Fairness (in ML)
Concrete Problems in AI Safety — inherited
An adjacent research area concerned with ensuring machine learning systems do not discriminate.
Whether machine learning systems discriminate, and how to keep them from doing so, is the business of fairness research. It appears in this corpus for a modest reason: Concrete Problems in AI Safety (2016) lists it among the related safety problems that fall outside its own study of accidents.
What follows sets the field aside just as the paper does, then explains what the field actually studies and why the paper draws its boundary of accidents around it rather than through it.
A field the paper deliberately sets aside
Fairness research asks how to ensure machine learning systems do not discriminate, and it long predates Concrete Problems as an active area within and beyond machine learning; the paper cites it only to exclude it, listing it in one line among several "related problems in safety" that fall outside its own five-problem taxonomy of accidents (concrete-problems, §"Fairness:", p. 21). Its role in the corpus is entirely as a boundary marker rather than as an object of study.
What the field actually studies
Fairness in machine learning is generally concerned with disparate treatment or disparate impact across groups defined by protected attributes such as race, gender, or age, and the field has developed multiple, sometimes mutually incompatible, formal criteria for what "fair" means: demographic parity requires a model's positive-prediction rate to be equal across groups, $P(\hat{Y}=1 \mid A=a) = P(\hat{Y}=1 \mid A=b)$; equalized odds instead requires equal true-positive and false-positive rates across groups, conditioning the equality on the true label $Y$ rather than ignoring it. These criteria can conflict with each other and with predictive accuracy, which is part of why fairness is treated as its own research area rather than folded into any single fix.
Why the paper places it outside accidents
The distinction Concrete Problems draws is definitional, not a judgment of importance: Accidents in machine learning systems is defined as unintended harm from poor design, while a fairness failure is a specific kind of harm, discriminatory treatment, that can arise even from a system performing exactly as its designers intended if the designers' objective itself encoded or permitted disparate outcomes. This places fairness-research-area within the theme Drawing the boundary: what accident risk is not alongside Privacy (in ML), Security (attacks against ML systems), Abuse (of ML systems), Transparency (in ML), and Policy (economic/social impacts of ML), the theme through which the paper states that it believes work on these topics has "both urgency and great promise" (concrete-problems, §"Related Problems in Safety:", p. 21) even as it declines to develop them itself.