How would you measure harm? Any normal software system will occasionally hurt people who don’t fit the world model written into the system, so zero hurt isn’t r

How would you measure harm? Any normal software system will occasionally hurt people who don’t fit the world model written into the system, so zero hurt isn’t reasonable. How then? Probably by using some statistical measure, like “less than X% of human users are negatively affected by the system”. Ok, sounds good. But if X is 1 and your system hews that number by negatively affecting 20% of a specific group your system is ostensibly not targeting, that’s still a problem. That’s why specific protected groups are called out.