| Holy Christ, Are you seriously asking why ethics and concern for how the systems you design interact with end users and targets of those systems might be a worthy consideration? Let me give you a concrete example: Imagine you are an software engineer tasked with working on a facial recognition system to help police identify known criminals to help find suspects near the time and location of a crime. It observes nearby people and assigns a probability to them being a known criminal. Police department demands 80% accuracy for the product. You design such a system using some blackbox facial recognition AI, and you get the following results: Overall 78% accuracy with: 6.5% False Positive rate
31% False negative rate Not too bad, you tweak some things, hit your 80% accuracy without messing with the false positives too badly, and you meet the specification provided by the client. Mission accomplished and you're ready to ship right? Makes the company money? No problems? Cool. Except, because you didn't really care that much about how the technology you deployed would be used or the ethics surrounding its use, you failed to consider the right performance targets despite what your client asked for and your system is nearly 100% racist. What happened? You trained on equal numbers of prison mugshots, and mugshot like photos of people with no criminal records. You failed to consider that black people are over represented in the US prison system. (38% of prisoners but 13% of US population) Your classifier just learned to label someone a likely criminal if they were black and essentially no other criteria. Yet, the actual likelihood the people identified by the system as "criminals" in fact have a criminal history is at most somewhere ~33% despite the fact your system labels it as 80% likely. Worse, even if we have a hypothetical situation where blacks and non-blacks are represented in their average proportions, there's a near equal number of black and non-black people with criminal histories in the vicinity of the crime! Worse still, since people tend to be more segregated than that, when blacks are in even more of a minority there will be more non-blacks with criminal histories around. When blacks make up a greater proportion, the likelihood of being falsely accused goes up even more. And FYI... such systems with similar flaws have actually been built and deployed in the past. How do you think that plays on trust in the company and the technology in general in the long run? Considering end-use ethics brings value. |