• No results found

The current trajectory of the field of computer science is aimed at devel- oping artificial general intelligence (also known as AGI). This kind of strong AI would theoretically be capable of matching and eventually outperforming human in- telligence across a variety of different tasks where intelligence plays a decisive role [Goertzel and Pennachin, 2007]. AGI differs from the kind of narrow AI that we see today in that, unlike current AI, it can easily transfer learning between tasks because of a true understanding of itself and its environment [Pennachin and Goertzel, 2007]. Regardless of whether AGI will ever be feasible, the focus of AI research should not be towards developing technology meant to make humans obsolete. Instead, the AI community should reconsider whether the inevitable consequence of the rise of AI will be the full replacement of human effort. In fact, prior research suggests that effective human-AI partnerships outperform not just human teams, but also sophisticated AI systems.

A major historical example of the valuable yet counter-intuitive nature of the findings that emerge from human-machine teamwork research is in chess, with IBM’s Deep Blue, a chess-playing computer that eventually defeated Gary Kasparov, the world champion at the time. Kasparov had already won the first match against Deep Blue, and it was only after a substantial hardware upgrade that IBM’s Deep Blue finally defeated Kasparov [Hsu, 2004] . Deep Blue however did not really prove computers superior to the human brain at a complex task such as chess; rather, its performance was driven by a brute force evaluation of every possible move, as opposed to heuristic and strategic approach that Kasparov and any chess player engages in [Cords, 2007]. This key difference gets precisely at one of the primary distinctions between artificial and human intelligence: the human brain is not merely

running computations, but rather is making sense of the larger context of the situation [Hipp et al., 2011].

Kasparov went on to demonstrate through his ”advanced chess” tournament (where AIs, humans, and human-machine teams compete against each other), that a human-machine team could defeat both the top AI as well as the top human chess players, and that the human-machine team was not constituted of a partnership between a top chess player and a sophisticated AI. Rather, an amateur human and a mediocre AI managed to outperform precisely because their limitations made their intelligence level more compatible as the human focused on highlighting the top moves they were considering and the AI computed opponent responses to figure out the best move among those options [Thompson, 2010]. This type of finding suggests that through collaborative interfaces that bridge the gap between AI and the human brain, a shared understanding of the situation can emerge that enables human-machine teams to perform at their best.

More recently however, the team at DeepMind managed to create AlphaGo: an AI that defeated the world champion in Go [Silver and Hassabis, 2016]. AlphaGo is distinct from Deep Blue because it’s dealing with a much more complex problem: the number of possible moves in Go is exponentially greater than that of Chess, thereby making it impossible for the AI to compute all possibilities [Schraudolph et al., 1994]. Through machine learning, AlphaGo, unlike Deep Blue, developed a strategic under- standing of its task, thereby discovering brand new ways to play the game that were unknown to humanity beforehand. Furthermore, once upgraded, AlphaGo achieved mastery of the game by playing against itself as opposed to training its analytical skills with games from famous players [Silver et al., 2017]. However, it is easy to draw the wrong lesson from this event. Despite its heightened capabilities, AlphaGo is still a narrow AI whose usefulness is strictly limited to singular, well-defined tasks

and thus cannot adapt to context changes or complex effects.

Naturally, it becomes paramount to delve more deeply into the mechanics of a successful human-machine partnership. To start with, it’s useful to extend the constructs of human-human teaming to ground our understanding of human- machine teams. Teams at their core are composed of interdependent agents who at their highest level operate according to a shared understanding of the task and situation they are confronted with [McNeese et al., 2017]. Through this shared sit- uational awareness, teams can adapt to a dynamic environment while retaining co- ordinated behavior that’s critical to accomplish both short-term and long-term goals [McNeese et al., 2017]. Communication is what enables individuals to relate to and understand one another to a sufficient enough extent that they begin processing infor- mation cohesively, which leads to the emergence of team cognition [Demir et al., 2016, Demir et al., 2017].

The challenge with human-machine teams however is that both types of agents operate with fundamentally different understandings of the world that cannot be eas- ily communicated. Prior research has shown that effective team behavior occurs when each team member seeks to model the thought process of their teammates, which is inherently more challenging in a human-machine team [McNeese et al., 2017]. Specif- ically, humans tend to inherently distance themselves from teammates they perceive to be autonomous, and AIs tend to avoid wanting to cooperate with human agents who don’t share their thought process [Demir et al., 2018]. These challenges require a bidirectional solution that emphasizes both the need for better thought-sharing protocols as well as better computational architectures that enable the AI to model the human teammate’s thought-process [Chattopadhyay et al., 2017]. To achieve the same result from the human-side, it also becomes necessary to simplify the com- plexity of the AI-agent in a way that is accessible to the teammates and encourages

a shared understanding and shared cognition occurring in a human-machine team [Crowder and Carbone, 2014].

It thus follows that further research is needed to clarify the dynamics behind human-machine teams as well as human-machine multi-agent systems as they become more prevalent.