Games like HL2 and FEAR are legendary for having excellent AI. Recent titles (as of 2018) leave much to be desired. With improvements in our understanding of probability related state machines, and general machine learning (or AI as it's called sometimes) improves, it generally feels many commercial titles forego the development of good AI. Anecdotally, I have personally experienced in recent titles (where the AI is dumb as per say - far cry, wildlands, etc.). I feel having dumb AI may sell games more (giving a more accomplished feeling, pushing players towards multiplayer, etc.). So, let me pose the question, does any research show that bad AI leads to more sales?
Bad is highly subjective & requires some context. In an FPS for instance, the an AI that never misses a shot would have optimal scores, but arguably wouldn't be very fun.
Academic research in this areas has some significant difficulties. First, there's very little publicly available data about the sales of games. Second, companies rarely report such things when they don't reflect favorably on themselves. This sort of biased small pool of data is problematic for conducting rigorous statistical research.
That being said, there is some research that draws conclusions about player perception of AI & the experiences provided by games. In particular, "What You See Is Not What You Get: Player Perception of AI Opponents" by Baylor Wetzel and Kyle Anderson in Game AI Pro 3 draws the following conclusions:
- Players are not particularly good at estimating AI difficulty
- News is mixed regarding whether or not its worth investing time to develop complicated AIs
- Perceived difficulty correlated highly with realism but the correlation between actual difficulty & realism was not equally high
- Players were not particularly good at detecting cheating
- AIs that did cheated routinely went undetected
- Players readily accused AIs cheating when they did not
- Players were not good at inferring even basic AI strategies
- Players will project / invent stories onto AIs & see motivations, biases, cheating & desires where no exist
I'd wager to say that AI is one of those aspects which contribute very little to game sales because it is an aspect which is very difficult to use for marketing.
The easiest aspects to advertise are graphics and aesthetics. A single screenshot is enough to tell the customer at first glance "this game looks awesome".
The text description of your game is a good way to impress with the amount of content, because you can simply express it in numbers (100 levels, 10 different environments, 200 side-missions, over 80 hours of playtime...).
A trailer is a good way to give the player an impression about how the gameplay feels like. You can use it to highlight some unusual game mechanics (although not every game mechanic can be presented well in just a few seconds of screentime). It can also be used to give a glimpse of the soundscape.
But how do you market a good AI? Whether an AI is fun to play against or not is something you only notice after several hours of gameplay. It might be worth writing about in a longer review article. You could show a few interesting tidbits in a more documentation-oriented trailer. But unless your AI does something truly innovative and extraordinary, then those won't be particular interesting.
So if you design your game for maximum marketability, then AI is one of those aspects where you get the least return of investment for your development budget.
And yes, developing good AI is expensive. The better your AI, the more complex its programming becomes. More complex programming means more obscure interactions between different game features which means more corner cases for bugs to hide in. So your testing effort increases exponentially with the complexity of your AI. A complex AI is also really annoying for the level designers. The level designer might say "I want a huge battle to take place in this location, so I design it as the perfect battle arena with lots of interesting cover and tactical opportunities and put some enemies in it". But then the AI decides "Having a battle here would be strategically unwise, I will relocate and fight the player in this bland and boring environment instead". So you need to keep testing all the levels you designed to make sure that the AI still behaves the way the level designers intended.
What is good A.I.?
If we can agree on bad A.I. is one that accomplishes nothing. Then the problem is defining good A.I. because that will depend on what is the objective of A.I.
Often times, the objective is not to beat the players. Usually, we are not making games expecting players to pay to be defeated, and if we were, we would just make the game unfair.
Instead, the objectives of A.I. in games are to keep the game:
Usually, game are more successful when A.I. can simulate the skills of a regular human player – instead the best human player. I will remind you that human players are not perfect. In fact, a human player will not have full knowledge of the environment, will have limited reaction time, can be tricked or deceived, can be emotional and irrational, and has a tendency to fall to routines.
In addition, you can usually go without one of these - depending the genre. For example in a crowd-combat fighting game, the A.I. can get away with dumb agents - for the most part - because you kill them right away anyway. Other games will have easily exploitable A.I. – Some streamers may like that, and thus, provide a kind of publicity - but that is not a good fit for every genre or setting.
You have a budget
Well, you have a few budgets (a monetary one, the time to the deadline, and the work per frame), and you have to decide on what to spend them.
You will find that A.I. will be at the bottom of the priorities. Visuals, audio, UX, and writing usually come first. That is without talking about all the underlying technology, such as proper logging, networking, and so on.
I want to make a point about writing: It is much cheaper and much more effective. Spending some extra time in thinking what the NPCs do - when they are not talking to you - and what their relationships are, and some extra time in the history of items, and some extra time in writing more than two lines in narrative, dialogs, flavor text, and lore... will go a long way in selling the world as bigger and alive.
Futhermore, if you can tie good writing to a simple A.I. it becomes much more effective. For example, if your enemies express their state and actions in dialog and animation (think, for example, in classic Metal Gear, how the enemies express they noticed something or are searching for the player) it makes the NPCs seem smarter.
Should you aim to have a bad A.I. as an attempt to sell more? No, you should not. You should set reasonable and viable goals for your A.I. and make A.I. that is good at those goals. Even if the goal is to make the NPCs seem stupid. Because, I assure you, that is a thing. It can even be part of the lore of the world. In fact, you will learn to see that as be good A.I. once you accept that the objective is not to beat the player.
Addendum: A.I. should be able to beat the player. I want to say that it is better if it feels fair to the player. When I say that the objective is not to beat the player, I do not mean that A.I. should try to not beat the player... instead that you are not designing A.I. with the objective to beat the player.
On making better A.I.
