💬 NLP

Writing in Symbiosis: Mapping Human Creative Agency in the AI Era

This paper analyzes longitudinal writing data to propose a "Dual-Track Evolution" model, demonstrating that human creative agency is coevolving with AI through diverse adaptation patterns—including stylistic convergence, divergence, and stability—rather than succumbing to the feared homogenization of human expression.

Original authors: Vivan Doshi, Mengyuan Li

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vivan Doshi, Mengyuan Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you and a friend are both learning to paint. Suddenly, a super-smart robot joins the art class. This robot can paint anything perfectly, instantly, and in a very specific, polished style.

The big question everyone is asking is: Will everyone start painting exactly like the robot, making all art look the same? Or, will humans find new ways to paint that are still uniquely theirs?

This paper, written by Vivan Doshi and Mengyuan Li, investigates exactly that. They looked at thousands of people writing online (in chat rooms and academic papers) before and after AI writing tools became popular. Instead of just saying "AI is taking over," they discovered something much more interesting: Humans are reacting in three very different ways.

Here is the breakdown of their findings, using some simple analogies:

1. The "Dual-Track" Discovery

The researchers found that human writing is evolving on two separate tracks at the same time:

Track A (The Topic): Everyone is talking about AI. Just like how everyone suddenly started talking about the weather when a storm hits, writers are naturally discussing AI tools, prompts, and how they work. This is a convergence (everyone moving together).
Track B (The Style): This is where it gets wild. While everyone is talking about the same thing, how they are writing is splitting apart. Some people are sounding more like robots, some are trying to sound less like robots, and some are finding a middle ground.

2. The Three "Creative Archetypes"

By analyzing the writing styles of over 2,000 authors, the researchers found three distinct groups, or "archetypes," emerging from the crowd:

🛡️ The Resistors (The "Hand-Painted" Purists)

Who they are: About 21% of the writers.
What they do: They are actively fighting to keep their writing sounding human. Even though AI tools are available, they are doubling down on complex, messy, and unique sentence structures that AI usually avoids.
The Analogy: Imagine a baker who sees a machine that can make perfect, identical bread. The Resistor is the baker who says, "No thanks," and starts kneading the dough by hand, adding weird spices and making the bread look slightly imperfect on purpose. They value authenticity over speed.

🤖 The Adopters (The "Co-Pilot" Collaborators)

Who they are: About 18% of the writers.
What they do: They have fully embraced the AI style. Their writing has become smoother, more predictable, and statistically very similar to what a robot would write.
The Analogy: These are the bakers who say, "Let's use the machine!" They aren't trying to hide it; they are happy to let the machine do the heavy lifting. They are exploring a new way of creating where the line between "human" and "machine" is blurred. They are co-creators.

⚖️ The Pragmatists (The "Smart Editors")

Who they are: The largest group, about 41%.
What they do: They are the most interesting. They talk about AI a lot (using the tools to explore new topics), but they keep their own unique writing voice. They use AI for the "what" (the ideas) but keep the "how" (the style) strictly human.
The Analogy: These are the bakers who use the machine to mix the ingredients quickly but then take over to shape the loaf, add the final glaze, and decide the final look. They treat AI as a tool, not a replacement. They are the "editors" of the AI age.

3. The "Stigma" Twist

The paper also noticed a funny trend over time.

Early 2023: When AI first arrived, many people's writing started to sound a bit more like AI (maybe because they were using it to help write).
Late 2023/2024: Once people realized that "AI-sounding" writing was getting flagged or looked down upon, many of those same people started changing their style back. They started writing in more complex, human ways to prove they were real humans.

It's like a fashion trend: When everyone started wearing a specific hat, it became cool. But once everyone wore it, the cool kids started taking the hats off to stand out again.

Why Does This Matter?

The main takeaway is that humans are not passive. We aren't just being turned into robots by AI. Instead, we are adapting in smart, strategic ways.

For AI Detectors: This is bad news for simple "Human vs. Robot" detectors. If a "Resistor" writes a complex story and an "Adopter" writes a simple one, a detector might think the Resistor is a robot and the Adopter is human! We need smarter ways to understand these different styles.
For Creativity: It's good news for humanity. It shows that even with powerful tools, people still value their unique voices. We are negotiating a new relationship with technology, finding a balance between using the tool and keeping our soul.

In short: We aren't losing our humanity to AI. We are just learning how to dance with it, and everyone is dancing to their own beat.

1. Problem Statement

The rapid proliferation of Large Language Models (LLMs) has shifted the human-machine relationship from passive assistance to active collaboration. While existing research focuses on the homogenization of writing styles (the emergence of a uniform "AI style") and the development of AI detectors, these studies often treat human-AI interaction as a unidirectional influence. They frequently fail to account for human agency—the adaptive choices authors make in response to AI availability and detection awareness.

The core research question is: How is human creative agency evolving in a symbiotic relationship with AI? Specifically, does the presence of LLMs lead to a monolithic convergence of style, or do authors exhibit structured, diverse adaptation strategies?

2. Methodology

The authors employed a rigorous, three-phase methodology to analyze longitudinal writing data across formal and informal genres.

