Games with good AI

NOTE: This page is still under construction, there are more games to be added.

Defining Artificial Intelligence

Artificial Intelligence (AI), a term coined by John McCarthy in 1956, refers to systems designed to perform tasks that typically require human intelligence, such as reasoning, problem-solving, and decision-making. AI includes a wide array of algorithms and approaches aimed at mimicking intelligent behavior. Over time, AI has evolved into two primary paradigms—symbolic AI and machine learning—each playing a distinct role in its development and application, especially in video games.

Symbolic AI

Symbolic AI, also known as "classical AI," was prominent from the 1950s to the 1980s. It models problems as structured "state spaces" using facts, rules, and actions, with systems searching these spaces to solve specific tasks. Techniques like finite state machines (FSMs), behavior trees, and goal-oriented action planning (GOAP) are key examples. Though largely overtaken by machine learning for complex, dynamic problems, symbolic AI remains vital in video games for structured tasks like pathfinding (e.g., using the A* algorithm) and rule-based NPC behavior.

Machine Learning

Machine Learning (ML), introduced by Arthur Samuel in 1959, enables systems to learn from data without explicit programming. ML algorithms improve iteratively by analyzing data and minimizing errors, excelling at complex, adaptive tasks. Unlike symbolic AI’s reliance on predefined rules, ML identifies patterns and refines solutions independently. The evolution of conversational AI, which ML later revolutionized, saw early attempts like ELIZA (1966). Developed by Joseph Weizenbaum, ELIZA was a rule-based program (an example of early Natural Language Processing within symbolic AI) that simulated conversation through pattern matching. Much later, Microsoft's Tay (2016) was an experimental chatbot that used machine learning to learn from user interactions.

• 1950s–1960s: The perceptron, a basic neural network, laid the groundwork but was limited in capability.
• 1980s: Backpropagation enabled multi-layer neural networks, paving the way for deep learning.
• 1990s–2000s: Support vector machines (SVMs) and decision trees advanced supervised learning.
• 2010s: Deep learning, powered by large datasets and GPUs, revolutionized AI with large-scale neural networks.

ML includes key subfields relevant to gaming:

• Deep Learning: Uses complex neural networks for tasks like Google DeepMind’s AlphaStar, which mastered StarCraft II.
• Large Language Models (LLMs): Advanced models like GPT-3 generate human-like text, enhancing dialogue and content creation.
• Reinforcement Learning (RL): Trains AI to optimize decisions, as seen in AlphaGo and gaming AIs.
• Generative AI: Creates original content (e.g., text, images) using models like ChatGPT, increasingly used in game development.

AI in Video Games

AI has been integral to video games since their inception, evolving from simple rule-based systems in games like Pac-Man to advanced ML-driven technologies. It enhances gameplay, generates content, and streamlines production. The field of Game AI combines symbolic AI for efficiency (e.g., navigation meshes) with ML for adaptability, prioritizing immersive player experiences while maintaining performance.

AI for Player Experience

AI in games serves three main functions:

AI that Plays

This governs NPCs, opponents, and gameplay systems:

• Pathfinding: A* and navigation meshes ensure efficient movement (e.g., in DOOM).
• Behavior Systems: FSMs (The Last of Us), behavior trees (Halo), and GOAP (F.E.A.R.) control NPC actions.
• ML-Driven Opponents: Examples include Gran Turismo’s SophyAI and Forza’s Drivatars, which mimic player styles.

AI that Creates

Procedural Content Generation (PCG) crafts levels, narratives, and assets using intelligent rules. Generative AI enhances this by producing novel content, such as NPC dialogue or art. Innovations include:

• Real-Time Rendering: NVIDIA’s DLSS upscales graphics using deep learning.
• AI-Generated Assets: Tools create dynamic content, as seen in RetroArch’s AI Service for text translation.

AI that Models

ML analyzes player behavior to improve experiences:

• Matchmaking: Systems like Xbox’s TrueSkill ensure balanced matches.
• Analytics: Data-driven insights refine game design.

