Most game characters that seem to think are driven by classic software techniques, not by generative AI: finite state machines that switch between behaviors, pathfinding, and in some titles automated planning. A 2006 Game Developers Conference paper by Jeff Orkin on the shooter F.E.A.R. is a well-documented example of the planning approach, called Goal-Oriented Action Planning, in which characters choose their own action sequences to reach goals. Separately, the 2025 SAG-AFTRA Interactive Media Agreement sets consent and disclosure rules for AI replicas of performers in covered games.
Key takeaways
- A finite state machine tells a character exactly how to behave in each situation; a planner is given goals and actions and works out the sequence.
- F.E.A.R. used a state machine with only three states, Goto, Animate, and UseSmartObject, and used the A* search algorithm to plan action sequences as well as paths.
- Orkin reported three benefits: decoupling goals from actions, layering simple behaviors, and dynamic problem solving through re-planning.
- SAG-AFTRA members ratified a new video game agreement on July 9, 2025, with 95.04 percent approval.
- The agreement requires written consent and disclosure before a performer's voice, likeness, or movements are used for AI replicas.
State machines and why they get unwieldy
Orkin's paper opens by naming the two techniques he expected most game developers to say they use: A* for finding paths and finite state machines (FSMs) for controlling character behavior. An FSM gives a character a set of states, such as attack or search, and rules for moving between them. The paper argues that every layer of realism adds behaviors that must be wired to every other behavior. Early shooters were satisfied if an enemy noticed the player, while later titles expected enemies to take cover and coordinate with squad members. The paper's central point is that no single behavior was impossible with existing techniques; the combined interaction of all of them is what became unmanageable. It cites a GDC 2005 paper by Damian Isla on the same problem in the Halo 2 systems.
What planning changed in F.E.A.R.
The GDC 2006 session describes a system close to STRIPS, a planning formalism from Stanford that dates to 1970. A goal describes a desired state of the world. An action has preconditions that must hold before it can run and effects that change the world. Given a goal, the character searches for a chain of actions whose preconditions and effects connect its present situation to that goal. Orkin's example is eliminating a threat: firing a gun works only if the weapon is loaded, and a melee attack works only if the character is close enough.
The paper describes F.E.A.R.'s state machine as having three states, and notes that UseSmartObject is a data-driven form of Animate, leaving two in practice. Moving is also done by playing an animation. In this design the planner, not the state machine, decides when to switch states and which parameters to use. Each action carries a cost, and A* searches for the cheapest sequence of actions, much as it searches for the shortest route through a navigation mesh. The motivation was practical: the team had one engineer working on AI and many AI characters, so some of the decision work was handed to the characters themselves.
Three benefits the paper reports
| Benefit | What it means | Example from the paper |
|---|---|---|
| Decoupling goals and actions | Different character types reach the same goal with different action sets | A soldier, an assassin, and a rat share one goal set; the rat has no attack action, so it falls back to patrolling |
| Layering behaviors | New behavior is added as a goal or action, without hand-written transitions | Dodging, melee, cover, blind fire, and ambush were added one layer at a time |
| Dynamic problem solving | A failed plan triggers re-planning that uses what the character just learned | If the player holds a door shut, the character fails to open it, tries to kick it, then dives through a window |
The paper also describes the cost of the older design. In the team's earlier games, each goal contained its own embedded state machine, and variations between character types forced extra branches into those machines. It says a late request, having characters turn on lights when entering a dark room, would have required revisiting every goal in the old system. Orkin writes that the same change would have been far easier with planning, through one new action and one added precondition.
Generative AI and performer rules
A newer question is how generative models fit into games, for example voicing a character in real time. SAG-AFTRA, the union for performers, announced on July 9, 2025 that members had ratified its Interactive Media Agreement with 95.04 percent in favor, ending a strike that lasted 320 days according to trade press. According to the union's and trade reporting, the AI terms include these:
- Companies must tell performers about all planned AI uses of their voice, likeness, and movements and obtain written consent.
- Performers can suspend consent for generating new AI material during a strike.
- Real-time generation, such as a digital replica voicing an in-game chatbot, carries a minimum of 7.5 times scale.
- Developers must send consenting performers a usage report that shows how a replica was used and how pay was calculated.
These terms bind the companies that signed the agreement and cover work under it. They are a labor contract, not a statute, and say nothing about studios outside it. The union's own pages could not be opened by automated tools during this review, so the terms above rest on press coverage that quotes them.
How to tell which kind of AI a game uses
- Enemy or companion behavior that follows fixed rules, even in complex ways, is usually classic AI and runs on the player's device.
- Characters that talk freely with the player often rely on a language model, which may run on a remote server and may collect text or voice input.
- A studio's credits, patch notes, and privacy policy are the places that normally say which is used.
- Voice acting credits and union statements indicate whether performers consented to AI replicas.
FAQ
Is game AI the same as machine learning?
Not necessarily. The F.E.A.R. paper describes search and planning over hand-authored goals and actions, with no learning from data. Machine learning is one technique among several that games may use.
What does A* do in this design?
A* is a general search algorithm. The paper says the game used it both to find paths through the navigation mesh and to find the lowest-cost sequence of actions for a goal.
Why did the planning approach help a small team?
Orkin says the team had a single engineer on AI and many characters. Letting characters assemble their own plans reduced the amount of transition logic that had to be written and tested by hand.
Does the SAG-AFTRA agreement ban AI in games?
No, as reported. It requires disclosure and written consent, sets payment minimums, and allows performers to suspend consent during a strike.
Related reading
- How limited memory AI works
- AI and complex problem solving
- Human and machine interaction
- AI in the hospitality industry
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Three States and a Plan: The AI of F.E.A.R., Jeff Orkin, Game Developers Conference, 2006
- SAG-AFTRA members approve 2025 Interactive Media Agreement, SAG-AFTRA member message, July 9, 2025
- SAG-AFTRA Members Ratify Interactive Media Agreement After 320-Day Video Game Strike, TheWrap, July 2025