A. Corpus Curation

Data Sources: A large-scale corpus of 50,728 documents derived from:
- Formal Discourse: Computer science preprints from arXiv (filtered for permissive licenses).
- Informal Discourse: Public messages from the "Discord Unveiled" dataset (2015–2024).
Temporal Boundaries: The data is split into a Pre-LLM era (before Nov 30, 2022) and a Post-LLM era (after Nov 30, 2022).
Controls: To isolate stylistic changes from topical shifts, the authors used topic-controlled stratified sampling and residualized thematic content using a 100-topic transformer model.
Reference Baseline: A corpus of purely AI-generated text (ShareGPT-90k, Dolly-15k) was used to establish machine stylistic baselines.

B. Feature Engineering: The "Perplexity Gap"

To distinguish stylistic signals from thematic content, the authors developed a novel metric called the Perplexity Gap ( $\Delta$ ppl):

Pre-LLM Judge: A GPT-2 Medium model trained exclusively on pre-2022 human data (847M tokens), ensuring it has no exposure to AI-generated patterns.
Current Baseline: Llama-3-8B-base, representing modern LLM capabilities.
Metric Calculation: For each document $x$ $x$ , the gap is calculated as:
$\Delta \text{ppl} = \frac{-\ln(p_{\text{GPT2}}(x))}{|x|_{\text{chars}}} - \frac{-\ln(p_{\text{Llama}}(x))}{|x|_{\text{chars}}}$
- Interpretation: A high positive gap indicates text that is "easy" for modern LLMs but "hard" for pre-2022 models, signaling the adoption of LLM-era linguistic patterns. A low or negative gap suggests text remains challenging for modern models, preserving distinct human complexity.
Additional Features: The vector includes changes in Type-Token Ratio (TTR), Flesch-Kincaid Grade Level (FKGL), passive voice ratio, first-person pronoun frequency, punctuation density, and sentence length. All features are z-scored within-author across the temporal boundary.

C. Analytical Framework

Macro-Level: Used the PELT algorithm for change-point detection to identify structural breaks in thematic and stylistic trends over time.
Micro-Level (Author Archetypes): Applied HDBSCAN clustering to the standardized change vectors of 2,100 authors. This unsupervised approach allowed distinct behavioral patterns to emerge without forcing predefined categories.
Validation: Used cross-validation, bootstrap resampling, and the Adjusted Rand Index (ARI) to ensure cluster robustness.

3. Key Contributions

Dual-Track Evolution Hypothesis: The paper proposes and validates a model where human-AI coevolution occurs on two tracks simultaneously:
- Track 1 (Thematic): Universal convergence on AI-related topics.
- Track 2 (Stylistic): Structured divergence, where authors adopt distinct strategies rather than a single homogenized style.
The Creative Archetype Map: The identification of three distinct authorial archetypes based on their adaptation strategies:
- Resistors (21%): Maintain pre-LLM linguistic complexity. They show low/negative perplexity gaps, suggesting a deliberate preservation of unique human stylistic signatures, prioritizing authenticity over efficiency.
- Adopters (18%): Exhibit high perplexity gaps. Their writing becomes significantly easier for modern LLMs, indicating the adoption of smoother syntactic structures and predictable lexical choices characteristic of the LLM era.
- Pragmatists (41%): The largest group. They show moderate stylistic adaptation but high engagement with AI-related themes. They appear to use a "division of labor," using AI for thematic exploration while maintaining their personal voice.
Dynamic Temporal Analysis: The study reveals that stylistic adaptation is not linear. It shows an initial convergence (early 2023) followed by a stylistic avoidance trend (late 2023–2024), particularly in formal writing, likely driven by the stigma of AI detection and editorial pressure.

4. Key Results

Thematic Convergence: Both formal and informal discourse saw a statistically significant, sharp increase in AI-related terminology starting Q1 2023.
Stylistic Divergence:
- Mean perplexity gaps initially rose (23% in social, 15% in formal writing) in early 2023.
- By late 2023, these gaps declined (18% in social, 12% in formal), indicating a reversal where authors actively differentiated their writing from machine patterns.
Cluster Stability: The HDBSCAN clustering yielded a silhouette score of 0.426 and an Adjusted Rand Index of 0.891, confirming robust, distinct archetypes.
Predictive Power: The identified archetypes could predict future author texts with an AUC of 0.813.
Detection Implications: The findings challenge binary "Human vs. Machine" detection. An "Adopter's" text may be statistically closer to AI output than a "Resistor's," suggesting that current detectors may fail to account for human adaptation strategies.

5. Significance and Implications

Redefining Authorship: The paper argues that human creativity is not being erased but is being renegotiated. Authors are active agents developing diverse strategies (collaboration, resistance, or curation) rather than passive recipients of technological influence.
Limitations of Binary Detection: The existence of "Adopters" suggests that simple AI detection tools may be insufficient and potentially biased, as they may flag human authors who have intentionally or unintentionally adopted AI-like styles.
Policy and Tool Design: The findings suggest a need for "archetype-aware" models in AI detection and policy. Furthermore, AI tools should be designed to enhance diverse human voices rather than enforce a single "optimal" style, protecting linguistic diversity and cultural nuance.
Future Research: The study highlights the need to move beyond aggregate statistics to understand individual agency, offering a framework for studying human-AI coevolution in other creative and professional domains.

In conclusion, the paper demonstrates that the human response to AI is not a monolithic trend toward machine-like writing but a complex, structured evolution characterized by thematic convergence and stylistic differentiation.