AI on Android: available information does not confirm which features are processed on the device

Android’s AI page brings together several features and services, including Gemini and Circle to Search. Its presentation focuses on possible uses: getting help with tasks, interacting with an assistant, or searching for items shown on the screen. It provides an overview of features, but it is not a complete technical account of how they work. Source: Android

The distinction matters because “working on Android” is not the same as “being processed on the phone.” Android is the operating system and environment in which these features are offered; that label alone does not clarify whether computation takes place on the device, on remote servers, or through a combination of both. Nor does it identify what information leaves the phone to handle a request. The overview page is not enough to classify each feature as on-device or cloud-based.

Accordingly, the evidence located does not provide grounds to announce a recent shift to fully on-device AI, or to claim that the cited features always require a connection. The conclusion is narrower: promotional information can identify features, but cannot establish the processing route for each one. This limitation does not prove that on-device processing does not exist; it means this source does not establish it.

Gemini is an assistant experience that Android presents as helping users write, create, plan, and learn. Those capabilities describe tasks, not the architecture that performs them. The dedicated Gemini on Android page explains how to start using the assistant, but the material available here does not provide enough detail to show which operation is processed locally, which requires remote services, or what network requirements apply in each case. Source: Android

The same caution applies to Circle to Search. Being able to start a search based on what someone sees on the screen does not reveal where items are recognized, what content is transmitted, or whether the feature behaves the same way on every phone. A general page that groups features together is no substitute for a feature-specific technical description covering requirements, compatible models, and data handling.

For users, the useful verification step is to look for documentation tied to the exact feature and device. Check the connection requirement, availability conditions by model and region, and any official explanation of how shared content is handled. A claim about one feature should not be applied to all the others, even when they appear together under the Android AI label.

Google maintains a help page titled “About Android AICore.” This documentation is relevant when investigating AI-related components on Android, but the title and existence of a component do not prove that every AI feature on the platform uses it, or that a particular operation runs locally. Reaching those conclusions requires a documented link between the component, the feature, and the device model. Source: Android Help

In a rigorous assessment, AICore would be a technical lead requiring context, not a general answer about privacy. It would be necessary to know which process invokes it, which model and version are involved, what data is handled, and what happens without a connection. The evidence gathered does not complete that map or allow a comparison of behavior across different phones.

It is also important to distinguish processing, storage, and synchronization. The fact that some computation takes place on the phone does not, by itself, show that information is not stored, sent for another stage, or linked to an account. Conversely, missing detail on an overview page does not justify concluding that data is sent. These are separate questions and require separate documentation.

Android’s developer documentation presents tools and resources for building AI experiences in apps. This material is intended for software developers and is not, by itself, a statement about how each consumer service shown on Android’s page works. It is not valid to extrapolate an ability to build AI apps into a claim about the processing route used by Gemini or Circle to Search. Source: Android Developers

To assess a privacy claim, its scope must be identified: which feature, version, data, and processing stage it describes. “On-device” may refer to where one part of a model runs; it does not necessarily describe the entire journey of a request. Likewise, having a connection available does not prove that all content is sent to the cloud. Without a feature-specific official explanation, either interpretation would be an unverified inference.

If a decision depends on using a feature offline or limiting the sharing of content, the most prudent step is to consult that feature’s help information and applicable privacy details, then check the requirements for the specific model. If those documents do not answer the question, the right course is to preserve the uncertainty: do not assume complete local protection, and do not describe a data transfer as fact.

The search examined does not support saying that Android has announced a recent change turning its AI features into on-device processing, nor does it support claiming that they all depend on the cloud. The general page lists features, and the AICore documentation points to a related component, but a verifiable link is missing between each feature, the compatible device, and where its operations run.

As a result, the appropriate editorial approach is not to present the feature list as news about privacy or offline operation. It is to explain the limit: the documentation consulted does not determine how each feature is processed. Closing that gap would require dated official pages specifying connection requirements, compatibility, and data handling for each feature, as well as evidence of a recent announcement if the story is presented as news.

For people using Android, the practical rule is simple: assess each tool separately and do not infer its architecture from its name, the Gemini brand, or the presence of AI in the system. This article does not evaluate behavior in use or provide technical tests; it defines what the cited sources support and what remains undocumented in the available evidence.