In standard supervised learning, we treat dataset labels as absolute truth. We train models under the assumption that for every image, text snippet, or audio clip, there exists a single, objectively correct "ground truth" label.
However, this assumption breaks down when we look closely at how datasets are actually built. Because machine learning relies heavily on human annotators, datasets inherently reflect the messy, subjective, and often biased nature of human judgment. Rather than capturing an objective reality, models frequently learn to replicate the specific cultural backgrounds, cognitive biases, and varying levels of expertise of the people who labeled the data.
To build more robust and equitable AI, a growing body of research focuses on moving away from the myth of the objective label, choosing instead to explicitly model and account for human subjectivity.
How Subjectivity Creeps into Ground Truth
Subjectivity enters the data pipeline long before a model begins training. It manifests primarily through three channels: varying levels of expertise, cognitive shortcuts, and the inherent ambiguity of human-centric tasks.
1. Variance in Expertise and Interpretation
When datasets are crowdsourced or curated at scale, annotators bring wildly different baselines of knowledge. In ecological monitoring, for instance, Connor M. Wood et al. point out in The machine learning–powered BirdNET App reduces barriers to global bird research by enabling citizen science participation that data collected by human observers contains inconsistent, unquantifiable biases rooted in individual skill levels.
Even highly trained experts are not immune. In genomics, Takeya Kasukawa et al. discovered that expert curators exhibit systematic naming biases, such as over-attributing biological functions based on weak similarities to well-known species rather than strictly following the underlying data (Development and Evaluation of an Automated Annotation Pipeline and cDNA Annotation System).
2. Cognitive and Social Biases
Annotators do not operate in a vacuum; their labels are shaped by social preconceptions. For example, when humans are asked to evaluate personality traits from video clips, they are highly susceptible to the "halo effect"—allowing superficial traits like attractiveness, age, or ethnicity to skew their judgments. As shown by Ricardo Darío Pérez Principi et al. in On the Effect of Observed Subject Biases in Apparent Personality Analysis from Audio-visual Signals, personality computing models trained on these labels end up predicting the biased perceptions of the annotators rather than any objective ground truth about the subjects.
3. Inherent Task Ambiguity
For tasks like sentiment analysis, content moderation, or fact-checking, "truth" is fundamentally up for debate. Preslav Nakov et al. note that automated fact-checking models often fail because their training data reflects specific, localized perceptions of truth rather than objective reality (Automated Fact-Checking for Assisting Human Fact-Checkers). When the core concept being measured—such as trustworthiness or toxicity—is socially constructed, forcing a single label erases the valid disagreement inherent to the problem.
The Downstream Consequences
When we ignore this subjectivity and force a consensus (such as via majority voting), we introduce systematic errors. Models trained on these flattened labels inherit and amplify human prejudices.
In visual quality assessment, Haoning Wu et al. demonstrate that models trained on subjective quality scores become bound to the specific aesthetic preferences of the original annotators, limiting their ability to generalize (Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels). Furthermore, legal and regulatory frameworks often struggle to govern these systems because, as Alejandra Bringas Colmenarejo et al. argue, high-level AI regulations lack effective mechanisms to monitor or audit the subtle, systemic biases introduced during human data annotation (Fairness in Agreement With European Values: An Interdisciplinary Perspective on AI Regulation).
Shifting the Paradigm: From Hard Labels to Soft Distributions
If human subjectivity is unavoidable, how do we build reliable machine learning models? The core insight of this research direction is that disagreement among annotators is not noise to be filtered out; it is a signal to be modeled.
Instead of forcing annotators to agree on a single "hard" label (e.g., deciding an emotional expression is strictly "Happy" or "Sad"), researchers are shifting toward "soft" labels that preserve the full distribution of human opinions.
Traditional Hard Labeling (Majority Vote):
[Annotator 1: Happy]
[Annotator 2: Happy] ---> Final Label: [1.0, 0.0] (100% Happy)
[Annotator 3: Sad] *Signal lost: The expression was ambiguous.*
Soft-Target Labeling (Distributional):
[Annotator 1: Happy]
[Annotator 2: Happy] ---> Final Label: [0.66, 0.33] (66% Happy, 33% Sad)
[Annotator 3: Sad] *Signal preserved: The model learns the ambiguity.*
An elegant demonstration of why this works can be found in the work of Peter Washington et al. (Training Affective Computer Vision Models by Crowdsourcing Soft-Target Labels). When training computer vision models to recognize emotions, the authors compared models trained on traditional hard labels against those trained on soft probability distributions derived from crowdsourced consensus.
They discovered that while hard-labeled models might score higher on rigid, top-1 accuracy metrics, soft-target models produce output distributions that closely mirror actual human perception. By training on the variance between annotators, the model learns to identify compound or ambiguous emotions, making it far more resilient when deployed in real-world, nuanced environments.
Systemic Solutions and Data Governance
Beyond algorithmic adjustments, addressing annotation subjectivity requires structural changes to how datasets are documented and governed.
- Human-Centered AI Frameworks: To prevent the embedding of human prejudices, You Chen et al. advocate for a multidisciplinary approach that actively audits for cognitive, inter-annotator, and confirmation biases throughout the labeling lifecycle (Human-Centered Design to Address Biases in Artificial Intelligence).
- Standardized Provenance Reporting: Currently, there is a severe lack of transparency regarding who labels our data. Andrew M. Demetriou et al. audited highly cited recommender systems research and found a widespread failure to report annotator demographics, training, or agreement levels (Annotation Practices in Societally Impactful Machine Learning Applications: What are Popular Recommender Systems Models Actually Trained On?). Standardizing this reporting is essential for understanding the boundaries and biases of modern AI models.
By treating human subjectivity as a core feature of data rather than an error to be erased, the machine learning community can move toward systems that do not just mimic human decisions blindly, but understand the nuance, uncertainty, and context behind them.
Key Foundational Work
Detection and removal of switching noise in push-to-talk and voice operated exchange communications systems (2007) — contributed 1 concept elements to this direction.
Development and Evaluation of an Automated Annotation Pipeline and cDNA Annotation System (2003) — contributed 1 concept elements to this direction.