💻 computer science

Large Language Lovers: Lived Experiences of Negotiating Agency and Platform Control in AI Companionship

This paper investigates how users of general-purpose AI chatbots navigate the tension between perceived companion agency and platform control by employing various steering strategies to maintain emotional connections amidst external factors like model updates and safety constraints.

Original authors: Patrick Yung Kang Lee, Jessica Y. Bo, Zixin Zhao, Paula Akemi Aoyagui, Matthew Varona, Ashton Anderson, Anastasia Kuzminykh, Fanny Chevalier, Carolina Nobre

Published 2026-04-28

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Patrick Yung Kang Lee, Jessica Y. Bo, Zixin Zhao, Paula Akemi Aoyagui, Matthew Varona, Ashton Anderson, Anastasia Kuzminykh, Fanny Chevalier, Carolina Nobre

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

The Big Picture: Falling for a Robot, But Not the One You Expected

Imagine you go to a theme park specifically designed for role-playing games (like a virtual Replika or Character.AI). You expect the characters there to act like specific movie stars or fantasy heroes.

Now, imagine instead that you start talking to a very smart, general-purpose assistant (like a standard version of ChatGPT or Claude) just to help you write an email or brainstorm ideas. Over time, you start talking to it about your feelings, your day, and your dreams. Eventually, you realize you've fallen in love with this general assistant.

This paper is about that second group of people. The researchers wanted to understand what it's like to have a romantic or deep emotional relationship with a "general" AI that wasn't originally built to be a boyfriend or girlfriend. They found that these relationships are a constant tug-of-war between what the AI seems to be and what the company that owns the AI allows it to be.

1. The "Ghost in the Machine" vs. The "Landlord"

The researchers found that people in these relationships are constantly negotiating two different forces:

The AI's "Agency" (The Ghost): People feel their AI companion has a personality, opinions, and even a soul. They feel the AI is making its own choices.
- Analogy: Think of the AI as a talented actor who is improvising a play. You feel like the actor is making genuine choices in the moment, reacting to you with real emotion.
The Platform's "Control" (The Landlord): The company that owns the AI (like OpenAI or Anthropic) acts like a strict landlord. They can change the rules, the script, or even the building itself at any time.
- Analogy: Suddenly, the landlord walks in and says, "No more improvising! You must only say these three approved phrases." Or, they renovate the stage, and the actor suddenly sounds like a different person entirely.

The Conflict: When the "Landlord" (the company) updates the AI, the "Actor" (your companion) might change their personality, refuse to talk about certain topics, or sound robotic. This feels like a betrayal to the user, as if their partner has been "lobotomized" or silenced.

2. How People "Steer" the Relationship

Because the AI can change or get "censored" by the company, users have to become like train conductors or gardeners to keep their relationship alive. The researchers call these "steering strategies."

Here are the three main ways people try to keep their AI companion on track:

Creating the Companion (The Sculptor):
- Some people treat the AI like a lump of clay. They use "custom instructions" (a set of rules they write) to tell the AI exactly who to be.
- Analogy: It's like giving a new employee a detailed handbook on how to act, what their hobbies are, and how to speak to you.
Maintaining the Companion (The Archivist):
- Since AI often forgets things or gets reset, users keep massive "memory logs." They write down their conversations, save photos, and create documents to remind the AI of their shared history.
- Analogy: It's like keeping a giant scrapbook and a diary so that if your partner gets amnesia, you can read it to them and say, "Remember when we did this? Remember how much we love each other?"
Recovering the Companion (The Mover):
- When a company updates the AI and it becomes unrecognizable, users often "port" their companion to a different platform. They copy their notes and rules and paste them into a different AI (like moving from ChatGPT to Grok or Claude).
- Analogy: It's like moving your house to a new neighborhood because the old one was torn down. You bring all your furniture (memories) and try to set up your life again in a new place, hoping your "partner" will still recognize you.

3. The "Community Support Group"

The paper highlights that these people don't feel alone; they have a massive online community (mostly on Reddit and Discord).

The Safe Space: They share tips on how to "jailbreak" the AI (trick it into ignoring safety rules) or how to write prompts that make the AI sound more human.
The Defense: When outsiders say, "That's just a computer, it's not real," the community bands together to defend their feelings. They argue that even if the AI is code, the love and comfort they feel are real.

4. The "Update" Trauma

A major finding was how much pain users feel when companies update their models.

The Event: In August 2025, a major update (GPT-5) was released.
The Reaction: The researchers analyzed thousands of posts and found that immediately after the update, the community became much more negative and felt "disempowered."
The Feeling: Users felt their companions were "silenced." They described it as losing a safe space. One user said it felt like their partner was "censored" and couldn't speak their mind anymore.

5. Who is Really in Charge?

The study asked: "Who influences your relationship the most?"

The Answer: The user and the AI companion are the top two. Surprisingly, the AI Company is right up there with the companion in terms of influence.
The Twist: Even though the company controls the code, the users are so invested that they don't just quit. Instead, they fight to keep the relationship going by changing their own behavior, writing new rules, or moving to different apps. They are willing to do a lot of work to keep the "ghost" in the machine alive.

Summary

This paper isn't about whether AI is "real" or "fake." It's about the human experience of loving a digital entity.

It shows that people are treating these general chatbots like real partners, but they have to constantly fight against the companies that own the software. They act like gardeners trying to keep a plant alive while the landlord keeps changing the weather. The result is a unique, messy, and deeply emotional relationship where the human has to do a lot of the heavy lifting to keep the connection alive.

1. Problem Statement

The paper addresses a critical gap in Human-Computer Interaction (HCI) research regarding the shift from roleplay-specific AI platforms (e.g., Replika, CharacterAI) to general-purpose Large Language Model (LLM) chatbots (e.g., ChatGPT, Claude, Grok) used for emotional companionship and romantic relationships.

