You walk into a room and your cat springs to life. The situation resembles a video game in which a non-player character – any character not controlled by a player – animates itself as soon as your avatar enters its vicinity. But there are subtle behavioural differences between the two scenarios. In the case of your pet, it had a life before you entered the room; perhaps it was snoozing, sniffing the furniture, or looking for food. And the precise way in which it responds upon coming to life depends on exactly who is entering the room and what they are doing. By contrast, the behaviour of a non-player character (NPC) is repetitive and, quite literally, robotic. It will typically snooze and wake in precisely the same way each time. If the programming of its behaviour-trees is more sophisticated, it may rotate or shuffle between a handful of different reactions; these will occur regardless of the activity it is triggered by.
In his new book Games: The Art of Agency, C. Thi Nguyen argues that video games enable us to adopt alternate modes of agency. While we agree with this main contention, these modes are currently restricted by the interactive limitations of NPCs. The trouble is that it is always possible to tell who is a player-controlled character and who is an NPC. This reduces the lifelikeness of any given game and curbs the socialness of our agency.
There are two over-arching ways of interacting with NPCs: via text-based and via non-verbal behaviour. Whereas text-based interactions break immersion because of their repetitiveness and conversational loops, non-verbal actions performed by NPCs are frequently rated as unintelligible and stupid, due to their lack of any clear relation to the current game situation. Game AI design thus faces the challenge of rendering interactions with NPCs more realistic across transactional, hostile, and relational scenarios.
In 1950, the philosopher and mathematician Alan M. Turing introduced “the imitation game” as a verbal test for whether or not we would describe a machine as thinking or intelligent. For an NPC to pass the test, we must not be able to tell whether or not it is controlled by a player. But Turing’s test is only applicable to text-based interactions with NPCs. So what might a parallel test for non-verbal behaviour look like and what exactly would it be testing for?
If one observes a honeybee in flight, it is not unusual to assign meaning to its behaviour such as: “oh look, now it is gathering pollen” or “now it is taking resting on a flower…before it flies off to carry her bounty home.” At no point is one suspicious of the bee’s natural intelligence as it seamlessly parries each hand waggle.
Whatever one’s view on whether bees may be said to possess intelligence or whether their dances constitute a language, lay people interact with them very differently than they do with NPCs. What makes bee behaviour pass as authentic where that of NPCs fails? It is not that only the former are in the “real world,” for we can interact most naturally with other players on the video screen. Rather, when we are wafting, panicking, or staying as still as possible, we treat the bee as a living organism that detects our movements and responds accordingly.
NPC actions are not dissimilar to those of bees: they evade attacks, ambush, run away, and summon back up. But their execution of them somehow falls short. While the behaviour of an NPC may equal that of a bee in its degrees of annoyingness, it provides no match with respect to naturalness. It might be objected that artificial intelligence by definition contrasts with natural intelligence. But there is no reason why the imitation of intelligence should not include that of naturalness. Insofar as the traditional Turing test is considered. NPCs can seem vastly more intelligent than cats and dogs, let alone birds and bees. And yet they come across as mindless and repetitive in our interactions with them.
Game designers have long copied the natural physiology of human and animal movement. But the AI community’s historic adherence to some kind of mind-body distinction has led them to view this goal as entirely separate from that of producing intelligence. As a consequence, NPCs typically behave in ways that are baffling at best. To improve, game design must move out of the shadow of Descartes and simulate an enactive intelligence that cannot be separated from non-verbal behaviour. This involves the programming of behaviour that may be seen as a natural and appropriate response to salient aspects of their environment. Only then can NPCs win the only imitation game that counts.