Let’s talk about AI because obviously everyone has to. When it comes to gaming, AI has been around since the earliest of days. The very first usage for AI was a checkers-playing program that debuted—wait for it—71 years ago.
Since then, AI and gaming haven’t separated. Old fogey chat rooms and gamer forums are filled with what I’m confident are the earliest and most wide-spread consumer discussions about AI, with gamers ranting about “AI” enemies that were overly dumb or hard, decades before investors decided it was cool.
While lots of people (including yours truly), have written about how awesome the newest fad—generative AI—could be for game makers in speeding development or for gamers to harness powerful UGC tools to create their fantasies, I want to focus on the above: the lowly in-game AI. Particularly, how AI can unlock new categories of games by allowing in-game characters to vary more and react more nimbly to more diverse player behavior and manipulated environments. And I’ll throw in how web3 can play a role in this, because why not.
The NPC
Right now, the most common application of AI in gaming is the NPC (non-player character). Typically, NPCs serve specific functions in a game’s world:
Some are neutral, like a character walking along a roadway
Some are helpful, like a combat companion who fights by your side
Some are hostile, from run-of-the-mill chess opponents to big game bosses
And, normally, they stay in that specific role—though some games (from Minecraft to GTA) have examples of neutral characters that will turn hostile if threatened or attacked. However, if you leave them alone again long enough, they tend to reset to their default state, i.e. they won’t hold a grudge and take it out on you “days” later in the gameplay.
They also tend to have niche domains and lives; the shopkeeper stays in the store and the Yeti stays in the ice biome. They’re trained on certain tasks repeated with consistency every day. They don’t have families, they don’t get sick, their tastes in music don’t change.
Even in spectacularly advanced open worlds like Red Dead Redemption II, where many NPCs do have comparatively vast lives (home, work, lunch, work, nightcap, home), they tend to rinse and repeat the same processes every day, regardless of the weather or day of the week or if their dog died the night before.
AI needs to be trained on information, in the sense that you can’t just say “be an enemy” and hope a villain knows what to do. In-game enemies are created based on a set of rules they have to follow (how to move, attack, etc.), with models of how an enemy in a given game should act (this is how you shoot an arrow or throw a grenade), and that behavior reacts to models trained on how human players will engage with it. The datasets that get fed into this are naturally limited, so there won’t be any wildcard actions or reactions expected. This means that even the most advanced AI enemies are ultimately predictable, even in cases where they can adapt—to an extent—to how a player plays (like Metal Gear Solid, where players that repeatedly infiltrate enemy bases at night will be met by more and more hostiles with night vision goggles).
But models and data sets are improving so dramatically, that developers can now harness NPC structures that are dynamic and vast in possibilities.
Reactive worlds
While we’re here, the environment of a game is a kind of NPC in itself. Some worlds you can manipulate (e.g. build a house, chop down a tree, leave a footprint in mud), but in general worlds tend not to react to player events, and certainly not with persisting, macro changes. I’m unlikely to cause an avalanche with a gun blast, or burn a forest and see the soot darken all the windows of nearby houses.
And the changes that are allowed tend not to impact the behavior of character NPCs. You can’t starve your enemies to death by introducing an invasive plant that then wipes out the food source for the only protein they consume: deer.
Narratives of the future
In my mind, all the other applications of AI that are getting attention right now (like the ability to have characters say anything in real-time that wasn’t recorded by a voice actor) should be in the service of creating the most complex storytelling and unique gameplay that’s ever been seen (not taking voice actor jobs: hot take), driven by NPCs and environments that react in new ways. I hesitate to say “in the most realistic” ways, which is where some AI conversation goes. I don’t really want characters to behave like everyday humans, but I want the range of responses and behaviors to balloon, and my ability as a player to be at the mercy of—and react to—those to increase. (Of course, this won’t and shouldn’t work for all games.)
I’m imagining a game’s final boss, who follows and thwarts and threatens me at every turn of the game’s journey. But who also has a life. He eats, sleeps, knows other people, shops. With huge NPC AI capabilities and an adaptable and reactive world, there are potentially infinite ways for me to take him down.
I could steal something from a store he frequents and frame him, so the shopkeeper gets mad and has him thrown in jail overnight, buying me time.
What if I follow him and find his best friend, whose son I kidnap and throw in a well, knowing that the villain—secretly empathic in his private life—will want to help the friend track down his son, leading him into my trap.
How about sabotaging a water main at the start of the game, so it slowly weakens the foundation of his house, leading to expensive repairs and his ammo stores put into storage.
I slowly isolate him from his network, his wealth, and all his evil resources. I turn neutral NPCs into unwitting allies for me, and my enemy’s world against him.
Obviously, he’d get to know my tricks and react. Go to all his friends and tell them if they see a guy like me, don’t trust him. Move his family to a hotel. Discover my favorite routes around the world and leave purposeful misdirection in the path. Hell, the game could get meta and make note of me, as the gamer, getting sloppy in my controls at the end of my usual two-hours sessions, and using that as the exact moment my nemesis pops a task that requires intense dexterity.
These examples aren’t necessarily that amazing in themselves; I’ll leave it up to actual game designers to create something that isn’t a banal life simulator, or otherwise tedious and frustrating. But you get the gist. If you’ve played Deathloop, it’s like that on crack.
Web3
Now here’s where things get beyond funky. As more of the consumer world starts to have on-chain underpinnings, my wallet will increasingly be a digital fingerprint of my existence across every domain of life, from the music I love to the car I own.
AI + web3 will create the most customized gaming experiences in existence.
How about a game that lures me into a false sense of security by playing my favorite jams, so while my nemesis sneaks in through a back window I’m bopping around to Rusted Root?
Maybe it notices I’ve got a bunch of shooter game NFTs in there (so I’m probably pretty decent with a gun) but no puzzle games, and it opts to give me a real twisty decoder task rather than a hallway of henchmen.
Leaving possible dystopian (or at least intensely unnerving or annoying) gameplay to the side, the potential for game mechanics that counter my in-game tactics and play with my IRL emotions and behaviors opens the doors to some next level narratives and NPC development.
Which is exactly where all the generative AI capabilities come into play. As an on-chain-verified nature lover, the game might opt to construct a jungly level rather than an antiseptic hallway. A Duolingo NFT pass for Spanish lessons could throw some dialog into Spanglish mode. The opportunities are terrifyingly endless, but I love the idea of developers having more tools to mess with me.
And the nice thing is that gaming and AI have been joined at the hip since the beginning, so designers are already primed to think about how to use this technology to build increasingly cool experiences. Bring it on, Hal.