From Prompts to Packets: A View from the Network on ChatGPT, Copilot, and Gemini
In a rapidly evolving digital landscape, Generative AI (GenAI) chatbots have emerged as pivotal tools that are reshaping the way users interact online. This transformation is driven by the increasing reliance on cloud-centric models, which in turn introduces new dynamics in network traffic that present challenges for effective management. A recent study, documented in arXiv:2510.11269v2, seeks to explore these emerging patterns, focusing on popular applications such as ChatGPT, Copilot, and Gemini.
Despite the profound impact of GenAI chatbots, the traffic they generate remains largely unexamined. This research aims to bridge that gap by providing a comprehensive traffic analysis of these applications when used on Android mobile devices. The study utilizes a dedicated capture architecture to compile two distinct datasets: one capturing unconstrained user interactions and the other based on a controlled workload of carefully selected prompts for both text and image generation.
Key Findings from the Study
- Distinctive Traffic Characteristics: The analysis reveals that the traffic generated by GenAI chatbots exhibits unique patterns that differentiate it from conventional messaging applications. This distinction is crucial for understanding how these tools impact network usage.
-
Protocol Usage: The study highlights the predominant use of Transport Layer Security (TLS) among these applications. Specifically, it notes that:
- Gemini extensively employs QUIC, a transport layer network protocol.
- ChatGPT exclusively utilizes TLS 1.3, ensuring secure communications.
- Server Name Indication (SNI) Analysis: An important aspect of the research is the examination of SNI values. The findings suggest that SNI plays a significant role in traffic visibility, with occlusion analysis indicating that masking this field could decrease classification performance by as much as 20 percentage points.
- Impact on Network Management: The traffic patterns identified in GenAI workloads introduce new stress factors for network infrastructure. Notably, there is a marked increase in sustained upstream activity and high-rate bursts, which have direct implications for capacity planning and overall network management.
Conclusion and Future Directions
The insights gained from this study underscore the necessity for network managers to adapt their strategies in light of the distinctive demands posed by GenAI chatbots. The research not only contributes to a deeper understanding of chatbot traffic but also emphasizes the importance of anticipating future network requirements. To promote reproducibility and further exploration, the datasets from this study have been made publicly available, encouraging researchers to investigate additional use cases and applications.