Understanding Google’s ‘Black Box’ AI

Google has offered further insight into how artificial intelligence is being used within Search, with one of its senior engineers explaining why these systems can sometimes feel difficult to fully understand.

Nikola Todorovic, Director of Software Engineering at Google Search, recently discussed the challenges of integrating machine learning into Search during an episode of the Search Off the Record podcast.

He explained that advanced AI models can behave “like a black box”, meaning engineers do not always have complete visibility into how decisions are made inside the system. This lack of transparency can make it harder to diagnose issues or refine performance when changes are introduced.

According to Todorovic, this is one of the main reasons why Google could not simply roll out machine learning across all areas of Search at once. Compared to traditional systems, these models are more complex, making them harder to troubleshoot and maintain over time.

SafeSearch as an early testing ground

One of the first areas where AI was introduced successfully was SafeSearch. This worked well because it could operate separately from the core ranking system, allowing engineers to test and improve models without affecting the wider search experience.

SafeSearch uses machine learning to analyse images and videos, helping to identify and filter explicit content. By running independently, it gave Google the flexibility to refine its models more safely and gradually.

Todorovic noted that advancements in image recognition, particularly through neural networks, played a key role in making SafeSearch an effective early use case for AI within Search.

How AI Overviews are built

When it comes to newer features like AI Overviews, Todorovic explained that these are layered on top of Google’s existing search systems rather than replacing them entirely.

The underlying process still relies on traditional retrieval and ranking methods. However, AI Overviews add another layer that gathers and summarises information from multiple sources.

One technique used is known as “fan-out” queries. This involves generating additional related searches based on the user’s original query, running them simultaneously, and then combining the results into a single response.

The system then pulls together key details such as page content, snippets, and titles to create a summarised answer for users.

AI Mode and its growing role

Todorovic also touched on AI Mode, which follows a similar approach but operates with more independence compared to AI Overviews.

While still connected to the main Search system, AI Mode is being developed with a broader framework, suggesting it could evolve into a more standalone experience over time.

This distinction may become more important as Google continues to expand its AI-driven features.

What this means for Search

Although the “black box” description has drawn attention, the broader message is more practical. Todorovic was not suggesting that Google lacks control over its AI systems, but rather explaining the technical challenges involved in deploying them at scale.

His comments reinforce that traditional search infrastructure still plays a central role, even as AI features become more prominent. Ranking systems, data retrieval, and core SEO principles remain relevant behind the scenes.

Looking ahead

As Google continues to develop AI-powered features, the balance between automation and transparency will remain a key focus.

The differences between AI Overviews and AI Mode, in particular, could shape how visibility, performance, and optimisation are measured in the future.

For now, the takeaway is clear: while AI is changing how search results are presented, the underlying systems that power Search are still very much in place—just with an added layer of intelligence on top.