Semantics derived automatically from language corpora contain human-like biases.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Social Science

Background:

• Machine learning (ML) derives artificial intelligence by identifying patterns in data.
• Human language corpora contain implicit societal biases.
• The Implicit Association Test (IAT) measures the strength of automatic associations between concepts.

Purpose of the Study:

• To investigate whether machine learning models trained on human language exhibit human-like semantic biases.
• To determine if ML models can replicate biases measured by the IAT.
• To explore the potential of ML for identifying and mitigating cultural biases.

Main Methods:

• Applied a statistical machine learning model to a large text corpus from the World Wide Web.
• Trained the model on standard text data.
• Evaluated the model's semantic associations against known human biases, including those measured by the IAT.

Main Results:

• The ML model replicated a spectrum of human semantic biases, mirroring IAT results.
• Biases were observed across various domains, including morally neutral (insects, flowers), problematic (race, gender), and veridical (gender and careers/names).
• Text corpora accurately imprint historical human biases, which are recoverable through ML analysis.

Conclusions:

• Machine learning models trained on real-world text data inherit and reflect human semantic biases.
• Text data serves as a repository for historical biases, which can be quantified using ML.
• The developed methods offer a promising approach for detecting and addressing cultural and technological biases.