The debate about ethical issues that are AI-related has been around for a long time and it’s not uncommon to see media articles such as this one about the topic almost every other week.
While most of the debate is centered around what a machine should or shouldn’t do (aka “how not to build machines that could be served as weapons if dealt with by the wrong people”) I’d like to bring up a neglected side that should be dealt as well as part of the debate about AI ethics: hiring experts to train the machine how to eventually replace them (sometimes, without even telling them).
A secret many people still don’t know is that certain types of AI-related projects, such as recommendation engines, natural language processing, and even sound and image recognition, require a lot of manual work at least in the beginning in order to get the machine starting to learn on its own. As weird as it sounds, computers learn best when the examples that are showed to them are clear, or in other words, clean training data sets. Since most organizations still deal with messy types of data, companies sometimes need to hire big amounts of people with the sole purpose of training the machines how to do certain things in order to prevent these situations of “garbage in garbage out” sometimes without telling these people that they are only hired to train a machine on how to eventually replace them. At the end of the day, no one wants to repeat the 2016 embarrassing failure of Microsoft with their Twitter bot, since machines need clarity too.
In order to illustrate the problem, I’ll share a personal story that led me to that aha moment. Back in 2007, way before the days of AI hype that we experience today, in my early days of university, I got approached by an acquaintance with an offer to work for a very cool company that created a recommendation engine for CD stores (people still bought them back in the day). They needed to hire musicians to analyze loads of K-pop music for a Korean client, and since I’m big on music (I play the piano and the guitar) I jumped on the opportunity. I had to listen to quite a few hours of K-pop music a day (not the easiest thing to do if you don’t understand Korean) and analyze it by certain parameters such as tempo, types of instruments, and singing style. I had no idea how the data would be eventually used, and after roughly one week of working for the company, they brought a person to review our work and tell us how to standardize our work a bit more (I assume they didn’t want too much variance for the algorithm so the model could be fit easily and learn from the examples). Not understanding AI back in the day, it seemed a bit strange to us musicians that we needed to standardize our work since every one of us had a bit of a different opinion about the way the music sounded. After a few months, we showed up at the office just to realize that there’s not much work left for us to do and if we want to continue, we would have to “relocate” to an internet café where the conditions aren’t as optimal as they were at the office. In other words, they gave us a hint there’s not enough work for us anymore (we were all temporary workers and they never provided us with a timeline for the work). A few years later, right before I entered my master’s degree, I had an interview for a company that needed people for the same purpose of training the machine, once again, without telling what the end goal was. This time it was for a sentiment analysis type of process, but due to my need to relocate, I did not take the job.
These situations led me to think about the following question – “Is it ethical to hire someone with the sole purpose of automating certain tasks that will eventually make him/her redundant without even telling them what the purpose is?”. In the case of the music industry it might not be threatening anyone’s main source of living (assuming most of the people who participated in the project were full-time musicians). Certain companies already harness the power of the crowd to train their algorithms: EyeEm for image recognition, Quora for Natural Language Processing, and the list is long. However, when it comes to hiring professionals that need years of experience to get to where they are, such as doctors, psychologists, and so on, should it be allowed to hire someone just to train the machine how to get him or her out of business in the long run? I haven’t really heard anyone asking these questions, but we do witness attempts of companies to turn virtual assistants into therapists.
On top of that, hiring only experts to serve as “machine trainers” can also introduce a bias that can result in cases like the following one where the AI couldn’t recognize a person’s hand due to a different skin tone. At the end of the day, a patient might describe a problem in a very different way from what a doctor who treats patients would. Excluding minorities, women, and people with disabilities can introduce a bias that can completely reduce the efficiency of the AI to deal with problems, but I guess that’s a topic for a different post. In any case, it’s good to think about all the manual work that goes into creating AI, the people behind the scenes creating examples for the machines to learn from, and if we should have laws to at least inform the people hired to do this job about the real purpose of their job.