Abstract
Pseudo Artificial Intelligence bias (PAIB) is broadly disseminated in the literature, which can result in unnecessary AI fear in society, exacerbate the enduring inequities and disparities in access to and sharing the benefits of AI applications, and waste social capital invested in AI research. This study systematically reviews publications in the literature to present three types of PAIBs identified due to: a) misunderstandings, b) pseudo mechanical bias, and c) over-expectations. We discussed the consequences of and solutions to PAIBs, including certifying users for AI applications to mitigate AI fears, providing customized user guidance for AI applications, and developing systematic approaches to monitor bias. We concluded that PAIB due to misunderstandings, pseudo mechanical bias, and over-expectations of algorithmic predictions is socially harmful.
Introduction
The term "AI Bias" has become a ubiquitous label for any undesirable or controversial output from machine learning models. While genuine algorithmic bias—stemming from skewed training data or flawed model design—is a critical challenge, the current discourse often fails to distinguish between inherent technical flaws and human-centric errors. This paper introduces the concept of Pseudo AI Bias (PAIB): instances where "bias" is attributed to an AI system but actually originates from human misunderstanding, operational misuse, or unrealistic expectations.
Labeling every error as "bias" risks exacerbating technophobia, misallocating research resources, and obscuring genuine systemic inequities. This study provides a conceptual framework to help researchers and the public differentiate between machine flaws and human-centric "pseudo" biases.
Methodology and Definitions
To establish a rigorous foundation, the authors analyzed authoritative literature across social and natural sciences. They identified three core properties of bias: deviation from ground truth, systematicity (it is not a random error), and tendency (favoring one group over another).
Using these criteria, the study performed a meta-analysis of high-profile "AI bias" cases, including the COMPAS recidivism scores, Amazon’s recruiting tool, and medical imaging datasets. This analysis led to the development of a new taxonomy of Pseudo AI Bias.
A Taxonomy of Pseudo AI Bias
The authors categorize PAIB into three distinct types, illustrated by common scenarios where training data and testing environments diverge.
1. Misunderstanding of Bias
- Error vs. Bias: A single mistake is a bug or error. It only constitutes bias if it occurs systematically across a population.
- Fairness vs. Bias: In cases like COMPAS, achieving mathematical accuracy and social "fairness" simultaneously is often impossible. Choosing one metric over another is a social policy decision, not an algorithmic failure.
2. Pseudo-mechanical Bias
This occurs when AI is operated incorrectly or applied to contexts it wasn't designed for:
- Distributional Shifts: Using a model on data it never saw during training (e.g., applying an English-trained model to Chinese cultural contexts).
- Proxy Variables: In healthcare, using "spending" as a proxy for "health" led to biased outcomes for Black patients who historically had less access to care. The bias was in the human choice of the proxy variable, not the machine's calculation.
3. Over-expectation
Humans often expect AI to solve subjective dilemmas that lack a ground truth. If an AI judges a beauty contest and the results lack diversity, it often reflects the subjective labels provided by humans or the inherent impossibility of quantifying a subjective concept like "beauty" without bias.
Consequences and Solutions
The proliferation of PAIB is socially harmful as it shifts accountability from human decision-makers to the "black box" of the machine. To mitigate these effects, the authors propose:
- User Certification: Professional AI users (e.g., judges, doctors) should be certified to understand the mechanics and limitations of their tools.
- Customized Guidance: Developers must provide clear "operating envelopes" to prevent out-of-sample application.
- Systematic Monitoring: Utilizing open-source audit toolkits like Aequitas to distinguish between biased actions and biased outcomes.
Conclusion
While addressing real algorithmic bias remains paramount, distinguishing it from Pseudo AI Bias is essential for the responsible advancement of technology. Future work should focus on empirical validation of this taxonomy and the development of "AI literacy" programs to help society navigate the nuances of machine-assisted decision-making.
Summary updated