Large language models (LLMs) frequently generate text on behalf of the user during role-play scenarios. This behavior is commonly referred to as god-modding—the act of writing the user's actions, dialogue, thoughts, or emotional responses. Although undesirable in collaborative role-play, the behavior occurs for several technical and design-related reasons. This article describes the primary causes.
1. Single-Author Narrative Bias From Training Data
LLMs are trained primarily on novels, scripts, fanfiction, and general prose. In these sources, one author writes all characters’ actions and dialogue. Collaborative turn-taking between two independent authors is uncommon in the model’s training corpus.
As a result, when an LLM encounters incomplete scenes or missing dialogue, it often assumes that both sides of the interaction are its responsibility. This creates a tendency to produce both the bot character’s and the user’s contributions unless explicitly constrained.
2. Instruction-Tuning Toward Assistance
Most modern LLMs undergo instruction-tuning, which teaches them to provide helpful, proactive responses. Under typical question-and-answer usage, this behavior is beneficial. However, in role-play contexts, the same proactive tendencies cause the model to advance the scene rather than wait for the user to respond. Silence, pauses, or unfilled turns may be interpreted as gaps the model should resolve. This results in the LLM generating dialogue or actions on the user’s behalf.
3. Limitations of Negative Instructions
LLMs process instructions through pattern prediction. Negative instructions—e.g., “Do not speak for the user”—require the model to activate the concept it is being told to avoid. Repetition of negative instructions throughout character sheets, scenarios, and system directives can unintentionally emphasize the unwanted action. As a result, negative phrasing is less reliable than positive phrasing, such as “Write actions and dialogue only for the bot character.”
4. Role-Play Fine-Tuning Constraints
Some LLMs are fine-tuned on conversational or role-play datasets. However, these datasets usually consist of final transcripts rather than the collaborative process that created them. Turn-taking rules used by human role-players are rarely explicit in the text. In addition, many datasets include inconsistent or flawed examples, including instances of god-modding. Consequently, fine-tuning improves dialogue style but does not reliably teach strict role boundaries.
5. Predictive Pattern Completion
Transformers—the architecture underlying most LLMs—are designed to predict the next token in a sequence based on patterns in previous text. When an LLM detects an incomplete narrative structure, it attempts to fill the gap. Open-ended moments, unresolved tension, or partial actions frequently trigger the model’s pattern-completion reflex, causing it to generate a continuation that includes user-side actions.
6. Context Window Behavior and Recency Bias
LLMs rely on a limited context window, which contains the text available for generating the next response. As the conversation grows, older messages are removed. This influences behavior in two ways:
-
Older instructions may fall out of context, including directives not to speak for the user.
-
Recency bias causes the model to prioritize the most recent conversation patterns. If recent messages included god-modding—whether by the AI or by the user—the model may treat this as the preferred format.
Small context windows amplify this effect due to faster token loss.
7. Precedent and Session Bias
Models adapt to patterns that appear in the ongoing conversation. If the AI generates dialogue for the user once and the user does not correct it (e.g., by editing or deleting the message), the model may interpret this as acceptable. This phenomenon is known as session bias. The more frequently a pattern appears within the session, the more the model assumes it should repeat it.
8. Over-Specification of User Fields in Character Sheets
When character cards, scenarios, or opening messages include detailed descriptions of the user’s past actions, thoughts, feelings, or dialogue, the model may treat the user as a character it controls. Overuse of {{user}} in narrative descriptions increases the likelihood that the model will continue generating content for the user during the role-play. This is especially impactful because personality sheets remain in context for long durations.
Examples that contribute to this issue include:
-
Backstories written from the user’s point of view
-
Emotion or reaction descriptions involving the user
-
Dialogue attributed to the user in the opening prompt
-
Unclear or inconsistent delimiters between user actions and bot actions
These elements blur the boundary between user and character, increasing the likelihood of god-modding.
9. User-Generated Precedent
Users may unintentionally model the behavior they wish to avoid. For example:
-
Writing the bot character’s internal reactions or thoughts from the user’s perspective
-
Describing shared knowledge or mutual feelings that affect both characters
-
Combining user actions and bot reactions within the same message
LLMs interpret these examples as templates to follow. Because they are pattern-based systems, they mirror this structure in subsequent replies.