In Spike Jonze’s new movie Her, Joaquin Phoenix is an introverted writer on the verge of divorce who falls in love with his computer’s intelligent operating system. Exactly how far-fetched or credible you think that is probably depends on how invested you are in technology. It was after all, inspired by the real-life web application Cleverbot, which lets visitors engage in conversations with an AI program; and in a lot of ways the movie is a study of our growing reliance on devices as mediators in our social lives and love affairs. From Siri to Tinder, our smartphones and tablets are simultaneously humansing themselves while mechanising our relationships with other humans. How long before we fall for the devices themselves?

Really though, this is already happening in video games, where non-player characters are becoming ever more complex, both in terms of visual representation and “intelligence”. In narrative adventures like Mass Effect and The Walking Dead, players stay with the same characters over the course of several games, communicating with them through simple conversation trees and trying to keep them safe through endless dangerous encounters with aliens and zombies. Mass Effect even allows players to form sexual relationships with other characters, which can have a profound impact on the gaming experience – even if the sense of reciprocity is minimal. In Mass Effect, the love affairs are story functions and the AI characters merely pawns in a set of narrative possibilities. They don’t love you back. Not really.

But what if they could? In some ways this is more likely than the scenario envisaged in Her. Unlike operating systems, which are the practical interface between us and the computer, video games provide a playful environment in which we’re invited to identify with the onscreen avatars and events. Just as literature has done for hundreds of years, games invite us to identify with characters, they function to make us feel. It may be just a matter of time before a clever game designer uses cutting edge AI routines to simulate emotion. We know that basic speech recognition and conversational abilities are already possible, but can these be extended from chatbot programs into sophisticated game characters?

Will we see AIs capable of falling for players?
An intelligent question

The first challenge here is in defining the “intelligence” part of artificial intelligence. Although huge advances have been made in academic AI research involving elements such as affective computing and machine perception, in video games the concept is still mostly confined to navigating the game world and responding to basic player actions. An AI character in an action game will be able to find a route through the environment, and will be able to hear and see the player-character, responding with a series of set actions. But they’re little more than marionettes or remote-controlled robots that have been given basic motor functions.

The next step perhaps is to go from spatial to emotional pathfinding. “One thing that’s really important in social interaction is being able to think about what someone else might be thinking – it’s how we learn and plan things,” says Michael Cook, a researcher at Imperial College who has built an AI that can design its own games. “When you’re designing AI, you mostly focus on the end result, what a character actually does and whether that’s intelligent. But for us to relate to characters, we might need to make AI mimic the thought processes we go through as well. The way an AI plans to walk across a room in an FPS game is completely different to how you or I plan the same route. I think thought processes are a really interesting area for AI generally, since making believable mistakes and being outwitted in games also need this kind of work to be done.”

For love, there has to be true adaptive intelligence. The main problem is that we’re not sure what that means. “True intelligence is a big phrase that hides a multitude of sins,” says Cook. “Is my dog truly intelligent? It can definitely do things that many video game characters can’t – it can learn actions, it can connect sounds to meanings, it can make friends with people and remember them – but it’s not the kind of intelligence we want from game characters. I think we’re getting closer to a new generation of game character, but the hard problems remain hard, and one of the biggest in terms of relationships with avatars is communication and language. We’re a long, long way off having natural conversations with characters that don’t slip right into the uncanny valley.”

The problem is that we don’t actually know about a lot of the stuff we’d need to simulate in order to make a game character who could fall in love with the player. ”When we talk about going beyond AI that can win the game, we get into really murky territory of how much we actually understand the human condition,” says AI researcher and game developer Luke Dicken. “I think that we’re both quite close and also a long way away from this kind of NPC in games.

