Modeling Intentional Causality in Automated Storytelling

61 concepts39 papers1 proposal

If an AI generates a sequence of events—“John walked to the store. John bought a knife. John went home.”—the text is grammatically correct and chronologically ordered, but it lacks the essence of a story. Why did John buy the knife? Was he preparing a dinner party, or planning a betrayal?

Traditional language models and early AI systems often treat narratives as statistical word sequences or linear timelines. However, human readers perceive stories through the lens of intentional causality: the web of internal goals, plans, and psychological states that motivate observable behavior. When a narrative lacks this internal scaffolding, characters feel like cardboard cutouts, acting without clear purpose, and the plot becomes a disjointed sequence of random occurrences.

To build systems that can truly generate and understand stories, researchers have shifted from modeling what happens to modeling why it happens.

The Foundations of Personal Causation

The realization that narratives require a formal model of human psychology dates back to the early decades of artificial intelligence. Researchers recognized that standard physical cause-and-effect models were insufficient for narrative comprehension. To bridge this gap, Charles F. Schmidt et al. sought to formalize a computational model of "naive psychology" and personal causation, reverse-engineering how humans infer intentions and beliefs behind a person's choices.

This perspective directly inspired early generative systems. Instead of relying on rigid linguistic templates or story grammars, James R. Meehan introduced TALE-SPIN, a system that modeled storytelling as a simulation of autonomous agents navigating a world governed by physical and social rules. In TALE-SPIN, narrative coherence emerges organically as a chronological log of characters attempting to satisfy internal needs (like hunger or companionship) through rational problem-solving and social interaction.

[Internal Need / Goal] ──> [Rational Problem-Solving] ──> [Character Action] ──> [Emergent Plot Event]

As the field evolved, researchers realized that human behavior is rarely a straight line of simple problem-solving. Michael C. Dyer advanced this by introducing Memory Organization Packets (MOPs) in the BORIS system, moving away from rigid, linear event scripts toward flexible structures that account for overarching social contracts, goals, and emotional obligations. Understanding why a character refuses to pay for poor service, for instance, requires tracking these hidden social expectations. Similarly, Eric A. Domeshek demonstrated that for an automated system to provide meaningful social advice, it must index stories not by surface-level keywords, but by the underlying chains of goals and plans that led to specific outcomes.

Enforcing Motivational Coherence in Planning

In symbolic AI and narrative planning, the challenge shifts from simulating a world to mathematically enforcing that characters act only when they have a logical reason to do so.

As reviewed by Arnav Jhala et al., classical planning models focus primarily on physical possibility—whether an agent can perform an action based on world prerequisites. Narrative planners extend this by requiring motivational coherence. Under this framework, every character action must be justified by a sub-plan aimed at achieving that character's personal goals. If an action does not serve a character's internal motivations, the planning algorithm treats it as a flaw that must be resolved during the search process.

This abstraction allows systems to evaluate and manipulate stories based on narrative logic rather than low-level syntactic details. Adam Amos-Binks demonstrates the utility of this approach in interactive storytelling, where a player's choices frequently force the system off-script. By prioritizing high-level character goals and their causal dependencies over rigid action sequences, an experience manager can dynamically select new story paths that preserve the underlying psychological logic of the narrative, even during player-driven deviations.

Even when attempting to acquire these formal planning models directly from natural language text summaries, bridging the gap between logical prerequisites and narrative intent remains a primary focus for ensuring believable generation, as shown by Thomas Hayton et al..

Grounding Neural Models in Psychology and Commonsense

While symbolic planners excel at logical consistency, modern deep learning approaches attempt to capture the fluid, creative nature of human storytelling. However, standard neural language models frequently lose the plot because they treat characters as mere sequences of words. Even when models maintain dynamic vector memories for individual characters to track their involvement, Elizabeth Clark et al. note that they still fail to capture the underlying motivations behind their actions without explicit high-level story planning.

To resolve this, recent architectures explicitly ground neural generation in psychological frameworks and external commonsense knowledge. The PICS framework, developed by Yunpeng Li et al., and the COMMA framework, introduced by Yuqiang Xie et al., integrate psychological theories of motivation—such as Maslow’s Hierarchy of Needs—directly into the generation pipeline.