AI for Game Production

AI also transforms development processes:

• Testing: AI automates QA, with tools like modl.ai finding bugs (e.g., Ubisoft’s The Division).
• Content Creation: Motion matching (Hitman) and texture upscaling (Mass Effect: Legendary Edition) enhance efficiency.
• Analytics & Fairness: ML informs design (Tomb Raider: Underworld) and combats cheating (Counter-Strike).

AI that Plays

Used in quality assurance (QA) and testing, AI automates repetitive tasks and employs curiosity-driven algorithms to identify bugs. Companies like Ubisoft (The Division) and Electronic Arts have integrated AI-driven testing, while startups like modl.ai make these tools accessible to indie developers.

AI that Creates

• Animation: Motion matching, as seen in Hitman and The Last of Us Part II, uses AI to generate realistic animations efficiently.
• Texture Upscaling: AI enhances low-resolution assets during development, as in Mass Effect: Legendary Edition.

Generative Tools: Tools like GitHub Copilot accelerate code development, while generative AI creates subtitles or text-to-speech for accessibility, reducing production time.

AI that Models

• Player Analytics: ML analyzes player data to inform design decisions, as seen in Tomb Raider: Underworld.
• Anti-Cheating Systems: Valve’s use of deep learning in Counter-Strike detects cheating behaviors, ensuring fair online play.

AI’s evolution from symbolic systems to machine learning and generative models has revolutionized both its theoretical foundations and practical applications. In video games, AI enhances creativity, interactivity, and immersion, from rule-based NPCs to generative content creation. Game AI, as a specialized field, balances technical efficiency with engaging player experiences, while production-focused AI streamlines development processes. As AI continues to advance, its synergy with gaming not only elevates player experiences but also drives broader AI innovation, with games serving as dynamic testing grounds for new algorithms.

Cheap tricks used in video games for AI

• Ignoring game rules and physics such as bypassing vehicle handling model or using rubber-banding in racing games

Example: In a racing game, you're in first place by a significant margin, but suddenly the AI cars behind you (especially those that were struggling) gain an impossible burst of speed on a straightaway, allowing them to catch up or even overtake you in a way that defies the physics of their vehicles or your own. Or, an AI car takes a corner at full speed without drifting or losing traction, something a player's car would never be able to do.

Player Perception: "How did that car just do that? It was miles behind!" or "The AI cars don't seem to be affected by the same physics as mine."

• Hitbox manipulation

Example: In a fighting game, an enemy's punch appears to miss your character on screen, but you still take damage because the hitbox for their attack is much larger than the visual representation. Conversely, your perfectly aimed shot at an enemy might pass right through them without registering a hit because their hitbox is smaller or oddly shaped.

Player Perception: "That was a phantom hit!" or "My attack clearly hit, but it didn't register."

• Static modifier based buff/debuff system instead of implementing complex buff system which has a series of classes for each modification.

Example: When an enemy casts a "Poison" spell, your character instantly gets a fixed "-2 HP per second" debuff icon, regardless of their armor, resistance, or the enemy's intelligence stat. There's no visible calculation or interaction between different game systems (like an elemental resistance system mitigating the poison's full effect); it's just a direct, predetermined numerical penalty. Similarly, an "Attack Boost" buff on an enemy simply gives them a flat "+20% damage" increase, rather than a more nuanced system where the boost interacts with their specific weapon type, skill level, or your character's defense.

Player Perception: "This buff/debuff always does the exact same thing, no matter what." or "There's no real depth to how these status effects work."

• Enemy level scaling/leveling manipulation

Example: In an open-world RPG, you return to an area where enemies were previously easy to defeat, but now that your character has leveled up, those same enemies have also magically leveled up to match your strength, negating the feeling of progression. Or, an area designated for early-game players suddenly has high-level enemies if you enter it with a powerful character.

Player Perception: "No matter how strong I get, the enemies always feel the same difficulty." or "Why are these goblins suddenly as tough as dragons?"

