AI Unplugged: The Secret Life of a Background Character

NPCs have been a part of role-playing games for decades. Thanks to #AI, they may finally get the chance to fulfill their true potential.

What's a NPC?

Non-Player Character (NPCs) were originally created for the tabletop role-playing game Dungeons & Dragons. They were all the other characters and creatures played by the game master. Their name is a hint -- they're everything not played by the players themselves, who (usually) only role-play as one character.

As a result, NPCs could be as rich or two-dimensional as the game master playing them. It's a running joke that background characters who may seem unimportant can easily become the focus of the game, with players adopting NPCs as pets or sidekicks.

When video games arrived, NPC interactions were primarily limited to foes to be dispatched. The mechanics of combat were much easier to replicate than say, parley (a valid tactic that could be done when engaging with dragons). As video games evolved, NPCs evolved with them, but they are largely relegated to decision trees, in which players select one of several branching options to engage with a NPC. The outcome creates a new branch of options (or might limit the options, or might result in combat), with random checks providing some variability as to the outcomes.

The interest in NPCs perhaps most visible in the recent Baldur's Gate 3, which is an official D&D-licensed game. Players love the ability to romance companions, a feature that has been available in prior games, but has come to the forefront both with the maturity of the subject matter, better graphics, and even more complex conversational outcomes.

To get a sense of how much work is involved in this process, Red Dead Redemption 2 provides a template. Its main story consists of 2,000 pages, with 2,200 days of motion capture work, requiring 1,200 actors. The finished game includes 300,000 animations and 500,000 lines of dialogue.

An AI could considerably cut down on all these costs, eliminating much of the motion capture work, acting work, and dialogue creation (something SAG-AFTRA and the WGA are rightly concerned about). No wonder then, that many companies see an opportunity to further create immersion with NPCs powered by AI.

The AI NPC

Generative AI can already engage in persona play. That is, you can train many of the generative AI to adopt a persona as if they were speaking and acting as someone else. Some AI resist this more than others; notably, convincing an AI to "be someone else" is key to also getting it to break its own protocols, so some AI resist this strongly (Gemini and Copilot being two examples).

Once an AI has instructions to be someone else besides a friendly, helpful assistant, it doesn't take much for it to then act as that character. The challenge isn't so much in getting a peasant farmer to talk -- it's to ensure the peasant farmer only talks about farming. Or to put it another way, generative AI flips the typical NPC challenge on its head: instead of worrying about covering every possible conversational outcome, the key is restricting the NPC's knowledge to what it should plausibly know about.

The other problem is that generative AI is a network of systems, not a portable piece of code. Connecting it to an online database requires constant queries, one for every player and for every NPC. This can add up and obviously requires the game to be played in the cloud.

And of course, there's plenty of other factors to consider. NPCs are typically bound by code and geography, so an AI innkeeper can't just ask you to visit their garden when there 1) is no garden to visit, and 2) they can't leave the confines of their inn.

Despite those limitations, an immersive NPC who can interact with players within a framework is still feasible, and many video game startups are banking on it.

The Possible

Co-founder Artificial Agency demonstrated just how feasible a NPC in a game world might be with a five minute demo featuring a Minecraft NPC named Aaron:

At one point, Kearney’s in-game character asked Aaron to gather supplies for a scary mining adventure, and though it wasn’t programmed to do so, the NPC visited multiple chests to gather armor, helmets, tools and food, and delivered the supplies back to Kearney’s character. And when Kearney told Aaron she was gluten-free after it brought back some bread, the NPC apologized, and offered a gluten-free option instead: cooked chicken. The simple demo showed how Artificial Agency’s AI NPCs could not only talk, but perform complex actions without being explicitly told to do so.

This kind of interactivity comes at a cost. CEO of Artificial Agency Brian Tanner estimates the demo costs $1 in AI inferencing costs, but that the price will come down to a 1 center or less. Artificial Agency is not alone:

The field is still very new, but it’s extremely hot. In 2022 the venture firm Andreessen Horowitz launched Games Fund, a $600 million fund dedicated to gaming startups. A huge number of these are planning to use AI in gaming. And the firm, also known as A16Z, has now invested in two studios that are aiming to create their own versions of AI NPCs. A second $600 million round was announced in April 2024.

To get a sense of what an AI/NPC narrative interface might look like, Microsoft recently released a paper, “GENEVA: GENErating and Visualizing branching narratives using LLMs(opens in new tab),” presented at IEEE Conference on Games 2024. The tool uses graphs to visualize high-level narrative descriptions and constraints, building on the generative capabilities of GPT-4. You can see how it operates in this demo:

Oh Shut Up

Go too far with this line of thinking and we end up with secondary characters taking over the spotlight.

The reality is that not every game needs to be connected to ChatGPT's powerful network to be convincing enough to pass as a full-fledged personality. Indeed, part of the fun of a game is playing through a story, and if NPCs can derail it at any time, it's less of a game and more of an experience. This can certainly be fun in its own right, but there are downsides:

Sometimes, the possibility of everything is too much to cope with. No Man’s Sky, a hugely hyped space game launched in 2016 that used algorithms to generate endless planets to explore, was seen by many players as a bit of a letdown when it finally arrived. Players quickly discovered that being able to explore a universe that never ended, with worlds that were endlessly different, actually fell a little flat. (A series of updates over subsequent years has made No Man’s Sky a little more structured, and it’s now generally well thought of.)

The compromise is likely somewhere in the middle: a game that has a database of relevant info that feels native to the game world, without having every NPC know every aspect of the entire universe. By bounding the game's creativity enough so that it keeps the player on track but giving the NPCs enough leeway to express those facets of the world in their own way, that might be just enough to make a believable fantasy immersion that's still fun.

We plan to test this out ourselves with the Multi-User Dungeon, RetroMUD. Wish us luck!

Please Note: The views and opinions expressed here are solely my own and do not necessarily represent those of my employer or any other organization.