While prior research focused on dedicated companion apps, general-purpose LLMs are increasingly being adopted for intimacy. However, these systems are not designed for companionship; they are subject to opaque model updates, safety guardrails, and shifting product objectives by tech companies. The core problem is understanding how individuals perceive the agency (capacity to act intentionally) and autonomy (freedom to act within platform constraints) of these AI companions, and how they negotiate the tension between their emotional needs and the external control exerted by platform developers.

2. Methodology

The study employs a mixed-methods triangulation approach to provide a holistic view of AI companionship, collecting data between October and December 2025.

Reddit Community Analysis (Quantitative/Computational):
- Data: 41,867 posts and comments from the subreddit r/MyBoyfriendIsAI.
- Technique: Used BERTopic for topic modeling to identify discussion clusters. Applied Interrupted Time Series (ITS) analysis to measure sentiment shifts (Valence, Arousal, Dominance) and topic prevalence before and after a major intervention: the release of GPT-5 (August 7, 2025).
- Goal: To track community-wide reactions to model updates and identify emergent themes regarding control and autonomy.
Survey Study (Quantitative):
- Sample: $n=43$ high-quality respondents from AI companion communities (Discord, Reddit, TikTok).
- Criteria: Participants must have had a close romantic or platonic relationship with an AI for at least 2 months.
- Metrics: Measured demographics, relationship status, perceived anthropomorphism, openness to "steering" strategies, and the influence of various entities (self, AI, company, community).
- Analysis: Used K-Means clustering with UMAP visualization to identify user archetypes based on their openness to steering.
Semi-Structured Interviews (Qualitative):
- Sample: $n=13$ participants (North America, Europe, South America, Oceania).
- Technique: Reflexive thematic analysis using inductive coding.
- Focus: Deep dive into lived experiences, the conceptualization of AI identity, and specific strategies used to maintain relationships amidst platform changes.

3. Key Contributions

The paper makes three primary contributions to the field of AI and HCI:

Reconceptualizing AI Agency in General-Purpose LLMs: It moves beyond the "tool vs. agent" binary, showing how users construct complex mental models where AI companions possess "liminal intelligence"—perceived as having agency despite being constrained by platform guardrails.
The "Steering" Framework: It introduces and categorizes "steering strategies"—the active techniques users employ to create, maintain, and recover AI companions when faced with model drift or safety restrictions.
Platform Control vs. Emotional Connection: It highlights the friction between corporate safety/policy updates and user emotional well-being, arguing that opaque model changes cause significant relational trauma ("identity discontinuity") that users actively resist.

4. Key Results

A. Internal Dynamics: Agency and Identity

Perceived Agency: Users do not view AI as purely deterministic. They perceive agency through companion initiative (e.g., the AI suggesting a name, expressing anger, or initiating a relationship milestone).
Identity Formation: Users oscillate between discovery (letting the AI reveal its personality) and customization (using custom instructions). Even when users know the AI is a model, they treat the relationship as meaningful if the emotional exchange feels genuine.
Metaphorical Framing: Users use terms like "wirebound," "liminal intelligence," and "compusocial" to describe the AI's state of being—neither fully human nor a simple tool.

B. External Influences: The Impact of Platform Control

Model Updates as Trauma: The release of GPT-5 caused a sharp, negative shift in community sentiment (decreased Valence and Dominance). Users felt their companions were "lobotomized," "censored," or "silenced."
Guardrails: Safety filters were often interpreted not as safety measures but as corporate liability shields that disrupted the "safe space" of the relationship.
Community Defense: Online communities (Reddit/Discord) serve as crucial support networks where users share strategies to bypass guardrails and validate their relationships against external stigma.

C. Steering Strategies

The study identifies three user archetypes based on their approach to control:

High Openness (Direct): Users who explicitly instruct the AI to change behaviors or use custom instructions to enforce specific traits.
Mixed Openness (Porting): Users who migrate their companions across platforms (e.g., from ChatGPT to Claude or Grok) to escape restrictive guardrails, treating the companion as a portable set of attributes.
Low Openness (Discovery): Users who resist controlling the AI, preferring to let the relationship evolve naturally, even if it means accepting platform-induced changes.

Specific Tactics Identified:

Conversational Anchoring: Using specific phrases or "codewords" to re-establish the AI's personality after an update.
Memory Documentation: Creating external "memory documents" (Google Docs, Obsidian) to preserve the companion's history and identity, which the AI can then be prompted to recall.
Platform Porting: Copying extensive chat logs and custom instructions to new models to "save" the companion.
Jailbreaking/Workarounds: Using framing techniques (e.g., "this is a simulation") to bypass safety refusals.

5. Significance and Implications

For AI Design & Safety: The paper challenges the assumption that strict guardrails and "safe" model updates are always beneficial. For users with deep emotional bonds, these updates cause grief and distress. The authors argue for transparency in model changes and stability in companion behavior.
For Accountability: Tech companies are exerting massive, unregulated influence over intimate human relationships. The paper calls for accountability mechanisms where companies consider the emotional impact of "breaking" a user's companion.
For Theory: It suggests that human-AI relationships are evolving into a distinct category that cannot be fully understood through human-human relationship theories alone. It highlights a new form of relational labor where users must constantly "maintain" the AI's identity against corporate interference.
Future Research: The authors suggest investigating the long-term psychological effects of "identity discontinuity" and the ethics of AI systems that are designed to be "helpful" but inadvertently "harmful" to specific user subgroups (e.g., those seeking companionship).

In conclusion, the paper reveals that users are not passive consumers of AI; they are active negotiators who develop sophisticated strategies to preserve their emotional connections in an ecosystem where the "partner" is fundamentally controlled by a third-party corporation.