• Momentum/Dynamic Difficulty Adjustment (DDA) and player handicapping (reducing player capability)

Example: In a sports game, if you're dominating your opponent, your AI teammates suddenly start making more errors, your shots become less accurate, or the opposing AI gets a sudden burst of improved performance (faster passes, perfect tackles). Or, in an action game, if you're consistently performing well, the game might subtly reduce your damage output or increase enemy health to make fights last longer.

Player Perception: "The game feels like it's actively trying to stop me from winning." or "My players suddenly forgot how to play!"

• Hardcoded and heavily scripted solutions for NPC behaviours or scenes instead of using situation-driven dynamic complex tree systems.

Example: An NPC in a quest always follows the exact same path and delivers the same lines of dialogue in the same order, regardless of what the player does or says. If the player tries to interact with them out of sequence, the NPC either ignores them, repeats an earlier line, or breaks the script entirely. Or, a cinematic cutscene always plays out identically, even if the player's actions in the preceding gameplay should logically alter the outcome or character reactions.

Player Perception: "This NPC feels like a robot, just following a script." or "My choices don't seem to matter; the story always plays out the same way."

• Input reading/reading user input and then splicing in correct responses ahead of the players input

Example: In a fighting game, as soon as you initiate an attack animation, the AI immediately blocks or counter-attacks with perfect timing, even if your attack is meant to be a surprise or have a delayed wind-up. Or, in a stealth game, an enemy immediately turns and spots you the moment you press the button to peek around a corner, before your character's animation even completes.

Player Perception: "The AI is reacting to my button presses, not my character's actions on screen." or "It's like the AI knows what I'm going to do before I even do it."

• Enemy spam

Example: In a wave-based combat scenario, instead of increasing enemy intelligence or varying enemy types, the game simply throws an overwhelming number of identical, weak enemies at the player, relying on sheer quantity to provide a challenge. Or, in an RPG, every few steps you take in a dungeon trigger an identical random encounter with the same group of enemies.

Player Perception: "This isn't hard, it's just annoying. There are too many of them!" or "Are these the only enemies in the game?"

• Spawning vehicles on the visible map

Example: In open-world games, a simple implemented system dynamically spawns vehicles just outside the player's field of view to maintain the illusion of a busy, living city. This system uses algorithms to determine the optimal placement and timing of spawns based on the player's position and the camera's orientation, ensuring that cars appear 'naturally' as the player explores the environment. In older games, developers even leveraged camera viewpoint limitations, spawning objects only within the visible frame to optimize performance.

Player Perception: "The city always feels populated with traffic, but occasionally, cars seem to materialize out of thin air."

• Perfect aim accuracy of the enemy

Example: A standard grunt enemy with a pistol in a shooter game is consistently headshotting you from across the map, even while moving, or through smoke and darkness, something a player with the same weapon would struggle to do. Even if you're behind cover, bullets seem to find you with uncanny precision.

Player Perception: "This enemy's aim is impossible; they never miss!" or "I'm behind cover, how am I still getting hit so accurately?"

Some of these if properly and correctly implemented, would be effective in AI design for video games which can significantly enhance gameplay by incorporating dynamic elements that reflect real-world scenarios. For instance, consider a sports game featuring a heavily underdog home team known for its passionate crowd and dedicated supporters (ultras). If this team scores a goal due to a stroke of luck, the resulting surge in adrenaline, dopamine, pressure and morale could dramatically shift the game's momentum. A well-designed momentum system would allow the underdog team to temporarily outperform their much more skilled worldstar opponents, simulating the psychological and emotional boost from the crowd's energy (a simplistic version of which was seen in early Pro Evolution Soccer games[1]). This creates a more immersive and realistic experience, making characters feel responsive to in-game events.

