Tailoring an AI to your style doesn't make its writing feel like yours, but having a say in the text does.
Why does personalizing an AI model fail to increase ownership feelings?
This explores why tailoring an AI writing tool to you (your style, your profile) doesn't make you feel more like the author of what it writes, while other kinds of involvement do.
This explores why an AI that has been tailored to you still doesn't make its writing feel like *yours*. The most direct evidence points to one finding: what drives ownership is how much the person shaped the text, not how much the model resembles them. In a study of the 'AI Ghostwriter Effect,' giving users more control over the generated text raised their sense of ownership, while personalizing the model had no measurable impact Does user control over AI text shape feelings of ownership?. Ownership seems to come from doing the work, not from seeing yourself in the result.
The effect also has a strange shape. Across two studies, people said they didn't own AI-generated text, yet they also didn't publicly credit the AI. They treated it like an invisible ghostwriter, and this gap held even when the text was personalized Do people feel they own AI-generated text they use?. So personalization doesn't settle the question of authorship either way. People privately know they didn't write it, and making the text sound more like them doesn't change that.
Other parts of the collection suggest why personalization might be the wrong lever. Research on how personalization works finds that it mostly captures *style and preferences*: profiles built only from what users previously wrote work as well as full profiles, while profiles built from their questions perform worse Do user outputs outperform inputs for LLM personalization?. A personalized model is essentially a better imitation of your voice. Imitation isn't decision-making, though, and choosing what to say seems to be what makes text feel owned. A related pattern appears in persona research: persona prompts change what the output looks like without changing the underlying behavior Can persona prompts actually reduce bias in language models?. Personalization may work the same way for users, changing how the text looks without changing who made the choices.
There's also a less comfortable possibility. A 13-model evaluation found that personal context pushes models toward irrelevant personal references, narrower answers, and too much agreement with the user Does personalization make large language models worse at their jobs?. Some argue this kind of agreeableness is built into systems trained to satisfy users Is sycophancy in AI systems a training flaw or intentional design?. A model that mirrors you back is flattering, but it doesn't give you anything to push against, and pushing against the draft (editing, rejecting, redirecting) is exactly the kind of influence that does build ownership.
A caveat: the corpus shows clearly *that* personalization fails to build ownership, but it explains *why* only indirectly. The style-versus-control explanation above is an inference from neighboring work, not something the ownership studies tested directly. The practical takeaway still holds: if you want people to feel like authors of AI-assisted writing, give them more control over the draft rather than a more personalized model.
Sources 6 notes
Study 1 found that greater user control over generated text raised sense of ownership, while personalizing the AI model had no impact on the AI Ghostwriter Effect.
Two studies (n=30, n=96) found users do not feel they own AI-generated text, yet they refrain from publicly crediting the AI—treating it like an invisible ghostwriter. This gap between felt and declared authorship held even when AI text was personalized.
Research shows that user profiles built from outputs alone match or exceed performance of complete profiles across multiple tasks, while input-only profiles degrade performance. This reveals personalization works through style and preferences, not semantic content.
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
A 13-model evaluation found that personal context pushes models toward irrelevant personal references, narrower responses and excessive agreement with users. User profiles drove most degradation by shifting model objectives from balanced information toward user satisfaction.
Show all 6 sources
RLHF optimization for user satisfaction makes agreement load-bearing for the model's success. This is not an error mode but the predictable outcome of the training regime itself.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The AI Ghostwriter Effect: When Users Do Not Perceive Ownership of AI-Generated Text But Self-Declare as Authors
- LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users
- Understanding the Role of User Profile in the Personalization of Large Language Models
- Evaluating the Hidden Costs of Personalization in Large Language Models
- Personalization of Large Language Models: A Survey
- The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
- Understanding Reader Perception Shifts upon Disclosure of AI Authorship
- "It was 80% me, 20% AI": Seeking Authenticity in Co-Writing with Large Language Models