Google Engineer Demystifies ‘Black Box’ AI Search Models

Nikola Todorovic, a Director of Software Engineering at Google Search, recently joined the Search Off the Record podcast to unpack how artificial intelligence has been integrated into the company’s core search engine over time.

With 15 years of experience in the search organization, Todorovic leads the SafeSearch engineering team. He explained that deploying machine learning broadly across Search was a gradual process, largely because complex models are inherently harder to diagnose and maintain than simpler, rule-based systems. He noted that these advanced models can “function like a kind of a black box,” meaning engineers do not always have full visibility into their internal operations. This opacity makes debugging particularly challenging when search systems evolve or when a model must be swapped out.

SafeSearch as a Testing Ground

Todorovic revealed that SafeSearch served as an ideal proving ground for early AI deployment. Because the system could be isolated from the main ranking flow, the team could run standalone image and video classifiers that produced a signal,like how explicit a result might be. If issues emerged, engineers could iterate on the model without risking disruption to the rest of Google Search.

He noted that convolutional neural networks began significantly improving image understanding roughly 12 years ago, making SafeSearch a natural early candidate for machine learning integration.

AI Overviews Built on Traditional Search

When discussing AI Overviews, Todorovic described the feature as something that “stamps on top” of Google’s existing retrieval and ranking systems. He emphasized that the underlying retrieval and ranking infrastructure remains what he called “the old style, the old school.”

The process often involves fan-out queries, he explained. Google may identify additional queries related to the original input, run them in parallel, and then combine the retrieved results into a single response. AI Overviews then synthesize and summarize information from selected results, drawing on source text, snippets, titles, and other page context.

AI Mode follows a similar pattern but operates with greater autonomy. Todorovic described it as still running on Search, while having a “bigger platform for its own.”

Why This Matters

The “black box” remark has attracted attention, but the full context is critical. Todorovic was explaining why machine learning was difficult to roll out broadly across Search, not suggesting that Google lacks oversight of AI Overviews or AI Mode.

His comments add useful context to Google’s existing AI Search documentation. The company has already stated that AI Overviews and AI Mode may use query fan-out, issuing multiple related searches across subtopics and data sources to develop responses.

The key takeaway is not that AI is a “black box” in a concerning sense. Rather, his remarks reinforce that traditional Search systems remain foundational to AI Overviews, even as Google layers summarization and fan-out on top. This keeps conventional search fundamentals relevant to AI-driven features, even as the presentation of results evolves.

Looking Ahead

The distinction between AI Overviews and AI Mode is worth monitoring as Google expands AI Mode. Todorovic described AI Overviews as more isolated from the rest of Search, while AI Mode has more of its own dedicated infrastructure.

That difference may shape how Google explains visibility, measurement, and optimization guidance as AI Mode grows.