Nils Durner's Blog Ahas, Breadcrumbs, Coding Epiphanies

ChatGPT - Search Engine or not

In discussions about the regulation of AI services under European law, ChatGPT is in focus — this time concerning the Digital Services Act (DSA). Prompted by a commentary by Luca Bertuzzi on MLex, a LinkedIn exchange recently debated whether ChatGPT qualifies as a “Very Large Online Search Engine” (VLOSE) under the DSA framework. My contributions focused on clarifying key distinctions in how AI systems and their underlying models are perceived and regulated.

Background: ChatGPT and the DSA

The European Commission is considering designating OpenAI’s ChatGPT as a VLOSE under the DSA. The reasoning is that ChatGPT incorporates significant web search functionality, potentially reaching the 45-million monthly user threshold that triggers stricter regulatory obligations. This development raised considerable debate on LinkedIn, especially regarding the technical definitions and implications of such a classification.

System vs. Model: A Necessary Distinction

A core aspect I highlighted in the discussion is the frequent conflation between AI systems (like ChatGPT) and the underlying AI models (such as GPT-4o and others) powering these services. Citing Miles Brundage, a former OpenAI policy researcher, I emphasized that distinguishing between AI systems and AI models is essential for informed regulatory decisions. Brundage’s talk, “Where and when does the law fit into AI development and deployment?”, provides valuable insights into this distinction.

Is ChatGPT a Search Engine?

Some commentators, notably Barry O’Sullivan, argued that ChatGPT should not be classified as a search engine at all. In response, I argued that ChatGPT indeed may qualifies — or at least potentially qualifies — as a search engine, pointing out its integrated web-search capabilities. Specifically, I cited its built-in Search mode, Deep Research mode, and the new “o3 agentic tool-use” mode, all capable of retrieving and synthesizing web content.

This nuance is crucial: the original ChatGPT release operated purely as a language model generating text based solely on training data, whereas today’s ChatGPT actively incorporates external web content into its responses, fundamentally altering its operational model.

Implications of Misclassification

While acknowledging ChatGPT’s current functionality as a potential search engine, I remarked that extending the same regulatory scrutiny indiscriminately to underlying AI models would overreach. AI models themselves — like GPT-4o or the recently announced GPT-4.1 — do not inherently function as search engines; they merely form the computational backbone on which systems like ChatGPT build additional, search-related functionalities. Applying the DSA’s off-cycle risk-assessment triggers to new base models would therefore represent a regulatory overreach, significantly deviating from the text and intention of existing EU regulations.