10 · Designing Emotionally Engaging Narratives with AI | Generative AI for Game Development
This chapter companion argues that emotion in games is a system rather than a scene. It covers five design techniques from well-known games, then builds an emotional loop with VADER sentiment scoring, trust and mood variables, and GPT dialogue.
This video is a companion to the tenth chapter of Generative AI for Game Development by Shesh Narayan Gupta. When developers say emotional, the video says, they usually mean a sad cutscene. The chapter argues instead that emotion is a system that listens and adapts. The thesis is that in film emotion is directed at you, while in games it is chosen by you, and a choice you made hits differently than a scene you watched.
Three dynamics behind emotion
The video names three dynamics that make this work: player agency, where decisions must change relationships; emotional pacing, where tension and calm alternate; and relational depth, where characters evolve with or against the player. It cites Journey as proof. With no dialogue and no script, only exploration, companionship and subtle cues, players still bond with anonymous strangers because the emotion came from design rather than exposition.
Five techniques from existing games
- Motivation. Characters need clear, human goals such as redemption, belonging or revenge. Ellie's obsession in The Last of Us Part II gives every player decision a moral edge because you understand her even when you disagree.
- Contrast and pacing. Joy feels brighter after fear, and grief lands harder after peace. Arthur Morgan's quiet hunting hours make the later beats hit harder.
- Reflection. Give players space to process, such as a quiet walk, a journal entry, or a radio call with Delilah in Firewatch.
- Mechanics as emotion. In Papers, Please, stamping documents becomes moral anxiety, and in Celeste, climbing becomes the struggle itself. When the controls feel the feeling, the player performs emotion rather than watching it.
- Conflict. Internal conflict, as in Disco Elysium, relational conflict such as betrayal in Mass Effect 3, and moral conflict such as impossible choices in Detroit. The video adds that resolution is not always happy, noting that Life Is Strange ends in reckoning rather than relief.
From conditionals to sentiment-aware AI
Before AI, there is the conditional. In the farewell exercise, a companion leaves and the scene has two dialogue variants, grateful if you saved their family and wounded if you did not. One variable produces a different emotion from the same scene.
The AI version adds layers. The VADER sentiment tool reads the player's typed tone and scores it from minus one to plus one. If the score is negative enough, the character's mood shifts to hurt and trust drops by 15 points. GPT then speaks, but not blind: the prompt injects the player's words, the current mood and the trust score, so the model performs the tracked state instead of inventing the emotion. The full loop is player input, sentiment analysis, emotional state update, then GPT dialogue and memory, so the character remembers how you spoke three scenes ago.
Captain Io and how to know it worked
The example character, Captain Io, tracks two numbers: trust from zero to 100, and a mood of encouraged, hurt or uncertain. After five interactions comes the ending. Trust above 70 gives one last mission and "I'm with you," below 30 she walks away, and in between the answer is maybe someday. The video notes that the ending is arithmetic but feels like a relationship. Signs it worked include hesitation before a cruel choice, players describing the character like a real person, and regret in post-play interviews.
Key takeaways
- Emotion in games becomes agency: a chosen feeling hits harder than a watched one.
- Anchor characters in clear human goals.
- Pace with contrast and leave room for reflection.
- Fuse emotion into mechanics.
- Never use random emotion; tie it to player action and track variables such as trust and guilt.
- Let players influence emotion, never dictate it.
- AI is a collaborator: the designer and player author the feeling.
Try it yourself
Do the chapter's own exercise in Twine and build the farewell. A companion is leaving, and you write two variants from one variable: gratitude if the player saved their family, and hurt if they did not.
