In traditional machine learning, we treat data labels as "ground truth"—an objective reflection of reality. However, when tasks involve identifying media bias, social harm, or political framing, this assumption collapses. Research across linguistics, social science, and AI suggests that for these subjective tasks, the label is not just a property of the data, but a product of the annotator’s own worldview.
The Myth of the Objective Label
The fundamental challenge in bias detection is that neutrality is not a fixed point. As Vlad Niculae et al. argue in their study of media quoting patterns, perceptions of neutrality vary significantly among different observers. This makes it difficult to establish objective benchmarks for media analysis without relying on the personal perspectives of the researchers themselves (QUOTUS: The Structure of Political Media Coverage as Revealed by Quoting Patterns).
When annotators are asked to label bias, they often exhibit low inter-rater agreement. For example, Gillian Dobbie et al. found that even expert annotators agreed on bias labels only 35% of the time when evaluating social harm in local contexts (Challenges in Annotating Datasets to Quantify Bias in Under-represented Society). This lack of consensus suggests that bias is not a binary feature that can be easily quantified, but a deeply subjective interpretation rooted in individual philosophical perspectives.
The Internal Filter: Worldviews and Ideology
Why do humans disagree so consistently? The evidence points to an "internal filter" composed of personal beliefs and prior knowledge. Christoph Hube et al. demonstrated that individuals tend to perceive statements they personally agree with as neutral rather than biased (LimitBias! Measuring Worker Biases in the Crowdsourced Collection of Subjective Judgments (short paper)). This psychological tendency causes workers to mistake their own viewpoints for objective common sense.
This effect extends to how we process news. Sora Lim shows that a reader's prior knowledge of an event and the political leaning of the news source significantly influence how bias is perceived (Annotating and Analyzing Biased Sentences in News Articles using Crowdsourcing). Similarly, Seunghyun Yoon et al. found that the perception of whether a headline matches its body text is heavily influenced by the reader's existing political and social biases (Detecting Incongruity between News Headline and Body Text via a Deep Hierarchical Encoder).
Systematic Bias vs. Category Indistinguishability
To improve data quality, we must distinguish between different types of disagreement. Rebecca F. Bruce et al. highlight the importance of separating systematic observer bias (where an annotator has a consistent personal preference) from category indistinguishability (where the definitions themselves are flawed) (Word-Sense Distinguishability and Inter-Coder Agreement).
In many cases, the categories we ask humans to label are simply too subtle for consistent detection. Marta Recasens et al. found that identifying specific words that induce bias—such as framing or epistemological bias—is a highly subjective task with low human agreement, suggesting these violations are often too nuanced for consistent manual labeling (Linguistic Models for Analyzing and Detecting Biased Language).
The Impact of Lived Experience and Context
Subjectivity is also driven by the annotator's physical and social context. In the realm of computer vision, Candice Schumann et al. showed that an annotator's geographic location significantly influences their "mental model" of skin tone, introducing subjectivity into labels used for auditing algorithmic fairness (Consensus and Subjectivity of Skin Tone Annotation for ML Fairness).
This geographic and cultural skew also affects how we interpret media framing. Gaku Morio et al. note that the interpretation of environmental "greenwashing" is inherently subjective and often exacerbated by the cultural biases present in training data (A Multimodal Benchmark for Framing of Oil & Gas Advertising and Potential Greenwashing Detection).
Implications for Machine Learning
The recurring evidence of low inter-annotator agreement—ranging from clinical case workups (Zwaan et al.) to conflict narratives (Al Khatib et al.)—suggests that the goal of a single "gold standard" may be misplaced for subjective domains.
Instead of forcing a consensus that doesn't exist in the real world, the authors of these papers suggest we must:
- Quantify worker bias by pairing tasks with opinion surveys (Hube et al.).
- Use prescriptive guidelines while acknowledging that interpretation remains a hurdle (AlEmadi et al.).
- Move beyond lexical analysis toward context-aware discourse modeling that accounts for informational bias and fact omission (Fan et al.).
By treating subjectivity as a feature rather than noise, researchers can build models that better reflect the complex, multi-perspective nature of human judgment.
Key Foundational Work
Kodak's consumer video benchmark data set: concept definition and annotation (2007) — contributed 1 concept elements to this direction.
Word-Sense Distinguishability and Inter-Coder Agreement (1998) — contributed 1 concept elements to this direction.