TLDR: A study analyzed the network traffic of ChatGPT, Copilot, and Gemini mobile apps, revealing distinct app- and content-specific patterns in data volume, protocol usage (e.g., Gemini’s QUIC adoption, ChatGPT’s TLS 1.3 exclusivity), and sustained high upstream activity, which poses new challenges for mobile networks compared to traditional messaging apps. Server Name Indication (SNI) proved crucial for classifying GenAI traffic.
Generative AI (GenAI) chatbots like ChatGPT, Copilot, and Gemini have become incredibly popular, changing how we interact online. However, their impact on network traffic has been largely unexplored until now. A recent study, titled “From Prompts to Packets: A View from the Network on ChatGPT, Copilot, and Gemini”, delves deep into the network traffic generated by these leading chatbots when accessed via their Android mobile apps for both text and image generation.
Conducted by Antonio Montieria, Alfredo Nascitaa, and Antonio Pescapèa from the University of Napoli Federico II, Italy, this research provides crucial insights into how GenAI chatbots utilize mobile networks, highlighting unique characteristics that differentiate them from conventional applications.
How the Study Was Conducted
The researchers set up a dedicated system to capture and label network traffic. They collected two main types of data: a 60-hour ‘generic’ dataset from unconstrained prompts and a ‘controlled’ dataset using identical prompts across GenAI apps. To enable direct comparisons, the controlled prompts and responses were also replicated using traditional messaging apps like WhatsApp and Telegram.
Key Findings: Distinctive Traffic Patterns
The study revealed significant app- and content-specific traffic patterns:
-
Volume and Direction: Copilot and Gemini showed a much higher network load, particularly in terms of data volume, when generating multimodal content (like images) compared to text. ChatGPT, however, generated substantial traffic even for textual responses, likely due to its method of sending responses incrementally in small batches of tokens. Across all apps, downstream traffic (data received by the user) consistently dominated, especially for ChatGPT, where it accounted for over 94% of the total volume.
-
Packet Dynamics: Analyzing individual data flows showed that multimodal content generally involved larger packet sizes. ChatGPT’s traffic often featured sequences of large downstream packets, reflecting its incremental transmission style. Gemini, on the other hand, displayed a pattern of alternating between very small and relatively large packets.
Protocol Adoption and Identification
The research also uncovered interesting differences in how these apps use network protocols:
-
Transport Protocols: ChatGPT and Copilot primarily relied on TLS (Transport Layer Security). Gemini, however, extensively leveraged QUIC, a newer internet transport protocol, in addition to TLS. This suggests varying architectural choices among the GenAI providers.
-
TLS Versions and SNI: ChatGPT exclusively used TLS 1.3, the latest version, while about 25% of Copilot’s traffic still used TLS 1.2. Gemini’s choice of TLS version even varied depending on whether it was generating text or multimodal content. Crucially, the Server Name Indication (SNI) values—which indicate the hostname a client is trying to connect to—were found to be highly specific to each app and content type. For instance, a particular SNI was exclusively linked to image generation in ChatGPT, and another for Gemini.
Classifying GenAI Traffic
The distinct SNI values proved to be powerful identifiers. The study found that GenAI apps could be classified from their network traffic with high accuracy (around 92% F1-score). When SNI information was masked, classification performance dropped significantly by up to 20 percentage points. This highlights the critical role SNI plays in distinguishing between different GenAI apps and even between text and image generation within the same app.
GenAI vs. Traditional Messaging Apps
A direct comparison with WhatsApp and Telegram using identical content revealed stark differences:
-
Higher Load: GenAI chatbots generated significantly more traffic, both downstream and, notably, upstream (data sent from the user’s device). This sustained high upstream activity, even for simple text prompts, represents a novel stress factor for mobile networks.
-
Transmission Strategies: Each GenAI app exhibited unique temporal profiles and transmission strategies. For example, ChatGPT showed multiple upstream bursts, while Copilot concentrated its peaks in the first half of each minute. Messaging apps, in contrast, had much lower rates and distinct, often bursty, patterns.
Also Read:
- The Evolving Landscape of Web Search: AI’s Impact on Sources and Content
- Dynamic Wireless Access: An AI-Driven Game Theory Approach
Implications for Network Management
The findings underscore that GenAI chatbot traffic constitutes a new and distinct category of mobile network usage. This understanding is vital for network operators to anticipate and manage the impact of these rapidly growing services. The sustained upstream activity, in particular, poses new challenges for network monitoring, planning, and resource allocation, emphasizing the need for infrastructure adjustments to accommodate GenAI-driven workloads.
