New research revealed that inspecting ChatGPT’s network traffic uncovered internal labels and specific data providers the artificial intelligence model uses for information retrieval.
The analysis provides a first look at the mechanisms ChatGPT employs to select its sources, challenging common assumptions about its underlying data acquisition processes, according to the researchers.
The study found that ChatGPT assigns an internal label, ‘result_source,’ to each web result, which can take one of four values: serp, labrador, bright, and oxylabs. Researchers Mark Williams-Cook, Metehan and Suganthan Mohanadasan conducted the analysis.
The ‘bright’ label directly corresponds to Bright Data, while ‘oxylabs’ refers to Oxylabs, indicating that ChatGPT uses these third-party data providers for some of its information, the researchers reported.
The methodology involved inspecting browser network traffic through HTTP inspection after TLS decryption, allowing access to JSON metadata associated with ChatGPT’s operations. This approach differed from traditional packet sniffing, the researchers said.
Researchers also identified six ‘turn_use_case’ values, which categorize different types of user queries, according to the study. This internal categorization helps dictate how ChatGPT processes information requests.
Text-based queries often bypassed direct web searches, while more complex ‘Thinking’ queries triggered numerous sub-queries, including site-specific searches and price verifications, the analysis found.
The frequency observations, such as percentages and rankings, were directional due to a small sample size focused on software as a service and tech-related queries, the researchers reported. However, the structural findings regarding internal labels and source types were deemed high-confidence.
Initial data capture was complicated by ChatGPT’s answer streaming over long-lived connections, and automated Chrome browsers used for the study were blocked by Cloudflare, the researchers noted.
OpenAI, the developer of ChatGPT, did not immediately respond to requests for comment on the findings.
Source: Search Engine Journal
