In the standard supervised learning narrative, "ground truth" is treated as an objective oracle—a set of labels that represent an absolute reality. However, as machine learning moves from objective tasks like identifying stop signs to subjective tasks like detecting hate speech, humor, or political bias, this foundation begins to crumble. Research across natural language processing and computer vision reveals that what we call ground truth is often just a collection of human opinions, shaped by the cultural backgrounds, personal biases, and cognitive limitations of the annotators.
The Myth of Objective Labels
The core problem is the "Semantic Gap": the distance between raw data and the linguistic labels we use to describe it. Fausto Giunchiglia, Mayukh Bagchi, and Xiaolei Diao argue in Visual Ground Truth Construction as Faceted Classification that even structured classification frameworks cannot eliminate human interpretation. When an annotator maps a visual feature to a category, they are not just recording a fact; they are performing a subjective mapping influenced by their own mental model.
This subjectivity creates "invisible uncertainties" throughout the analysis pipeline. Pantea Haghighatkhah and colleagues identify that these uncertainties stem from both the original content creators and the analysts themselves, leading to distortions that automated systems struggle to quantify Characterizing Uncertainty in the Visual Text Analysis Pipeline.
Cultural Context and Systematic Bias
When annotators lack the cultural context of the data they are labeling, the resulting "ground truth" becomes a vehicle for systemic bias. A landmark study by Maarten Sap and Noah A. Smith demonstrated that hate speech detectors are significantly biased against African American English (AAE), often mislabeling dialect markers as toxic because annotators do not understand the speaker's background The Risk of Racial Bias in Hate Speech Detection.
This issue is compounded by the "Garbage In, Garbage Out" principle. Bertie Vidgen points out that most datasets fail to document the social identity of annotators, obscuring the specific cultural lenses through which the data was filtered Directions in abusive language training data, a systematic review: Garbage in, garbage out. Whether it is distinguishing sex education from suggestive content on TikTok Enfa George or rating the "funniness" of standup comedy Anirudh Mittal, the label depends more on who is watching than what is being shown.
From Consensus to Pluralism
If ground truth is subjective, how should we build datasets? Researchers are moving away from the idea of a single "correct" label toward more nuanced paradigms:
- Descriptive vs. Prescriptive Annotation: Paul Röttger, Bertie Vidgen, and Dirk Hovy propose that researchers must choose between descriptive annotation (capturing the diversity of human beliefs) and prescriptive annotation (enforcing a specific policy) Two Contrasting Data Annotation Paradigms for Subjective NLP Tasks.
- Modular Definitions: To reduce noise, researchers like Urja Khurana and Eric Nalisnick suggest breaking down vague concepts like "hate speech" into specific, modular criteria Hate Speech Criteria: A Modular Approach to Task-Specific Hate Speech Definitions.
- Automated Proxies: In some cases, researchers bypass human labels entirely. For example, K. S. Krishnapriya and Kevin Bowyer found that automated skin tone metrics can be more consistent than human raters, who often exhibit internal bias Analysis of Manual and Automated Skin Tone Assignments for Face Recognition Applications.
The LLM Judge Dilemma
As we increasingly use Large Language Models (LLMs) to evaluate other models, the problem of subjectivity scales up. Lingzhi Wang and colleagues have found that LLMs like GPT-3.5 inherit the same ideological biases as human annotators, often misclassifying partisan content as neutral Investigating Bias in LLM-Based Bias Detection: Disparities between LLMs and Human Perception.
Ultimately, the research suggests that bias is not a "bug" to be mathematically solved, but a socio-cultural judgment. As Heather Cole-Lewis and Stephen Pfohl argue, evaluation frameworks must treat bias as a multifaceted human assessment rather than a binary classification task A toolbox for surfacing health equity harms and biases in large language models. Understanding why a model disagrees with a human—or why two humans disagree with each other—is becoming as important as the prediction itself.