Mikelis' Blog

Lessons from Playtesting Character AI

Playtesting AI early in production is essential. A large enough playtest provides a clear design vision and direction. It shows us how players will perceive the game once it’s released and out of our hands.

And to an uncomfortable degree, that’s all that matters. Designers influence perception of their work only in the team, while the game is in development. Once it ships, ratings, reviews, sales, and reputation belong to the players. What we say as designers doesn’t matter that much.

Playtests often bring bad news for the current design vision. Our ideas aren’t as strong as we believed, and sometimes players seem to misunderstand our creative intent. Nevertheless, the wise decision is to embrace feedback as a constructive force, helping the game succeed beyond our reach.

Misconceptions & Learnings

Several recurring assumptions about game AI collapse under player testing. To us, they may seem great – or at least worth trying – but players tend not to like them. Here’s what 5 years of AAA experience has taught me:

Realistic Logic Is Not Better

The most accurate, reality-approximating AIs exist in academia, and players don’t find them very fun. Nearly all game AI is deliberately unrealistic, biased toward fun, drama, spectacle, and theatrics. Trying to simulate the human brain is usually a waste of effort.

There are exceptions. In esports, realistic AI opponents have their place. Markov chains and machine learning models, for instance, drive AI moves in chess and strategy games.

But generally, players prefer AI characters that are more impressive, expressive, and larger than life, which rules out realism – or at least makes it a poor return on investment compared to alternatives.

Harder AI Is Not Better

Development teams playtest combat encounters endlessly. They become skilled at the game, and the game becomes boring. It seems that if they only made the encounters harder, they would have more fun.

But many players are far less skilled in the intricacies of the game. Dynamic difficulty often bridges this gap, adjusting encounters to player skill level. I cover this briefly in an article about player expectations.

Another reason game AI drifts toward difficulty is misreading other successful games. We often see players fight waves of grunts, special enemies, minibosses, and bosses. It seems like players love constant challenge. But this is misleading. Those games almost always shower players with resources, include strong AI hustling mechanics, and maintain pacing where players realistically face only one or two threats at a time.

Complexity is Undesirable

Complexity doesn’t guarantee better playtesting results. Scripted, theatrical AI characters often test better than advanced systems. A strong set of animation and bark-based reactions can easily outweigh dozens of combat states.

Focus on theatrics: barks, reactions, context, contrast, predictability, humanity. These are what stick with players. Complexity mostly matters in production costs – it’s expensive.

Players Can't Tell AI Difficulty

Players can’t usually tell how difficult AI opponents are. They judge difficulty based on how their encounters feel. This means designers can employ dynamic scaling, hustling, and other features to ensure players succeed while still feeling optimally challenged.

Cheating AI Can be Good

You may have seen this online – players write poor reviews about game AI because it cheats. But this is only the case when AI characters cheat against the player.

Even very evident cheating can be a good design choice. In Left 4 Dead 2, companions stuck in geometry teleport to the player. Resolving navigation issues across all maps could cost weeks of development, whereas teleporting is an appropriate solution. It also enables easier community map development.

But consider the opposite: a police character teleporting behind the player in Cyberpunk 2077 when wanted, which was universally disliked. It matters whether the cheating favors the player or not.

Randomness can be Good and Bad

Randomness also comes in two forms: input and output. Input randomness governs whether, when, or how often actions happen. Output randomness affects the final outcome.

Input randomness benefits AI. Characters use random delays, investigate random locations, or vary their hit chances, creating the illusion of intelligence and breaking up synchronization. Output randomness, however, undermines player control and emergent play. A bullet that hits should always do consistent damage. A scored football goal should always count. Changing outcomes feels like unfair cheating.

Varying when and how an action is performed saves development time and enhances believability. Varying the consequence of an action makes AI unpredictable, less relatable, and ultimately doesn’t facilitate emergent play. Players view it as cheating.

Conclusion

Playtests reveal more than we expect and often force direction changes, which is why testing early is vital. But some outcomes are predictable: realism, high difficulty, and excessive complexity rarely test well. There’s almost always a better return on investment in building AI with generally liked features. Cheating and randomness, when applied in the player’s favor, often test surprisingly well.

Want to learn more? Check out my other game AI development articles on this blog, where I talk about AAA-style game AI production and design. Or see these excellent community resources:

Special thanks to the following community contributors:
Giacomo Salvadori Co-editor

#Game AI #Game Design #Game Development #Game Production