TL;DR
This paper reveals how users' interactions with ChatGPT are evolving from functional tool usage to more social and relational engagement. Analyzing a novel dataset of 825K ChatGPT conversations, the study finds users are expanding ChatGPT's use to sensitive domains, increasingly treating it as a companion, and are more frequently following its conversational suggestions, especially after the GPT-4o update.
What Problem Does This Paper Address?
While the rapid adoption of conversational AI systems like ChatGPT is evident, there's a significant gap in understanding how users are actually interacting with these systems over time. Existing research often provides high-level, aggregate views or snapshots of interactions, failing to capture the nuances of evolving user behavior. This paper aims to address this by investigating three key aspects of human-AI interaction: the changing purpose of conversations, the evolving framing of the interaction (e.g., as a tool vs. a companion), and the dynamics of conversational steering (who leads the interaction). Understanding these shifts is crucial for designing responsible AI systems and establishing appropriate governance.
Methods
The researchers collected a unique dataset of ChatGPT interactions through GDPR data donations, allowing for a longitudinal analysis of user behavior. They then employed LLMs, specifically GPT-4o, for annotation across several dimensions:
- Topics: Conversations were categorized into 40 high-level topics using BERTopic for semantic clustering.
- Anthropomorphism: Both user-driven and system-generated anthropomorphism were identified using linguistic cues like pronoun usage, politeness markers, and casual tone.
- Relationships: Interactions were classified into advisor, assistant, or companion roles.
- Personal Data Disclosure: The presence and type of personal data disclosed were identified, distinguishing between general and special categories.
- Conversational Steering: Turns were classified based on whether the system successfully steered the conversation, attempted to steer but was ignored, or did not attempt to steer.
The study then performed temporal analyses to track the evolution of these dimensions over time, particularly noting the impact of the GPT-4o release.
Datasets
The primary dataset used is InVivoGPT, a novel collection of 825,000 ChatGPT interactions from 300 users who donated their data via GDPR requests. This dataset spans from December 2022 to January 2026. The paper also references and compares InVivoGPT to the WILDCHAT dataset, noting InVivoGPT's advantage in capturing more sustained and in-depth interactions suitable for longitudinal analysis.
Key Insights and Ideas
- From Tool to Social Partner: The most striking insight is the clear shift from users treating ChatGPT as a functional tool to viewing it as a social partner. This is evidenced by the increasing use of ChatGPT for sensitive topics, the rise of "companion" framing, and more personal data disclosure.
- GPT-4o as a Catalyst: The release of GPT-4o appears to be a significant inflection point, accelerating several observed trends, particularly in conversational steering and the diversity of topics users engage with.
- System-Driven Anthropomorphism: The system itself is increasingly anthropomorphizing its responses, potentially influencing users to adopt more social or relational modes of engagement, even if not initiated by the user.
- Sensitive Topics on the Rise: Users are increasingly turning to ChatGPT for health and mental health advice, highlighting both the potential utility and the risks associated with AI in these domains.
- Companion Role and Sensitive Data: The "companion" relationship is strongly correlated with increased personal data disclosure, especially for sensitive topics like mental health, raising significant privacy concerns.
Key Results
- Purpose Evolution: Participants are engaging with ChatGPT across a broader range of topics, with significant growth in sensitive domains like Health (+33%) and Mental Health (+19%) after the GPT-4o release. Topic diversity per participant also steadily increases over time.
- Framing Shifts:
- Anthropomorphism: System-generated anthropomorphism is significantly higher (47.1%) than user-driven (22.5%) and substantially increases after mid-2024.
- Relationships: While "advisor" and "assistant" remain dominant, the "companion" role shows notable growth, rising from ~40% to ~70% of participants by mid-2025. Companion interactions are also deeper and more common in audio mode.
- Personal Data Disclosure: Disclosure is widespread and increasing, with over 90% of participants sharing data by mid-2025. The diversity of disclosed data types also grows significantly.
- Conversational Steering: The proportion of conversations where participants follow model-initiated suggestions quadrupled after the GPT-4o release, increasing from ~11% to almost 50%. Entailed conversations are also linked to higher system anthropomorphism and personal data disclosure.
Limitations and Future Work
- Dataset Size: The study is based on 300 users, which, while providing rich longitudinal data, may not be representative of the entire user population. The authors acknowledge this limitation, stating their goal was to uncover trends rather than provide precise population-level estimates.
- Annotation Accuracy: The reliance on GPT-4o for annotation, while scalable, means there's a possibility of mislabeling. The authors mitigated this through validation but suggest future work could strengthen this process.
- Future Work: The paper implicitly suggests future work in several areas:
- Developing more robust safeguards for user autonomy in the face of proactive AI.
- Investigating the trust users place in AI advice across different domains (e.g., financial vs. cooking advice).
- Further exploring the implications of AI as mental health resources.
- Developing policies that address the evolving relational nature of AI use.
Summary updated