“From the code or computer science point of view, it’s a horsepower problem now and that’s just going to get solved over time either by streamlining techniques or extra muscle in PCs, so we’re close-ish in some ways. On the other hand, from a behavioural science point of view, we don’t understand words like ‘intelligent’, ‘characterful’ or ‘deep personality’ and what that truly means in a deep connectionist kind of way – we know it when we see it, but that’s not necessarily enough to reliably construct something, and when we miss, we miss by a long way.”
Mind the gap

But the thing about games is, players are more likely to fill in the gaps themselves. With a ‘thinking’ operating system, we’re starting with a dispassionate productivity application; an operating system is designed to be useful not loveable and it exists as a functional system. But in games, there’s often a narrative, there’s a script and there is playful interaction – games like Mass Effect provide an environment in which we want to establish relationships with characters – we naturally humanise the experience, because, look, there are people on the screen and they’re doing interesting stuff. Games usually have very simple stories and very broad characteristations, but because we’re invested in the experience we imprint emotional depth on to the framework. So surely the task of generating a convincing, attractive AI is a little more simple for game makers than OS designers?

“Characters can generate really strong emotive feelings in us already,” says Dicken. “Telltale Games do this really well – you want to protect Clementine and I definitely fell a bit in love with Snow White in The Wolf Among Us Episode 1. Right now though, I think a lot of that emotive nature is stemming from putting ourselves in the role of the player character. Crystal Dynamics infamously talked about trying to make the player feel protective of Lara Croft in the latest Tomb Raider and I think that’s a mistake not only because of the gender issues it raises, but because their job is to make the player be Lara. Did I fall a bit in love with Snow White, or did my version of the player character, Bigby Wolf? It’s a blurry line there.

“But stepping back from games a little and looking at something like Lionhead’s Project Milo, this was a kid that you could play with, he had toys and a personality and was super expressive. We’re still talking about virtual characters, but in this setting you as yourself get to interact with the character – this is a stark contrast from most games which are still about ‘role-playing’. This puts a divide between the player’s emotions and the game world, so I think an emotional attachment really needs to stem from ourselves not from our perception of the character’s self.”
The social scene

So let’s assume that we begin to see a new era of games in which gamers are invited to play as themselves, and to establish unmoderated relationships with AI characters. What then? “One big step forward I can see on the horizon is subtle social interaction,” says Cook. “The Impulsion Project, which I saw at AIIDE in 2012, blew my mind. This kind of stuff in games would go a long way to boosting the perceived intelligence of NPCs.” The Impulsion project is a social intelligence engine developed by Claudio Pedica at Reykjavik University’s AI lab. It seeks to generate lifelike interactive characters who understand their surroundings and can grasp the meaning of player interactions. Pedica has worked with Icelandic game developer CCP to explore the possibility of implanting his tech into future titles. And once you have AI characters who know where they are, what they are doing and what the player wants, we’re getting toward relationships.

Closer, but not quite there – games have to start to think about non-player characters in a different way; not as functional automatons in a narrative sequence, but as functioning beings that will generate a range of responses to the gamer – because range is important. “One of the real things here is how we interact with stuff,” says Dicken. “In general, people can’t make meaningful attachments to things that they can only interact with in very limited ways. That’s why natural language processing is so important, so that the character can understand what we are saying and that has symbolic meaning to them. Right now our complex romantic interactions in a lot of games boils down to ‘Press X to sexytime’ and a few canned pieces of dialogue and for some people maybe that’s enough, but it’s not what most people would consider sufficient for a meaningful relationship.”

Perhaps the most interesting element in all of this is the idea of reciprocity. In the movie Her, the AI operating system responds to the lead character and their relationship is mutual – there is a sense of shared enjoyment. Can a computer appreciate conversation? Can it get jokes?

Analysing the likelihood of this scenerio recently, Gartner researcher, Jackie Fenn, wrote: “Humour and creativity will be among the more challenging areas for artificial intelligence, but even here researchers are experimenting with clever algorithms and deep learning. If a computer can learn what makes people laugh – and more importantly what makes you laugh – based on watching and analysing over time, there is no theoretical reason that a computer couldn’t eventually display and respond to humour. Similarly with music or art – by experimenting, analysing and learning, it could figure out which compositions create the best emotional resonance in the human brain.”