List

Cover	Title	Notes SPOILER ALERT
	Alien Isolation	Tommy Thompson: The AI of Alien: Isolation
	Black and White	This game won AI awards because the 'creature' which acts as your pet could learn all sorts of skills and behaviours, which you could reinforce by rewarding or punishing him whenever he did stuff.
	Crysis and Crysis Warhead	The enemies treat you seriously. In a lot of FPS games, enemy npcs will unload 2 or 3 short bursts of gunfire (that usually miss) and hide back in cover, even though you’re right in front of them. In Crysis, the second you’re in line of sight you’ll have everyone mag dumping in your general direction (accurately). There’s usually a limit for how many enemies can attack you at once, but in Crysis it feels like anyone in the general vicinity will attack you all at once. It really sells the power fantasy of you being this high priority target that everyone is afraid of. It also punishes the player for not preparing for the encounter and ignoring the importance of energy consumption and proper positioning.[2]
	Dishonored 2	On higher difficulties.
	Divinity 2: Original Sin	Surprisingly good especially when the first big AI patch originally hit.
	Far Cry 2	Not the best on the combat but very tense, if they get the drop on you with no cover, it's probably over. The melee and random ai spawns were annoying, stealth was just ok. But there are immersive elements and behaviours.[3]
	F.E.A.R	See this paper or flevesque's and rocpapershotsun's articles for more information.
	Grand Theft Auto IV	Surprisingly good especially the cops and enemy AI compared to Grand Theft Auto V.[4]
	Halo: Combat Evolved	AI has a surprisingly robust decision tree that can respond to all sorts of things, like nearby cover, the type of weapon the player is shooting at them, proximity of other characters, and how hurt they are. On top of that, nearly every enemy encounter also has manual tweaks and scripting added to take advantage of things like the level's layout, special equipment lying around, and context-dependent things the player might be doing (such as going down a specific hallway).
	Half-Life 2	It has lots of different behaviour implementation to different situations.[5][6] But, unfortunately the guards have the mechanics are easily exploitable if you know what you're doing and some other little annoyances.[7]
	Haunting Ground	At the time (PS2 era) people complained because the chasers were pretty "smart". If you make too much noise, hide in the same places three times and other things, chasers would start to search for you in those places. Also, they could be tricked/discover your position because of the doors that you forgot to close. Hewie's AI was also great. Like a real dog. Hewie's AI gets better if you reward him for doing well or scold him for doing poorly. He gets pretty smart if you raise him well! And if you're mean to him it'll make him scared and mean back. You can even get him to attack you.
	Left 4 Dead and Left 4 Dead 2	The Director (sometimes referred to as The AI Director, or simply as AID) is the artificial intelligence of Left 4 Dead that features a dynamic system for game dramatics, pacing, and difficulty. (Similar to Alien: Isolation) Instead of set spawn points for enemies, the Director places enemies in varying positions and numbers based upon each player's current situation, status, skill, and location, creating a new experience for each play-through. The Director also creates mood and tension with emotional cues, such as visual effects, dynamic music, and character communication. Moreover, the Director is responsible for spawning additional health, ammo, weapons, and Special Infected, like the Witch or the Tank. It should be noted that there is another Director in the game, which controls the music on a per-player scale, called the Music Director. This article will only deal with the main one, the Director in charge of spawning enemies and items. [8][9]
	Metal Gear Solid V	If you happen to get into combat, the enemies push you tactically and give suppressing fire to each other. The enemy rarely stays in one place during combat. It's pretty easy to get out of immediately danger, but the enemies have your "last known position" which they immediately start to push and secure after they lose visual contact on you. If you happen to get out of combat phase, the enemies stay alert for a long time, and even signal other nearby outposts and checkpoints to stay on heightened alert. This can be prevented by destroying the comms in that outpost. On the opposite side is Ghost Recon Breakpoint. The AI pretty much has three phases. They patrol/search, stay in cover, or shoot at you from said cover. Do you tend to use a specific route often? Expect mines. Are you on alert state a lot? Expect riot shields, and in worst case, even suits that protect a lot of damage. Night vision or if you use lot of smokes and gasses, the enemy starts to use gas masks. When you mess up on a high-level FOB infiltration, then you're trying to hide or still move toward the objective, and you use binoculars and you see a mix of A) heavily armed soldiers SPRINTING sneakily from cover to cover in the dark B) heavily armed soldiers hunkering down in cover watching for you around corners in the dark C) them using radios to coordinate. The way they gasp when you sneak up on one, the sound/animation when you stun with flashlight beam at night, or the fact that if you interrupt a guard during a radio transmission, even if you do it silently, the commander obviously picks up on the fact that the soldier cut off and he raises the alert level. So you have to let any transmission finish before taking out the guard. But, unfortunately the guards have the mechanics are easily exploitable if you know what you're doing. For example, the AI has no counter to stun darting everyone in the legs in quick succession and then waiting. You can crawl ten metres in broad daylight in front of guards and they won't even see you.
	Middle Earth: Shadow of Mordor and Middle Earth: Shadow of War	Nemesis system
	Rainworld	The AI has so much range. There are predator plants that can basically one-hit kill you that are incredibly simple in the AI department, but that makes them terrifying as well: simple kill machines that just don't care otherwise. Then there are the scavengers. Social creatures, each with their own personality, that you can trade and better your relationship with each tribe, that hand signal you so you understand their current mood.
	Red Dead Redemption 2	unfortunately the enemy during combat have the mechanics are easily exploitable if you know what you're doing. But it has lots of different scripted behaviour implementation to different situations.[10]
	Selaco	Cites FEAR as an inspiration, and has the enemy AI as one of the explicit development goals.
	S.W.A.T. 4	Mitchell Godsend: Swat 4 - How Does the AI Work?
	The Forest	The behavior and actions of the cannibals change noticeably throughout the game, and they react based on your actions. But only if one of the cannibals escapes. If you kill all witnesses, they won't know about what happened. If you kill the leader of cannibal patrols, the rest tend to run away or kneel before you. You can even tame them. The cannibals can also apply non-violent methods through effigies, bizarre little constructions that serve to inform allies, warn the player, or other unknown purposes. There is a separate "faction" of cannibals, who are all emaciated and starving. They attack you on sight simply because they need to eat. If you incapacitate another cannibal, or leave out an animal corpse, they might start to try and eat them. If you cover yourself in red paint (and possibly if you light an effigy of your own on fire), then you possibly represent an unknown religious figure of theirs, and they will react very strangely; They might back away out of fear, while staying in sight. They might briefly kneel before you and bow their head. They might kneel before you and start a hum or chant. Sometimes a kneeling cannibal will remain kneeling regardless of what happens, presumably offering themselves as sacrifice, or surrendering themselves to you. If you slaughter everyone, the cannibals in general will leave you alone for a bit. The developers have stated that they added some "fuzziness" to the AI, to prevent us from entirely figuring out how it works.[11]
	The Last of Us	Intelligent AI from both humans and monsters alike. Makes every encounter intense and dangerous.
	Tom Clancy's Splinter Cell: Chaos Theory

