Google Engineer Explains ‘Black Box’ AI Models In Search via @sejournal, @MattGSouthern

Nikola Todorovic, Director of Software Engineering at Google Search, appeared on an episode of Search Off the Record to discuss how AI evolved inside Google Search.

Todorovic leads Google’s SafeSearch engineering team and has worked in the search organization for 15 years. He said machine learning was difficult to deploy broadly across Search because complex models are harder to understand and fix than simpler systems.

He was explaining why Google could not simply apply ML systems across Search at once. Todorovic said these models can “function like a kind of a black box” because engineers don’t always understand what happens underneath.

That makes debugging harder when search systems change over time or when a model needs to be replaced, he said.

SafeSearch As Proving Ground

Todorovic said SafeSearch was one of the first places where Google could deploy AI models in Search because the team could isolate those systems from the main ranking flow.

SafeSearch could run standalone image and video classifiers that produced a signal, such as how explicit a result might be. If problems came up, engineers could iterate on the model without disrupting the rest of Search.

Convolutional neural networks began improving image understanding about 12 years ago, he said, making SafeSearch a natural early use case for machine learning inside Search.

AI Overviews Built On Existing Search

Todorovic described AI Overviews as a feature that “stamps on top” of Google’s existing retrieval and ranking systems. He said the retrieval and ranking underneath AI Overviews is still what he called “the old style, the old school.”

The process can involve fan-out queries, he said. Google may identify additional queries related to the original input, run them in parallel, and bring the retrieved results back into one response.

AI Overviews then 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 operates with more independence, Todorovic said. He described it as still running on Search, while having a “bigger platform for its own.”

Why This Matters

The “black box” quote is getting attention, but the full context matters. Todorovic was explaining why machine learning was difficult to deploy broadly across Search, not saying Google lacks oversight of AI Overviews or AI Mode.

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

The useful point is not that AI is a “black box.” His comments reinforce that traditional Search systems still matter for AI Overviews, even as Google layers summarization and fan-out on top.

That keeps traditional Search fundamentals relevant to AI features, even as Google changes how results are summarized and presented.

Looking Ahead

The difference between AI Overviews and AI Mode is worth watching 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 infrastructure.

That difference may matter for how Google explains visibility, measurement, and optimization guidance as AI Mode expands.