Full transcript(auto-generated, with timestamps)
[0:00]Generative AI for game development, designing emotionally engaging narratives with AI. Make players feel, not just play. Hello, this is Liam in for Bear. When developers say emotional, they usually mean a sad cutscene. This chapter argues emotion isn't a scene, it's a system that listens and adapts. Six terms, then we're off. Agency, decisions change relationships. Emotional pacing, tension and common rhythm. Relational depth, characters evolve with you. Vader, sentiment analysis for player tone. Trust score, a number that remembers. Emotional loop, input, sentiment, state, dialogue. In film, emotion is directed at you. In games, it's chosen by you, and a choice you made hits differently than a scene
[0:41]You watched. That's the whole thesis. Emotion becomes agency. Three dynamics make it work. Player agency, decisions must change relationships. Emotional pacing, alternate tension and calm. Relational depth, characters who evolve with or against the player. Journey proves it. No dialogue, no script, just exploration, companionship, and subtle cues. Players bond with anonymous strangers. The emotion came from design, not exposition. Technique one, motivation. Characters need clear, human goals, redemption, belonging, revenge. Ellie's obsession in The Last of Us Part Roman II gives every player decision a moral edge because you understand her even when you disagree. Technique two, contrast and pacing. Joy feels brighter after fear. Grief lands harder after peace. Arthur Morgan's
[1:28]Quiet hunting hours make the beat trails hit harder. Rhythm is emotional texture. Technique three, reflection. Give players space to process. A quiet walk, a journal entry, a radio call with Delilah in Firewatch. Pauses don't interrupt the story, they deepen it. Technique four, mechanics as emotion. In Papers, Please, stamping documents becomes moral anxiety. In Celeste, climbing becomes the struggle itself. When the controls feel the feeling, the player doesn't watch emotion, they perform it. Technique five, conflict internal torn between duty and desire like Disco Elysium. Relational allies who betray trust like Mass Effect 3. Moral impossible choices like Detroit. And resolution isn't always happy. Life is Strange ends in reckoning, not
[2:12]Relief. Before AI, there was the conditional. The farewell exercise, one companion leaving, two dialogue variants, grateful if you saved their family, wounded if you didn't. Same scene, different emotion from a single variable. Then Vader reads the tone. The player types, "I did what I had to." Vader scores it from minus one to plus one. Negative enough and the character's mood shifts to hurt, trust drops 15 points. Then GPT speaks, but not blind. The prompt injects the emotional context, the player's words, the current mood, the trust score. The model doesn't invent the emotion, it performs the state you tracked. Best practices, never random emotion, tie it to player action. Track variables
[2:52]Like trust and guilt that evolve. Design arcs with room for growth or regression. And let players influence emotion, never dictate it. The full loop, player input, sentiment analysis, emotional state update, GPT dialogue and memory. So, the character remembers how you spoke to them three scenes ago. Meet Captain Io. She tracks two numbers, trust from zero to 100 and mood encouraged, hurt, uncertain. Every message you send moves them. She doesn't just hear your words, she remembers your tone. Five interactions, then the ending. Trust above 70, one last mission, "I'm with you." Below 30, she walks away. In between, maybe someday, but not today. The ending is arithmetic, but it
[3:31]Feels like a relationship. How do you know it worked? Watch for hesitation before a cruel choice. Listen for players describing the character like a real person. And in post-play interviews, look for regret. "I felt bad betraying him." That's the win. Emotion is not a scripted event, it's a system, choices, consequences, and interactivity looping through memory. With AI, that system can now scale without losing depth. But AI is not the author, it's a collaborator. The designer sets the emotional foundation. The player brings it to life. The AI simply listens and adapts. The system in six lines. One, emotion is agency, chosen feeling hits harder. Two, anchor characters in clear human goals.
[4:12]Three, pace with contrast and leave room to reflect. Four, fuse emotion into mechanics. Five, track trust and mood. Let Vader read tone and GPT perform the state. Six, AI collaborates. The designer and player author the feeling. Your turn, the chapter's own exercise. Entwine, build the farewell. A companion is leaving. Two variants from one variable. Gratitude if the player saved their family, hurt if they didn't. Designing emotionally engaging narratives with AI at Nick Bear Brown.
More from Generative AI for Game Development — Shesh Narayan Gupta
4:3311 · Emotionally Adaptive Storytelling | Generative AI for Game Development
4:0609 · Creating Branching Story Frameworks | Generative AI for Game Development
6:0112 · Dynamic Character Creation | Generative AI for Game Development
5:0108 · Ethical and Practical Considerations | Generative AI for Game Development
3:5313 · Integrating AI Models into Game Development | Generative AI for Game Development
3:31