External links

• YouTube series | AI 101 by AI and Games

Sources

• McCarthy, J. (1956). Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. Dartmouth College.
• Samuel, A. L. (1959). "Some Studies in Machine Learning Using the Game of Checkers." IBM Journal of Research and Development.
• Hinton, G. E., et al. (2006). "A Fast Learning Algorithm for Deep Belief Nets." Neural Computation.
• DeepMind. (2019). "AlphaStar: Mastering the Real-Time Strategy Game StarCraft II." deepmind.com.
• DeepMind. (2022). "AlphaCode: Competitive Programming with AI." deepmind.com.
• NVIDIA. (2018). "DLSS: Deep Learning Super Sampling." nvidia.com.
• Millington, I., & Funge, J. (2009). Artificial Intelligence for Games. CRC Press.
• Game Developers Conference (GDC). (2023–2024). Talks on AI-driven testing, motion matching, and generative AI in games.
• modl.ai. (2025). "AI-Driven Game Testing Solutions." modl.ai.
• Valve Corporation. (2018). "VACnet: Deep Learning for Anti-Cheating in Counter-Strike." GDC Talk.
• X Posts (2023–2025). Discussions on generative AI adoption, SophyAI, and Drivatars in gaming communities.
• Naughty Dog. (2020). "Motion Matching in The Last of Us Part II." GDC Talk.
• BioWare. (2021). "Mass Effect: Legendary Edition Texture Upscaling." bioware.com.
• How AI is Actually Used in the Video Games Industry | AI 101