At a
SOMA Salon last month, I presented on the ethics of design AI-based systems. Traditional design thinking is clearly an important part of designing products that use AI. But there are some special considerations that apply to AI-based systems: learning from the world as is and over-optmisation.
Two recent posts demonstrate this.
This research group
attempted to predict runway model success by looking at wide slews of data about models. It’s a reasonable example of the kinds of new approaches internet-connected data can afford the prediction problem. The researchers joined together disparate data: physical attributes, contractual elements, Instagram behaviour and even sentiment on social. Then applied this to what has traditional been an ambiguous problem of cultural filtering.
The paper, which is limited in its scope and conclusions, highlights many of the uncomfortable facts of machine learning: (1) we can only represent and reinforce the world as it was (2) the algorithm & feature choice determines the nature of the story we can tell (3) AIs are systems optimised for ruthless optimisation, so any biases in (1) and (2) will get magnified.
Undoubtedly the demand for computer-mediated filtering of social or cultural processes will increase. A more comical example is how these “horny nerds” used
deep learning to identify attractive women on Tinder. Here machine learning or AI was used to reduce friction in a process, in this case mate or partner selection.
Providing both a challenge and an opportunity is that much of the data we’ll need to explore the world is no longer publicly accessible. Instead it is held in the
data centres of some of the world’s largest technology firms.
The good news: these firms are steeped in the language of experimentation and so do at least exploit these data
at an experimental scale that would make most in the academy jealous, as Michael Schrage argues.
The bad news (as I argue) is that these commercial experiments are optimisations around very particular local maxima: that of generating whatever outcome a particular product needs to achieve (e.g more clicks on adverts, longer user-linger time to create more opportunities to show adverts, higher referral loops to drive more users to ultimately increase advertising inventory, and so on.)
Perhaps one day we’ll see large firms offer access to these data and their experimental infrastructure to improve academia’s ability to explore and test wide-ranging rather than commercial hypotheses.