Torn: AI-powered Player Retention

Background & Challenge
Torn is a huge, global, browser-based game (MMORPG) with a unique 20-year legacy. The game immerses players in an evolving virtual world that partially mirrors real life but with significant differences. DVDs exist, CDs do not; you can fly to Argentina, but not Brazil.
Over its lifespan, Torn has grown organically into an intricate, expansive game environment, making it richly engaging yet dauntingly complex for new players. Despite being a 20 year old Game, as of June 2025, Torn has just hit over 75,000 daily active players and is continuing to grow.
With increased popularity comes heightened demands: new players frequently encounter a steep learning curve, exacerbated by fragmented, outdated, or difficult-to-locate information sources. Information is scattered across decades-old forum posts, player-generated wikis, and unofficial sources, frequently outdated or incomplete.
- Player churn: Many new players quit early due to frustration and a general confusion of how to play.
- Information bottlenecks: Players relied heavily on community veterans and incomplete external resources for basic guidance.
- Limited support: The forums and wikis provide a wealth of information for new players, but it’s not easily found.
To tackle increased player churn and streamline onboarding, Torn required a strategic solution that could understand, contextualise, and clearly communicate the nuances of its intricate environment.
Foundations
Before we were able to start building out any solution, we needed to decide on several foundational technologies and methodologies. Given the advances of AI and LLMs, this seemed like an obvious fit. Obviously many AI products exist already (such as LLMs), and some needed to be custom built.
Like ChatGPT, Gemini, or Grok, we needed to use an LLM as the base. Large Language Models (LLMs) predict text by learning semantic relationships, enabling them to appear to reason and think, providing a natural conversational interface. It made little sense to build our own, instead focussing on the ability to have an existing LLM as the backend, but with the ability to swap the base model as and when new models arrived. New models will be cheaper, faster, and more accurate, and some will struggle with the context.
As Torn shares a large part of its world with the real world, it was only correct for an LLM to assume that you could buy CD’s wherever you can buy DVDs – but as we touched upon earlier, CDs do not exist in Torn. For this, we needed to ensure the AI was aware of Torn’s reality.
By using Retrieval-Augmented Generation (RAG) techniques, we were able to integrate up-to-date and context-specific Torn data (such as a lack of CD availability) into the AIs responses. We did this to improve the accuracy of the responses, and reduce the number of hallucinations.
Finally, Torn’s brand and style is very unique and important to maintain. Each item has its own Torn-specific branding; the game has its own design style; the language used is witty, dry, and at times macabre. For obvious reasons, we needed the chatbot to remain in context, on brand, and understand the differences between crime in the game, and in real life. For this we built a custom Sentinel Guardrail System. The sentinel categorises user intent, prevents exploitation (jailbreaking), and ensures correct and contextual accuracy and security.
Approach & Solution
Led by Pete Sergeant – ddx’s Director of AI – we built on these foundations and developed an AI solution specifically suited to Torn’s complex gameplay and unique user requirements.
It was crucial for players to be able to engage naturally and fluidly with the chatbot – using slang and in-game lore. Unlike conventional chatbots, players could interrupt or refine their queries mid-response without causing disruptions.
To further ensure accuracy, the chatbot implemented a sophisticated "Chain-of-Thought" validation process, wherein generated responses were cross-checked and validated against reliable, in-game sources before being presented to the player. Doing this allowed us to reduce inaccuracies and prevent the common pitfalls of AI-generated replies.
As the goal was ultimately player retention, we needed to work out how to ensure players stayed playing. Sending a message to a player is one thing, but if we could work out if a player needed a first aid kit, particular items, or even some in-game money, we’d have far more success keeping them.
Enter: "Scenarios Engine," a robust, state-machine-based system enabling the chatbot to securely perform actions that directly impacted the game world—such as sending these items or engaging in player negotiations—without risk of exploitation.
As the AI solution grew, so did the need for a comprehensive evaluation suite comprising thousands of automated tests. This testing ensured the chatbot maintained accuracy and responsiveness; allowing us to focus on retaining players even when swapping AI models.
Collectively, what was delivered was a one of a kind, advanced, context-aware AI assistant tailored specifically to meet Torn’s distinctive needs, significantly improving player support, engagement, and retention. Using our suite of technology, we were able to deploy several NPC’s, each with their own personality and understanding.
Impact & Results
We saw an immediate impact on player experience and business outcomes. What was interesting was the significant player retention uplift of 7% among players who interacted with the chatbot.
This improvement directly correlated with reduced player churn, reflecting the new proactive onboarding experience for newcomers.
Beyond quantitative measures, player feedback indicated high levels of satisfaction and engagement. Players frequently expressed surprise at the chatbot’s responsiveness and accuracy, often spending extended periods interacting with it:
“Dude, what AI did they write you in? You are a beauty. You are my favorite part of Torn. Holy shit!”
“I didn't expect the chatbot to be this good. Great work, Torn.”
“Hey George, just got to level 5. Quite excited to learn more and more about the city. You've been a golden chap in getting me there. You're very well programmed. Cudos.”
Operationally, our solution offered Torn a lot of flexibility and efficiency. The evaluation suite allows us to quickly integrate new AI models with almost zero disruption, allowing us to upgrade the chatbot's performance and accuracy without the need for extensive retraining efforts.
Ultimately, this project is a great example of when to invest in a custom AI solution vs taking something off the shelf. This core AI platform is now part of Torn’s long term strategic roadmap, and together, we’ve already identified the next round of features that we’ll be able to build on top.
Lessons Learned
Throughout the project, the team learned many lessons that will help when building any other tailored AI solutions within complex environments.
One critical lesson was recognising the limitations of off-the-shelf AI solutions in highly specialised contexts. The unique requirements and nuances of Torn demonstrated that customised development, paired with testing and continuous refinement, is necessary for success.
Another key insight was the efficacy of practical "dirty tricks", small, unconventional solutions that significantly enhanced the chatbot’s performance in practice, often outperforming more theoretically elegant methods.
Building the Torn chatbot has highlighted the importance of adaptability, fit & validation, and a willingness to iterate continuously to achieve practical, reliable outcomes in AI development.
Applicable Industries
The innovative solutions and methodologies developed through the Torn chatbot project have clear applicability beyond gaming, particularly in industries requiring high accuracy, robust security, and interactive user experiences. For instance, the healthcare industry could greatly benefit from secure, accurate AI-assisted record creation and patient interaction, minimising risks and ensuring compliance.
In the finance sector, reduced hallucination and enhanced accuracy in AI-driven advisory and compliance roles would significantly improve reliability and trust. Similarly, customer service across various industries could leverage these advancements to deliver personalized, accurate, and efficient support.
Ultimately, any regulated environment or industry requiring secure, auditable AI actions and interactions could find substantial value in adopting these tailored approaches developed through the Torn chatbot project.