Is This AI Trolling Me?

The video below is fairly short but it will help illustrate my point.

Warning: For my IS 425 class there’s some unprofessional PG-13 language in this video.

It is a montage of a few encounters the YouTuber Marbozir has with the Siam AI in Civilization 5. There’s a general perception among Civilization players that when the computer plays as Siam that the computer will “forward settle” the human player (in the video above, that’s Marbozir) and then the computer will complain that the human forward settled Siam. If an AI could ever gaslight a human this is the closest example I can think of.

Independent of how the AIs (computer players) were programmed in Civilization 5, this presents some questions: Can an AI troll a player? Was the AI programmed this way or is this just an accident (like Nuclear Gandhi)? Is there a distinction between a human perception of an AI and the actual design of the AI?

What is Artificial Intelligence?

There isn’t a consensus on what an artificial intelligence is. Here’s the definition I suggest (adapted from my IS 425 course):

An artificial intelligence refers to an attempt by computers to mimic human behavior. Various statistical techniques, algorithms, and machine learning techniques can produce an AI, but these are the how. AI refers to the WHAT.

Using this definition, terms like classification and association are machine learning techniques/algorithms but the computer player (or the enemy) in a video game is an artificial intelligence.

The AI Was Born This Way

Until recently, video game AIs were manually coded. That is, developers would manually program the computer player to perform specific actions if certain parameters were met in the game. The goal was to create a system that was good enough to create the illusion (for the player of the game) that the computer players or the enemies weren’t acting “dumb.” The developers would observe humans playing the game and then program the computer to mimic that behavior.

What Programs Do

Imagine any “ordinary” computer program. When a developer writes an application they are attempting to automate or facilitate organizational or personal processes that were already being done by humans. The purpose of the program is to save time or money. The program’s structure and characteristics reflect the people using the program and their needs. The end result is a program with logical statements (if/then/else, loops, etc.) and parameters that influence those logical statements.

The downside to manually programed video game AIs is that once human players have figured out the rules that the computer uses (the logical statements) the game becomes predictable. In a strategy game like Civilization 5 (or Civilization 6) this makes the game easy and less of a challenge for expert players like Marbozir. Games usually compensate for a limited computer AI by giving computer players artificial advantages that the human doesn’t get. So the hardest difficulty level in Civilization isn’t hard because the computer acts more intelligently but because the computer gets bonuses that the human player doesn’t get.

ML-Driven Artificial Intelligence

These days, machine learning algorithms create those logical statements automatically and faster. If we were going to use machine learning techniques to create an AI Civilization player, we could feed the logs of thousands of human civilization games into an algorithm (probably a neural network) to create a computer model. The more games you feed into the model the better the AI should be as a player. This would also allow the AI player to improve and adapt to the human players it is facing. That would create a better challenge for players like Marbozir and would most likely provide a more enjoyable challenge for him.

What the Human Perceives

Whether the AI was built with manually-coded logical statements or was created by the use of machine learning algorithms, what is important is how the human user of the system (in this case, a video game) perceives the AI. Did the developers of Civilization 5 program the Siam AI to troll the player? This could have been as simple as programming the Siam AI to forward settle cities near the human player’s capital (which is what happens at the beginning of Marbozir’s video). There probably wasn’t a line of code that says “be a troll” but in the end Marbozir’s reaction to the computer forward settling him was “what a troll move.” It allows the player to assign human characteristics like “troll” to the computer player.

A program that passes the Turing Test is one where a human interacting with the program cannot distinguish between an actual human and that computer program. We’ve learned that an artificial intelligence doesn’t need to exactly mimic human behavior; it only needs to be “good enough” to produce characteristics that feel “human.”

So an AI is what the human perceives.

Bonus: Simulating Awful AI Players

To the credit of the Civilization 5 developers, the AI is decent. For a casual player like myself, the AI will beat me on normal difficulty. If I was willing to take the time to figure out all of the rules the computer uses then I could probably beat the game on higher difficulties.

Another characteristic of the Civilization 5 AI is that it had to reflect the artificial rules of the game itself. How about a sports game that has to reflect the rules of a real-world game like NHL hockey? Take a look at the video below from 2BC Productions (this is the only one I could find where he doesn’t excessively curse at the game):

Fans of the NHL know the rules of hockey and they know how the computer players should act in an NHL-based video game. They know this because they’ve watched hundreds (or thousands) of real-life games. This means that it is easy to identify when the computer player is not playing “naturally.”

When NHL 15 came out it was easy to identify these mistakes as shown in the video above. It was inconceivable that the developers of the game would not have tested the game to identify these issue (like Sidney Crosby making a horrible blind pass that gets intercepted). Creating an AI for a game like NHL 15 should have involved observing how hockey players play the game, consulting with players and coaches on formations and strategies and developing the game to reflect those and to make sure that the AI players don’t make obvious mistakes.

It is a situation where the AI doesn’t reflect human behavior and it becomes a distraction for the human players. Even if the AI in NHL 15 was manually programmed it could still have been decent if mistakes like the ones mentioned in the video were fixed.