Making A.I. is hard. In particular making A.I. effective in beating the player - which, as I said, should not be your objective - is hard. At least if you want to go beyond planning algorithms and state machines, which can be very predictable.
For example, you can write planning algorithms that adapt to usual player behavior. It can even anticipate the player if you model the player behavior, which can be done with influence maps and similar technologies. Even that will be predictable, players will learn that if they keep doing X, the A.I will eventually do Y, and that can lead to exploitablity.
To break predictability, adding random to the algorithm goes a long way.
If that is what you consider “bad” A.I. you probably want some form of modern machine learning; you can try something like NEAT, and I would suggest finding an existing solution so you do not worry about implementation details. Then you probably will NOT have it run in real time because training agents takes time. Thus, the main use in games is to train agents to play the game and ship the trained agent.
Even if you were to train the agents during gameplay, you should start with decent agents, because if you start from scratch, it means the player will encounter dumb A.I.
As a consequence, usually the improvements during gameplay are not significant enough (in particular, if the player kills each enemy mere instants after it appears on screen). Meaning that this is often a waste of resources.
Futhermore, if the improvements during gamplay were effective, remember your difficulty curve. An A.I. that improve out of your control will make it hard to design for it and keep a good difficulty curve, leading to player frustration.
Ah, and generational machine learning - while results in improvements in the long run - can have dumber offerings. In addition, if you have seen anything about adversarial machine learning approaches, you know that even well trained A.I. can be exploitable.
I repeat, making A.I. is hard.
A.I in F.E.A.R? It is basically goal oriented planning with smart objects. A lot of the work is in the objects, not the agents. It is not much different than what you find in The Sims. In fact, we could argue that compared to The Sims, the A.I. in F.E.A.R is "bad" because the agents have less knowledge. Also, there is no machine learning in F.E.A.R. See Assaulting F.E.A.R.’s AI: 29 Tricks to Arm Your Game.
Will bad A.I sell more?
Truth resists simplicity
-- John Green.
Terrible A.I. will not sell. Because terrible A.I. results in NPC who stand in place and do nothing.
I believe there is a sweet spot where bad A.I. can sell. In particular, if you have a funky world, where dumb enemies make sense, or you are going for the comedic effect.
In general, good A.I. will sell better. Except good does not mean good at beating the player. A good difficulty curve is more important that A.I. effective at beating the player. Instead good and making the game interesting, believable and fun.
You can do plenty of other things that you can improve – aside from A.I. - that will be more effective for marketing. Although, that does not mean you should go for bad A.I.
Addendum on studies: Give me a metric on which we can quantify how "bad" A.I. is, and then we can talk about research on bad A.I. leading to more sales. Hmm... after controlling for all other variables of course... to isolate the effects of A.I. in sales we will also have to quantify visuals, audio, etc. I haven't found any study on game A.I. influencing sales (I find a lot on using A.I. for sale force automation) - I do not claim to have searched exhaustively - but won't be surprised if there aren't, in particular considering that a lot of game development happens without involment from academia.
Seeing the other answers, I may twist this question to biology rather than CS. While I work in the field of AI, this is much more about mammalian psychology when it comes to games (which seems to be semi-universal among the species.)
While I can't find the studies myself very well, here's a clip of Jordan Petersen taking about Games. https://www.youtube.com/watch?v=JI9n2yC14fU Around 1:40 he starts on the rat game. This is crucial to your question. Note: Whatever your thoughts on his views or arguments in general, this is unrelated. I simply couldn't find the study myself
The point of this is that studies are now showing that mammals play games in general, and that they only continue playing if they win an "ok" percent of the time (~33% for the rats.) If the AI is so difficult that it's impossible for the human (such as a good Chess AI against your average player); the human will stop playing.
If, instead, the human sees a challenge but can beat it, they are likely to continue. You need a glimmer of hope. Side note: while this is hardly academic wording; the more "casual" the player, the bigger the glimmer of hope needs to be
Why AIs are difficult What makes this difficult is what I just pointed out; people need different levels of difficulty based on how tenacious they are; which also means different approaches based on the game type.
Chess is simple: reduce the search space and (by virtue of choosing a random move among equal moves) introduce randomness; which is why Chess AIs are successful for casual and professional players. FPS is a bit harder; in general the AIs tend to work the same but they increase/decrease the amount of opponents and the damage scaling.
"Bad AIs" and Human Competition
The thing about more recent "Bad AIs" are more likely a trend of the increasing requirement of having multiplayer (or being ONLY a multiplayer game.) Players will naturally be able to beat some people and lose to others, and a skill-ranking system helps to sort them into buckets of satisfying difficulty.
AIs, thusly, are only really required for time-filling or for story-driven content; which is taking a backseat in the last decade by comparison to previous decades. Why build a convincing AI when you can generally get the result of an enjoyable game by pitting players against players?
The TL;DR version of my post: While this is an indirect answer; humans who want to beat things will often only continue playing if it's possible and rewarding. This tends to be easier by simply adding multiplayer than spending countless debugging and development hours making a convincing Ai.
Truthfully, 'bad AI' can be very loosely branded. The AI may not be a very difficult opponent, it may be very simple, or the AI writers were just plain lazy. I am going to go off of the simplicity side of it.
The reason why some games with simple AIs sell better is because we, as humans, want to understand things. If I were to write a several-thousand line file for the AI of an enemy, it would use a lot of random and conditionals, therefore making it complex. The patrons to my game who are not that good and win by learning the enemy's patterns will not be happy. They would be confused by all of the options that the AI has. This goes to negative reviews, and causes less game sales.
However, the simple AI found in games such as Mario, such as the goombas, makes it very easy for the average Joe to complete the levels. He can more easily understand how the AI works, how to avoid the enemy, and eventually kill it.