Coma AI

Suprise, I’m alive. No time to explain, here is a quick thought on Game Design. Game Designers and AI Theorist like to discuss the conditions that would accurately define what kind of qualities an Artificial Intelligence should exhibit. There are things like the Turing Test but that’s a pretty fuzzy one, focuses on text interaction and of little help for Game Design.

I came up with another approach. It’s not really practical but pretty fun… ok, it’s a very dark sense of humor but stay with me. There are several systems for categorizing the state of consciousness for coma patients. They have been based on the real impairments those patients exhibit. Interestingly, they faithfully reflect a bunch of things we take for granted when interacting with a person. They also describe the things we might notice as MISSING when a person fails to perform them.

For example, here is the Rancho Los Amigos Scale of Cognitive Functioning (more after the jump):

LEVEL 1 – No response. The patient appears to be in a very deep sleep or coma and does not respond to voices, sounds, light, or touch.
LEVEL 2 – Generalized response. The patient moves around, but movement does not seem to have a purpose or consistency. Patients may open their eyes but do not seem to focus on anything in particular.
LEVEL 3 – Localized response. Patients begin to move their eyes and look at specific people and objects. They turn their heads in the direction of loud voice or noise. Patients at Level 3 may follow a simple command, such as “Squeeze my hand.”
LEVEL 4 – Confused and agitated. The patient is very confused and agitated about where he or she is and what is happening in the surroundings. At the slightest provocation, the patient may become very restless, aggressive, orverbally abusive. The patient may enter into incoherent conversation.
LEVEL 5 – Confused, inappropriate but not agitated. The patient is confused and does not make sense in conversations but may be able to follow simple directions. Stressful situations may provoke some upset, but agitation is no longer a major problem. Patients may experience some frustration as elements of memory return.
LEVEL 6 – Confused but appropriate. The patient’s speech makes sense, and he or she is able to do simple things such as getting dressed, eating, and teeth brushing. Although patients know how to perform a specific activity, they need help in discerning when to start and stop. Learning new things may also be difficult.
LEVEL 7 – Automatic, appropriate. Patients can perform all self-care activities and are usually coherent. They have difficulty remembering recent events and discussions. Rational judgments, calculations, and solving multi-step problems present difficulties, yet patients may not seem to realize this.
LEVEL 8 – Purposeful and appropriate. At this level, patients are independent and can process new information. They remember distant and recent events and can figure out complex and simple problems.

Reading through the list, you might see what I’m getting at. Applying the list on to videogames works amazingly well but yields some troublesome results. Nowadays, the best you can expect from a Video Game AI is Level 7 (Automatic, appropriate). And even that is considered pretty advanced. Some of the things from Level 6 (Confused but appropriate) haven’t been done in games yet. I haven’t seen NPC’s dress themselves or brush their teeth. It’s already quite the feat when NPCs have a daily schedule like in Fallout 3 or Oblivion. But even then, I haven’t seen them perform any hygiene and I haven’t seen them change their clothes (ew!). And the need for help discerning when to start and stop? Defiantly something a AI Programmer can tell you stories about.

Level 7 mentions some thins like “difficulty remembering recent events and discussions. Rational judgments, calculations, and solving multi-step problems present difficulties, yet patients may not seem to realize this.” Well that sound exactly like the AI of Assassin’s Creed.

Level 5 (Confused, inappropriate but not agitated) is what you mostly get as “Wingman” AI for Team FPS Shooters. For example, this is the AI for your Teammates in Mass Effect. They talk nonsense, have no clue what’s happening but if you issue an order, they will follow.

Level 4 (Confused and agitated) is the standard AGGRO AI. You encounter any creature in Final Fantasy? It will attack you no matter who you are. Even games with advanced AI like Fallout 3 have them. Wandering through the Wastelands, People and Animals will attack me for no apparent reason. There is no way you can talk them out of it. I’m reminded of the epic “I want to heal the wolves” Escapist Article.

Level 3 (Localized response) is what was revolutionary about the AI of Half-Life 2. The NPC’s were actually looking towards the player while they were speaking. It’s pretty amazing how long it took us to get this right and how fundamental that feature is for considering somebody conscious. However, it’s far from being enough. Here is an example of the Oblivion AI exhibiting Level 3 consciousness. They look at things alright, but they fail to react to the general situation:

Level 2 (Generalized response) reads exactly like the description of the AI for most platformers. It’s pretty much the AI of a Goomba from Super Mario Bros.

So basically, game AI displays the whole spectrum of brain trauma victims. No game AI managed to get to Level 8 yet, although I think Facade did pretty well. We have a long way to go if we want believable characters in games. On a more serious note, I think looking at the medical definitions of consciousness might help to deal with the problems. The Scale I presented here is actually a very simple one. There are others which are much more detailed and might inspire new developments and help to get some priorities straight.

Krystian Majewski

Krystian Majewski was born in Warsaw and studied design at Köln International School of Design. Before, he was working on a mid-size console project for NEON Studios in Frankfurt. He helped establish a Master course in Game Design and Research at the Cologne Game Lab. Today he teaches Game Design at various institutions and develops independent games.

3 responses to “Coma AI”

  1. Yu-Chung Chen

    Hahaha, great find, how well the levels fit! Refreshing from all the (non-explanatory) citations of uncanny valley.

  2. Daniel 'sirleto' Renkel


    this is something new, for gamedev. hopefully some people with time and interest in AI as in adventures and such, where the ai "characters" play an important role, will stumple upon this…

    i personally believe, level 7 is not what games have reached. its more like ai looks like each level from 5 upwards but is actually 2 levels lower ;)

  3. Krystian Majewski

    @Daniel – well, it's ONLY about appearance and impression. Obviously, AI nowadays isn't even able to master bipedal walking properly.
    I would judge them for Level 7 in special cases and in special scenes. Otherwise, I'd agree that 5 maybe 6 is pretty much what you usually get for high-end AI.


The Game Design Scrapbook is a second blog of group of three game designers from Germany. On our first blog, Game Design Reviews we describe some games we played and point out various interesting details. Unfortunately, we found out that we also need some place to collect quick and dirty ideas that pop into our minds. Hence, welcome to Game Design Scrapbook. You will encounter wild, random rantings. Many of then incoherent. Some of them maybe even in German. If you don't like it, you might enjoy Game Design Reviews more.


follow Krystian on Twitter
follow Yu-Chung on Twitter
follow Daniel on Twitter