I would like to respond to the idea that the idea of human-in-the-loop design incorporate human judgment in effective ways.
First of all, the idea of humans in the loop design remind me very starkly of my time in New York over the summer at Spotify, where the emphasis of humans-in-the-loop music curation could not be more apparent; whereas Spotify has both Editorial ( human-curated) playlists and algorithmic playlists (ex. Discover Weekly), the company has been looking to merge the two to have humans-in-the-loop algorithmic playlists, essentially playlists where editors are curating a list of songs from an algorithmically curated pool of suggestions. I think this illustrates how human-in-the-loop can effectively merge human and technology driven sides of music taste to deliver more effective music recommendations. I do think that human skills including creative curation and understanding culture, equip us to be the project managers that AI systems can’t be in areas like music.
This also reminds me of the example of recruitment. While AI can quickly source a slew of relevant candidates via basic filters, many candidates will inevitably come through that aren’t the right fit for the position we’re trying to fill. A computer doesn’t know all the ins and outs of company culture.
However, I believe that there are many risks that can also be considered, whereas the reading paints a fairly positive image of humans-in-the-loop AI. There is opportunity both for human error and machine error which compounds the other. Air France 447 had a glitch which caused the plane to return to full manual control. The pilot made a mistake, which wouldn’t have been possible if normal safety overrides had stayed. Poor design choices can produce an AI that’s confusing to its human operators, so this is something we must also think about in building HITL systems.
