Many of us frequently encounter a service called Recaptcha when filling in a web form. The service is promoted as being a security measure, to avoid bots automatically spamming forms – but it's much more than that.
Consider the situation when Recaptcha offers up a 3x3 matrix of blurry images, some of which feature cars. To get through Recaptcha's challenge, you are required to confirm which of the 9 images contain cars. But, it's more than that. You are also confirming to Google, Recaptcha's owner, which images feature cars and which images it thinks might feature carsbut don't. Ironically, you have just become the service.
For AI to work, it needs to ingest and categorise massive amounts of data. Image processors like Recaptcha and similar services like fuzzy image search rely on such "understanding". No AI system has eyes and a brain; it just "sees" binary data. So, for it to figure out what an espresso machine is and what it is not, the system needs help. That help comes from training libraries. (NB Where to buy espresso machines)
In the case of image services, the software ingests millions of images which are used in its "training". However, it doesn't just end with a massive upload. Issues around bias have often surfaced over the years, as humans have chosen particular images, or subjects within images, and transferred their own biasses (not necessarily consciously) onto how the images are described. One might argue that it is impossible to have a 100% neutral AI system, because the ingested data is chosen by humans with some degree of bias.
Because of this, it's always more than just the cosmetic - although the time is coming where AI-classified images evolve into AI-produced images. Image-heavy e-commerce sites such as bovemlife.com will no longer require expensive staging of photos, or even the purchase of visually-sophisticated imagery from stock libraries. A natural language request will generate a perfect, new image every time. At that point, what AI understands has transcended visual data and has evolved into what AI understands about us.
Training Humans,an exhibition at Milan's Fondazione Prada, explores these themes. AI researcher Kate Crawford and artist Trevor Paglen have explored the relationship between the human directly training the machine, and the machine indirectly training the human. Within these relationships come issues of power, history, and the moral impacts of science. As an AI image classifier will tell you, nothing is at it seems.
Training Humans is at Fondazione Prada until 24/02/2020.
(This article contains paid placement links.)