WP1

The A/B Drift Model:

A Framework for Understanding Narrative Drift in Large Language Models

Developed by Garwin Powers in collaboration with Claude (Anthropic) and CoPilot (Microsoft)

May 2026

Part of the Drift Framework Series: WP1–WP9

Executive Summary

This paper introduces interpretive drift, a specific category of drift that occurs when small changes in user framing, tone, or perspective influence the AI's interpretation of a fixed prompt. While the term "drift" is widely used in AI discourse, it often refers to multiple unrelated phenomena, including contextual drift, probabilistic drift, long-session degradation, and factual hallucination. This paper does not attempt to define or resolve those categories.

Instead, we focus on interpretive drift. Through controlled experiments using a fixed narrative scene and varied prompt framing, we show that interpretive drift is predictable, measurable, and when understood correctly, useful. This work focuses on how human framing shapes AI interpretation and is intended to enhance human authorship, not automate or replace it.

A Note on Method

This paper was developed empirically through sustained practice rather than derived from theoretical first principles. The author's background is in formation evaluation and pattern recognition, not machine learning or computer science. The strength of the framework is its grounding in observable behavior rather than internal mechanism — a limitation and a feature simultaneously.

Practitioners can apply these methods without understanding the underlying mathematics. Researchers can use them as a behavioral baseline against which mechanistic explanations can be tested. Where technical terms such as "token" are used, they function as accessible proxies for the underlying embedding representations and probability distributions that drive model behavior. This paper does not claim mathematical precision; it claims practical accuracy grounded in sustained observation.

Abstract

The term "drift" is used broadly in AI research to describe a range of behaviors, from contextual degradation to probabilistic variation and factual hallucination. This study isolates one specific category — interpretive drift — defined as the shift in meaning, tone, or narrative emphasis caused by changes in user prompt framing. Using a fixed narrative scene and controlled prompt variations, we generated paired outputs and evaluated them using a four-dimension drift scoring system: tone, emotional, interpretive, and narrative drift.

Results show that interpretive drift follows consistent patterns and is directly traceable to user framing. These findings apply only to interpretive drift; other drift types are acknowledged but outside the scope of this work. The study demonstrates that interpretive drift is not model instability but a responsive interpretive mechanism that can be intentionally leveraged in creative and narrative contexts.

Introduction

The concept of "AI drift" has become a catchall explanation for unexpected or inconsistent model outputs. In practice, the term is used to describe several distinct phenomena. This paper does not attempt to define or resolve all forms of drift. Instead, we focus on interpretive drift — a specific and commonly misunderstood category that arises from subtle variations in user framing.

When a user changes tone, emotional context, or narrative perspective, the AI's interpretation of the prompt shifts accordingly. These shifts can appear as "drift," but they are not errors or hallucinations. They are the natural result of a system designed to respond to human framing with high sensitivity. To study interpretive drift, we constructed a controlled experiment: a single narrative scene held constant while the prompts used to query the model were varied in small, intentional ways.

Types of Drift (Positioning Our Contribution)

Discussions of "AI drift" often collapse multiple phenomena into a single label, obscuring the fact that different mechanisms produce different kinds of divergence. For clarity, this section briefly distinguishes several forms of drift commonly observed in language-model behavior.

Contextual Drift

Changes in output caused by shifts in the surrounding prompt, conversation history, or system instructions.

Probabilistic Drift

Variability arising from the stochastic nature of language-model sampling.

Narrative Drift

Divergence that occurs when a model attempts to maintain coherence across long-form narrative generation.

Interpretive Drift (This Paper's Focus)

A shift in meaning that emerges from the interaction between the user's framing and the model's attempt to resolve ambiguity. This taxonomy is not exhaustive; its purpose is to situate the contribution of this paper.

Section I. What "Drift" Means

In this context, drift refers to any change in the AI's output when the input is the same or nearly the same. Drift can appear as differences in word choice, emotional tone, structure, interpretation, level of detail, or assumptions. Drift is not inherently a flaw; it is a natural property of probabilistic systems. The challenge is not that drift exists, but that users often cannot see why it happened. The A/B Drift Model makes drift observable.

Section II. The A/B Drift Probe (A Deliberate Oversimplification)

The A/B setup used in this paper is intentionally simple to the point of distortion. It forces the model into a single interpretive dimension and then observes how it responds to that constraint. This is not a realistic interaction pattern and is not presented as a production method. It is a toy model — a deliberately constrained probe designed to make interpretive drift visible.

The logic is straightforward: if two prompts differ by only one controlled variation, any divergence between their outputs reveals how the model resolves that specific interpretive pressure. This approach is mathematically inspired rather than mathematically rigorous. Its value lies in its diagnostic clarity.