[Current Psychological State] ──> [Predict Next Motivation] ──> [Generate Action] ──> [Update Emotion]

By explicitly predicting a character's internal motivation before generating their subsequent actions, these models ensure that the text aligns with psychological transitions rather than just statistical word patterns. As Kyeongman Park et al. argue, grounding plot events in these specific psychological traits prevents the common failure mode where events simply happen to characters, ensuring instead that actions are driven by their unique internal logic.

Furthermore, integrating external causal and social commonsense knowledge provides a framework for tracking intentions and reactions over long contexts. Yuxin Wang et al. highlight that prioritizing "if-then" reasoning helps language models maintain character consistency and avoid logical contradictions. For instance, training models on external knowledge graphs that map intent and sentiment allows systems to better forecast script sequences and understand the downstream impacts of events, as demonstrated by Xiao Ding et al..

Crucially, a compelling story cannot progress entirely smoothly; it requires dramatic tension. To inject this programmatically, Youngrok Song et al. utilize defeasible inference—a form of reasoning that identifies how a goal might be thwarted. By structuring a formal mechanism to infer character intentions and intentionally introduce obstacles that weaken the causal link between a goal and its achievement, the system generates narratives with meaningful conflict rather than predictable, tensionless resolutions.

The Interplay of Character, Author, and Reader

Ultimately, modeling a narrative requires balancing three distinct perspectives: the internal goals of the characters, the rhetorical goals of the author, and the cognitive expectations of the reader.

This creates a fundamental tension in automated systems. If a system relies entirely on the local interactions of autonomous, reactive agents, complex plots can emerge, but they may lack dramatic structure (Vadim Bulitko). Conversely, Natlie Dehn argued that a system must prioritize authorial intent over pure world simulation. In Dehn's view, characters are "modeling clay" that the system should manipulate, introduce, or modify specifically to serve high-level narrative objectives like drama, suspense, and plausibility.

                  ┌────────────────────────┐
                  │    Authorial Intent    │
                  │ (Drama, Tension, Pace) │
                  └───────────┬────────────┘
                              ▼
                  ┌────────────────────────┐
                  │    Character Goals     │
                  │ (Motivations, Needs)   │
                  └───────────┬────────────┘
                              ▼
                  ┌────────────────────────┐
                  │     Emergent Plot      │
                  └────────────────────────┘

This deep intentional structure is also how human audiences comprehend and evaluate stories. Symbolic models of narrative planning show a high degree of alignment with human event cognition, accurately predicting how audiences interpret the causal relationships within a plot (Arnav Jhala et al.).

When systems lack this framework, their capacity for narrative comprehension drops significantly. For example, the PersonaBank corpus, developed by Stephanie M. Lukin et al., uses Story Intention Graphs (SIGs) to map out the emotional stakes and goals of personal narratives, demonstrating that reasoning about these underlying structures is essential for sophisticated tasks like stylistic retelling.

When evaluated on datasets like "TellMeWhy," standard language models struggle to answer basic "why" questions about character actions because they lack an explicit framework to represent the complex interplay of actions, plans, and goals (Yash Kumar Lal et al.). Similarly, extracting deep psychological profiles or personality traits from dialogue and scene descriptions remains a difficult benchmark for models that excel only at surface-level fact extraction (Yisi Sang et al.).

By shifting the focus of narrative AI from surface-level text generation to the underlying mechanics of intentional causality, researchers are closing the gap between machines that merely arrange words and machines that truly understand how to tell a story.

Key Foundational Work

TALE-SPIN, An Interactive Program that Writes Stories (1977) — contributed 3 concept elements to this direction.
TALE-SPIN, An Interactive Program that Writes Storie6 (1977) — contributed 3 concept elements to this direction.
TALE-SPIN, An Interactive Program that Writes Stories (1977) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_goal_driven_story_generation_b0701dd3,
  title = {goal-driven story generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/goal-driven-story-generation/rdr_b0701dd3a8d54d2886b265bc3fb04ec1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_goal_driven_story_generation_b0701dd3_v1,
  title = {goal-driven story generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/goal-driven-story-generation/rdr_b0701dd3a8d54d2886b265bc3fb04ec1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake