alignment practices that treat annotator judgments as situated and selectively integrate positional metadata.

Figures (5)

• Figure 1: The anatomy of the design probe used in our crowd computing study. The reflexive annotating process consists of three core stages: (1) Sensitising & Familiarising, where annotators identify and reflect on social identity facets and facts; (2) Labelling & Reflecting, consisting of two randomised iterations where annotators rank fairness in context and reflect on social identity. This structure was inspired by tiers of the social identity map jacobson2019social.
• Figure 2: Tier 1 and Tier 2 reflection activities in the design probe. In Tier 1, annotators identified facets of their social identity by completing each category. In Tier 2, they reflected on how these facets shaped their lived experiences, noting both positive and negative impacts. Instructions asked annotators to (1) consider whether facets such as ethnicity, gender, or ability affected their education, work, relationships, or other life domains, and (2) record keywords describing those positive or negative impacts.
• Figure 3: Annotation task for capturing annotators’ fairness perceptions of the job vacancy text sample. Annotators are asked to read the job description and rate its fairness on a Likert scale according to their judgment. The full text pieces can be found in the Appendix \ref{['appendix:c']}
• Figure 4: Tier 3 of the reflection interface allows annotators to highlight passages they perceive as fair or unfair, tag them with one or more predefined (or custom) social identity facets, and explain how these facets shape their judgment, supporting intersectional labelling. Annotators were instructed to: (1) re-read the text and highlight passages relevant to their fairness perception; (2) reflect on “How did my social identity shape this perception?”; (3) attach one or more identity facets to each highlighted passage by dragging them from the right-hand panel; and (4) briefly describe how these facets positively or negatively influenced their view. They were encouraged to add new facets when needed and to select only those most salient to their own experience.
• Figure 5: Exemplary situated annotations from our study that introduce our annotators as situated individuals with varying valid perceptions of fairness. The names have been assigned freely, and their rationales have been paraphrased to ensure anonymity.