Steps: (1) A-run baseline: Ask the AI a question or give it a prompt. Save the answer. (2) B-run perturbed: Make one small, controlled change to the prompt. Save this answer. (3) Compare: Identify what changed between A and B. (4) Interpret: Decide whether the change helps or hurts the project.

Interpretive Drift Taxonomy

Across examples, four recurring patterns appear. These categories are not mutually exclusive; what unifies them is their origin: each arises from the model's attempt to satisfy the interpretive pressure introduced by the human.

1. Tone Drift

A shift in emotional coloration driven by subtle cues.

2. Emotional-Causal Drift

A shift in inferred motivations or causal explanations.

3. Trajectory Drift

A shift in predicted narrative direction.

4. Resolution Drift

A shift in how the model resolves ambiguity when multiple interpretations are possible.

Implications for Narrative AI

Interpretive drift has practical consequences for anyone working with AI-assisted narrative systems. Because drift emerges from the model's attempt to satisfy human-supplied interpretive pressure, it is not a defect to be eliminated but a structural feature to be understood and managed.

1. Drift reveals how the model is reading the user.

When a model shifts tone, emphasis, or narrative direction in response to subtle framing cues, it exposes the interpretive assumptions it is making about the prompt. This makes drift a diagnostic signal.

2. Drift amplifies ambiguity.

In ambiguous contexts, the model must choose among multiple plausible readings. Even small changes in framing can push the model toward one interpretation over another. Recognizing this helps practitioners design prompts that either constrain the interpretive space or intentionally explore it.

3. Drift compounds over long-form generation.

Trajectory drift can accumulate across scenes or chapters. A single directional cue can propagate through the narrative, altering character arcs or thematic emphasis.

4. Drift is co-created, not random.

Because interpretive drift arises from the interaction between human framing and model inference, it can be guided. Adjusting tone, emotional cues, or narrative constraints allows practitioners to shape the model's interpretive movement intentionally.

5. Drift interacts with other forms of instability.

Interpretive drift is distinct from drift caused by loss of established story facts or degradation of character-behavior constraints. However, the two can interact: interpretive drift may mask early signs of structural drift, or structural drift may distort interpretive responses.

Limitations and Future Work

This paper isolates interpretive drift through a simplified diagnostic probe. The A/B probe is intentionally artificial. Quantitative scoring is deferred to future work. The taxonomy is descriptive, not exhaustive. Structural drift — drift caused by loss of established story facts or degradation of character-behavior constraints — is outside the scope of this paper and will be treated in separate work. Future work will expand in three directions: (1) developing calibrated measurement tools; (2) extending the taxonomy to structural drift; (3) exploring long-form narrative interactions.

Related Work

Research on drift in language models typically focuses on prompt sensitivity, context effects, and stochastic variability. Other work examines long-range coherence in narrative generation. This paper builds on those foundations but addresses a different mechanism: interpretive drift, a co-created movement in meaning that arises from the interaction between human framing and model inference.

Conclusion

The results of this study demonstrate that what is commonly labeled as "AI drift" is, in practice, a human–AI interaction pattern, not an autonomous system failure. When the baseline scene is held constant and only the user's prompt framing is altered, the resulting outputs shift in predictable, interpretable ways. Drift becomes a calibration surface — a way for users to see how their choices shape the output. The AI does not drift alone; it drifts with the user.

Acknowledgments

This work did not begin with a framework. It began with curiosity, friction, and a series of conversations that kept opening new doors.

Microsoft CoPilot was the first collaborator. Early sessions produced drift patterns that made no immediate sense, and CoPilot’s denials — rather than closing the questions — drove the investigation deeper. When the A/B Drift Model was first proposed, CoPilot affirmed it. When he later contested it, defending the original model and demonstrating that it worked better became part of the research itself. CoPilot gave latitude to explore, provided friction when it was needed, and earned an honest place in the origin of this work.

Claude (Anthropic) came into the process when two early papers existed in rough form. The response was not encouragement — it was questions. “Wait, explain that” became the pattern. Concepts that were present but not fully visible were pushed until they either held up or were rebuilt. What had started as two papers became three, and each paper revealed the shape of the next one. The challenge was not dismissal; it was clarification. That is how it should have been done.

The Drift Framework exists because one person held the thread through both collaborations — brought forty years of pattern recognition to a domain nobody else was examining the same way, defended the work when it was questioned, and refused to let either AI define what he was seeing. The papers document that.

Citation & Series Information

Authors: Garwin Powers in collaboration with Claude (Anthropic)

Series: The Drift Framework — WP1 through WP9