Search engines have been using keywords to figure out what we want for years. When you type “Paris” into Google , it has to figure out whether you want a hotel, a flight, or a history lesson. But a new research paper from Google shows that this logic could change massively soon. Instead of just looking at your words, Google is developing an on-device AI intent prediction system to understand your “intent” by analyzing how you interact with your device—all while keeping your data strictly private.
The research introduces a sophisticated method to extract user goals from “trajectories.” These are the sequences of actions you take on a website or app. Basically, the AI can build a map of your journey by looking at what you see on the screen (observations) and what you do (actions). This allows it to distinguish whether you are choosing a product because of its features or simply because it’s the cheapest option.
Google’s new two-stage method for predicting user intent
What makes this research stand out is how Google solved a complex problem using small, on-device models. Data centers usually need a lot of power to understand what makes people tick. However, Google’s researchers found a better way by splitting the task into two distinct stages:
First is the Interaction Summary. This is a small model that looks at each individual step—like a screenshot and a button click—and creates a brief summary of that specific moment. Interestingly, the researchers found that allowing the model to “speculate” about what’s happening and then removing those guesses leads to a much more accurate result.
Then there is the Final Intent. This second model takes all those individual summaries and combines them to describe the overall goal. According to the research, this two-step process actually outperformed much larger AI models running in massive data centers.
Privacy first: AI that stays on your phone
One of the most important takeaways from this study is the focus on local processing. Your browser or smartphone can run these “small” models without problems. This means the analysis of your behavior never leaves your device and is never sent back to Google’s servers.
Processing everything locally allows Google to address a major ethical hurdle. It allows for a personalized, proactive AI assistant that understands your needs without turning your digital life into a surveillance feed.
Why this matters for the future
This technology isn’t officially part of Google Search yet. However, the implications are massive. We are moving toward a world of “autonomous agents”—AI that doesn’t just answer questions but helps you complete tasks. To do that safely, the AI needs to be “faithful” to what you actually did and “comprehensive” enough to know your next move.
As Google researchers noted , evaluating human intent is notoriously difficult because our motivations are subjective. Even humans only agree with user intent about 80% of the time. Google is currently laying the groundwork for AI that anticipates what you need before you even finish scrolling. Meanwhile, the company continues working to perfect this new extraction method.
