The next major change in video games may not be better graphics or larger maps. It could be a computer-controlled teammate that understands a spoken instruction, reacts to the current match and changes its plan without waiting for a scripted trigger.
NVIDIA is already presenting several practical examples through its ACE technology. PUBG is working with an AI-controlled co-player capable of communicating in natural language and acting with greater independence. MIR5 is using an adaptive boss that learns from previous player tactics, while inZOI has intelligent residents designed to plan and reflect on their actions.
These systems will need clear boundaries. That principle is already familiar in structured digital products such as an online casino, where players expect the rules of a game to remain consistent after every input. An AI companion can behave less predictably, but it still needs to respect the rules of the world it occupies.
PUBG’s co-player faces a harder test than natural conversation
The most interesting part of PUBG’s planned co-player is not its ability to speak. Games have contained voiced companions for decades. The difficult task is making the character useful during a match without allowing it to become either incompetent or unfairly effective.
A player may ask the companion to search a building, cover a doorway or move towards a marked position. Each instruction appears simple until conditions change.
An enemy could enter the building while the AI is searching. Gunfire may make the original route unsafe. The human player might move elsewhere without explaining the new plan.
The companion must decide when to follow the instruction exactly and when to abandon it. If it ignores too many commands, it becomes frustrating. If it obeys blindly, it may walk into obvious danger.
Natural speech adds another problem. Players rarely deliver careful, complete instructions during a firefight. They use short phrases such as “watch upstairs” or “push left”. A human teammate understands these through the layout, previous conversation and shared view of the action.
An AI character needs access to similar context. It must identify which staircase the player means, what counts as the left side and how long it should hold the position.
A useful companion should sometimes be wrong
Developers may be tempted to make AI teammates exceptionally accurate. That would demonstrate the technology, but it could weaken the game.
A companion that always identifies enemies, selects the safest route and lands every shot is no longer behaving like a teammate. It becomes an automated advantage.
Players also need to understand why the character failed. A missed shot is easy to read. A strange tactical decision is more difficult, particularly if the AI cannot explain it clearly.
Good mistakes should fit the situation. A character may fail to notice an opponent hidden behind cover or misjudge the risk of crossing an open area. It should not forget a command seconds after acknowledging it or become stuck because two instructions conflict.
Consistency will determine trust. Players can adapt to a companion with limited aim or awareness. They cannot plan around behaviour that changes without an obvious reason.
This creates an unusual design problem. The goal is not to make the AI as intelligent as possible. It is to give it an understandable level of competence.
Adaptive enemies could become exhausting
MIR5 is taking a different approach by using an AI-powered boss that can learn from previous player tactics.
The idea is appealing. Boss encounters often become predictable once players identify a successful strategy. An enemy capable of recognising repeated attacks could force them to change weapons, positioning or timing.
There is still a risk of removing the satisfaction that makes boss battles enjoyable. Players expect difficult opponents to have patterns they can study. Learning those patterns and executing the correct response is part of the achievement.
A boss that changes too freely could make preparation feel pointless. If every successful tactic is immediately countered, the encounter may resemble an argument with the game rather than a test of skill.
Developers will need to decide what the boss is allowed to learn. It could adjust its positioning or place greater pressure on a frequently used part of the arena. It should not suddenly gain abilities that contradict its established design.
Players also need a chance to recognise the adaptation. A new animation, sound or piece of dialogue could signal that the boss has noticed their strategy. Without that feedback, a clever system may look like hidden cheating.
AI residents need memory with limits
The Smart Zois in inZOI address another long-standing problem. Residents in life simulations often follow fixed routines and forget important encounters as soon as the scripted event ends.
An AI-driven resident could remember an argument, change its daily plans or approach another character because of an earlier conversation. That would make the town feel less like a collection of schedules.
Long-term memory creates practical complications. If every character remembers every interaction, the system may become expensive to run and difficult to control. It could also preserve moments that the player considers insignificant.
Developers may need to give characters selective memory. A neighbour could remember a major disagreement but gradually lose details of a casual conversation. Relationships would then develop without requiring a permanent record of every sentence.
The system must also stop characters inventing facts. An unscripted conversation can feel convincing until a resident refers to a location that does not exist or claims knowledge the story has not provided.
Game worlds depend on boundaries. A character may speak naturally, but its knowledge should remain connected to the setting, current events and information it could reasonably possess.
AI advisers may arrive before fully autonomous characters
The AI adviser being developed for Total War Pharaoh is a more controlled use of the technology. It is designed to help players understand systems and mechanics by drawing on the game’s own data.
Strategy games are well suited to this kind of assistance. They often contain detailed tutorials, tooltips and encyclopaedia entries, yet players still struggle to find the right explanation during a campaign.
An adviser could answer a direct question about food production, public order or army composition without forcing the player to search several menus. It could also tailor the response to the current campaign.
Accuracy is essential. A plausible but incorrect answer could damage hours of planning. The adviser therefore needs to retrieve information from verified game data rather than generate a response based only on general patterns.
It should also distinguish between a rule and a suggestion. Explaining how a mechanic works is different from recommending a strategy. A player needs to know when the answer is factual and when it is one possible approach.
This narrower role may prove more useful than an AI character designed mainly to produce endless conversation.
Microsoft’s Muse is aimed at development as well as play
Microsoft has been exploring another branch of game AI through Muse, a model trained on Ninja Theory’s Bleeding Edge.
Muse was built to understand a 3D game world, including its physics and the connection between controller inputs and actions. Microsoft has discussed its potential for prototyping gameplay ideas and helping older games run on newer devices.
The preservation angle is particularly significant. Many older titles are tied to discontinued hardware or technical systems that are expensive to rebuild.
An AI model that understands how a game responds could support preservation work, although it cannot replace the original code and assets without raising questions about authenticity. A reconstructed game may look correct while behaving differently in less obvious situations.
For developers, prototyping offers a clearer short-term benefit. Teams could test the shape of an idea before committing artists, designers and programmers to a full production build.
That should not be confused with creating a finished game automatically. A prototype can show that a mechanic is possible. It cannot decide if the mechanic fits the story, remains interesting after ten hours or works for players with different abilities.
Local processing could influence which ideas succeed
Several AI gaming tools are being designed to run on the player’s own device. Local processing can reduce delays because every instruction does not need to travel to a remote server.
It can also help games remain playable during connection problems. An AI companion that stops responding whenever the internet becomes unstable would be difficult to trust.
The trade-off is hardware demand. Speech recognition, language generation, animation and game logic all compete with graphics for processing power and memory.
Developers may need smaller models built for individual tasks. One model could handle speech, another could choose actions and a third could generate facial animation. This is less dramatic than a single all-knowing character, but it may be more practical.
Console and PC specifications will also affect design. A feature that runs comfortably on a high-end graphics card may need a reduced version on older hardware.
Players need to know what the AI is doing
AI in games will not succeed through novelty alone. Players need to understand what information a system uses, what it remembers and how much control it has.
A co-player should make its current objective visible. An adaptive boss should signal a tactical change. A life-simulation resident should not claim knowledge it never received.
These details are ordinary game design problems made harder by a system capable of producing unexpected behaviour.
The strongest AI characters may not be those that can say anything or respond to every situation. They will be the ones whose decisions still feel connected to the world, the rules and the player’s actions.
A companion that makes a believable mistake and recovers intelligently could be more convincing than one that performs perfectly. That is the test facing PUBG, MIR5, inZOI and the other games now moving AI out of demonstrations and into playable systems.
Login to Geektown