Reporting Bias
Occurs when the data available for AI model training are not representative due to systematic omissions or non-random exclusion of certain data points or outcomes.
Reporting bias is a crucial issue in AI and statistics, where certain outcomes or data points are systematically under-represented or excluded, leading to biased model predictions and inferences. Such bias can emerge from the selective reporting of results, where only favorable outcomes are documented or from the lack of reporting certain information due to institutional or cultural pressures. This can significantly impact AI models, where training data must be comprehensive and unbiased for accurate performance. In practice, reporting bias can compromise the fairness and generalizability of AI systems, resulting in skewed outcomes that do not reflect real-world scenarios.
The term 'reporting bias' emerged around 1980, aligning with growing concerns about data integrity in research fields, and gained prominence with the rise of AI and ML technologies that rely heavily on vast datasets for training and accuracy.
Key contributors to the identification and exploration of reporting bias include researchers across various domains such as psychology, sociology, and computer science, with notable emphasis from statistical and AI communities that have worked to highlight and mitigate the impacts of biased data in model training and evaluation.