Positionality-Aware Machine Learning


There is no "view from nowhere".

A person’s position in the world and set of experiences shape their view of the world. It defines the bounds of their perspective and what assumptions may appear to them as "universal truths".

Positionality is the social and political context that influences, and potentially biases, a person’s unique but partial understanding and outlook on the world.


Positionality is not bias.

Current efforts to de-bias data and models suggest AI can be objective implying a view of AI as potentially better than human, impartial in principle, an arbiter where humans fail to agree. We give too much power to a future AI that is impossible in principle.

Positionality is propagated by AI systems

The scale of ML and AI systems makes it easier than ever to embed positionality in society. If the embedded positionality is not aligned with the current context, it could impact different sectors of the society through the scale of AI applications.

How does positionality impact ai systems?

Measuring Race and Ethnicity Across the Decades: 1790–2010 in US Census (source)

Classification systems are designed to organize items and ideas into discrete categories. When these “boxes” are designed, certain characteristics are inevitably brought into focus while rendering others invisible. Classification systems are contextual. They are informed by the perspectives, experiences, and knowledge of their creators. As most AI systems depend on classification systems as the information infrastructure, its positionality is also inherited in models and applications.

Sentiment Analysis Annotations Are Often Complicated! (source)

Data annotation is the process where crowd workers or domain experts assign additional labels to data samples. For many annotation tasks, choosing the label is a matter of individual judgement, and therefore inevitably informed by perspectives, experiences, and knowledge of the labeler. Categorizing all annotation disagreements as human errors silences the positionality of annotators. Instead, we need to discern the difference, and preserve positionality in annotated datasets.

Positionality can also impact AI systems through:

Data Collection

Data Sampling

Model design

Metrics design


Usage Analysis


Is Tomato a Vegetable or a Fruit?

If you want to build a classifier for vegetable and fruit, where do you put tomato?

It turns out - to botanists, it’s a fruit; to nutritionists, it’s a vegetable; to lawyers and judges, it’s a vegetable (according to the supreme court); to Computer Vision researchers, it’s a Misc (according to ImageNet)!

The classification of tomato is dependent on the application scenarios.

Positionality in International Classification of Disease (ICD-10)

Our circumstances and habits determine the ICD classification system. Why are ostriches, pigeons, crows, and swans not listed under the Contact with Birds sub-category? Why did it take 8 years to formalize medical coding for HIV? Why is it still not possible to code a patient's lack of access to electricity or a refrigerator, which could prohibit their use of insulin? Context, habits, and human life, as well as the need to count for, e.g., bureaucratic reasons, shape classification systems such as the ICD10.

Have more examples on how positionality influences your work?

workING with positionality

Reaching one "unbiased" AI to fit all is impossible.

Positionality is inevitable in AI systems. We explore how to design AI systems that can best fit the context of application.

Step 1: Uncovering Positionality

  • Identify classification systems, data, and models.
  • How does positionality enter the classification systems?
  • How does positionality impact the data labeling process?
  • How does positionality get embedded in the models and applications?

Step 2: Working with Positionality

  • Does the application scenario context fit with the data and model creation context?
  • How can the existing multitude of data and models of a domain benefit new applications?
  • How can we implement transfer learning effectively to facilitate the process of building context-aligned applications?

Step 3: Embedding Positionality in Workflows

  • Classification systems, data, and models, and context change over time.
  • How can we enable researchers and engineers to continuously iterate on their classification systems, data, and models in a real application development flow?

Join the conversation

Positionality is embedded in all application domains in artificial intelligence. Our white paper Towards Better Classification offers an overview on positionality and its effects on ML and AI. We have also been organizing workshops on specific application domains, where we work with domain experts to uncover positionality in their work. Follow our work through our papers and workshops!