Nikola Todorovic, director of software engineering at Google Search, appeared on an episode of Search Off the Record to discuss how AI has evolved within Google Search.
Todorovic leads Google’s SafeSearch engineering team and has worked in the search organization for 15 years. He said it’s difficult to widely deploy machine learning across search because complex models are harder to understand and modify than simpler systems.
He was explaining why Google can’t apply its ML system to the entire search at once. Todorovic said these models can “act like a kind of black box” because engineers don’t necessarily understand what’s going on inside.
That makes it difficult to debug if the search system changes over time or if the model needs to be replaced, he said.
Safe Search as a testing site
Todorovic said SafeSearch is one of the first places Google can bring AI models to search. Because teams can isolate these systems from the main ranking flow.
SafeSearch may run standalone image and video classifiers that generate signals such as how explicit the results are. If a problem occurs, engineers can iterate through the model without interrupting the rest of the search.
He says convolutional neural networks started improving image understanding about 12 years ago, making SafeSearch a natural early use case for machine learning within search.
Introduction to AI built on existing search
Todorovic described the AI brief as a feature that would “put us on top” of Google’s existing search and ranking systems. He said the search and rankings under AI Overview are still what he called “old style, old school.”
This process could include fan-out queries, he said. Google may identify additional queries related to the original input, run them in parallel, and return the retrieved results in a single response.
AI summaries combine and summarize information from selected results, including source text, snippets, titles, and other page context, he said.
AI mode follows a similar pattern, but works more independently, Todorovic said. He explained that it will still be able to run on search, even though it will have its own “bigger platform.”
why is this important
The “black box” quote has gotten a lot of attention, but the full context is important. Todorovic wasn’t saying that Google lacks oversight of AI overview or AI modes, but was explaining why it’s difficult to broadly deploy machine learning across search.
His comments add useful context to Google’s existing AI Search documentation. Google has already said that AI Overview and AI Mode may use query fan-out to generate responses by issuing multiple related searches across subtopics and data sources.
The useful point is not that AI is a “black box.” His comments confirm that traditional search systems remain important for AI overviews, even as Google layers summaries and fan-outs on top.
This way, even as Google changes how results are summarized and displayed, the fundamentals of traditional search will still be relevant to AI capabilities.
For the future
As Google expands on AI Mode, it’s worth noting the difference between AI Overview and AI Mode. Todorovic explained that AI Overview is more isolated from other searches, but AI mode has more of its own infrastructure.
This difference could be important in how Google describes visibility, measurement, and optimization guidance as AI modes expand.
