How to Check Whether Your Smartphone’s AI Runs on the Device or in the Cloud

“AI” on a phone’s feature list does not tell you where processing happens. A feature may analyse part of a request on the handset, send a more demanding task to a company’s servers, and then finish the result locally.

The most dependable check is feature-specific: identify the exact tool, read the manufacturer’s documentation, and look for an on-device control or official activity log. Airplane-mode testing and network monitoring can provide useful evidence, but neither proves that every stage of an AI feature runs locally.

Quick comparison

Verification method	What it can establish	Main limitation
Manufacturer documentation	Whether a named feature or model is designed for on-device or cloud processing	Behaviour can vary by model, software version, region and language
On-device processing setting	Whether supported AI functions can be restricted to local processing	Some features may stop working or lose capabilities
Activity or network log	Whether requests are recorded as being sent to a cloud service	The log normally covers only a specific system component
Airplane-mode test	Whether the tested operation can produce a result without an active connection	Caching and hybrid processing can give misleading results
Network monitoring	Whether an app communicates with remote servers during use	Encrypted, delayed or unrelated traffic can be difficult to interpret

Why processing location matters

On-device AI processes prompts, images, recordings or text on the phone. For supported tasks, this can reduce the amount of data sent away from the handset, work without an internet connection, reduce network use and avoid a server round trip. Google says Gemini Nano runs through Android’s AICore system service and supports local processing without sending data to the cloud for supported use cases.

Cloud AI sends at least some information to remote computers. Cloud models can provide more computing power and handle longer documents, larger inputs and more complex tasks. Android’s developer guidance distinguishes Gemini Nano from cloud models such as Gemini Pro and Gemini Flash, which run inference in Google’s data centres.

Neither approach is best for every task. Local models may be smaller and constrained by the phone’s memory, battery, storage and processing capacity. Cloud processing may be more capable, but it requires connectivity and creates an additional data-handling boundary.

Start with the exact feature

Do not begin with the phone’s general AI branding. “Galaxy AI”, “Pixel AI” and “Gemini” can refer to several functions with different models and processing paths.

Write down the feature you want to check, such as a call summary, photo editor, translation tool, writing assistant or voice recorder. Open its help, privacy or settings page and look for terms such as “on-device”, “cloud processing”, “online processing”, “network connection” or a model name such as Gemini Nano.

A specific model name is useful evidence. Google identifies Gemini Nano as an on-device model accessed through AICore. AICore does not have direct internet access, although model downloads and updates are routed through Private Compute Services. Simply seeing the word “Gemini” does not establish that a particular feature is local.

Samsung Galaxy: use the device-only control

On compatible Galaxy phones, open Settings > Galaxy AI and enable Process data only on device. On some software versions, the option may appear under Settings > Advanced features > Advanced intelligence.

Samsung says most translations and Gallery Edit Suggestions are handled on the device. It lists translation, time and weather Ambient wallpapers, and certain Google Messages functions as on-device examples. Auto-formatting, auto-summarising, auto-correction and image editing are listed as cloud-based functions.

The setting does not make every Galaxy AI feature local. Samsung warns that restricting cloud processing limits some functions, and available features can vary by device and software version. Check the menu again after a major One UI update.

iPhone: inspect the Apple Intelligence Report

Apple describes Apple Intelligence as “on-device first”. The iPhone evaluates whether a request can run locally; more complex requests may be sent to Private Cloud Compute, Apple’s server-based system using larger models.

To check for those cloud requests, open Settings > Privacy & Security > Apple Intelligence Report. Select a period covering the last 15 minutes or the last seven days, then export the activity. The report is saved as Apple_Intelligence_Report.json.

An empty report is not proof that every tested feature was fully local. It may simply mean that no Private Cloud Compute request was made during the selected period. The report identifies requests sent to Private Cloud Compute; it does not explain every network connection made by the phone.

Apple says Private Cloud Compute uses request data only to fulfil the task and removes it after the response is returned. Apple has also published technical documentation, selected source code, transparency information and a Virtual Research Environment for researchers to inspect aspects of the system. These materials concern Apple’s published cloud architecture and do not show that every Apple Intelligence feature always runs on the iPhone.

Android: check the Android System Intelligence log

Google provides a network-usage log for Android System Intelligence. Its features run inside Private Compute Core, which does not have direct network access. Private Compute Services acts as a controlled bridge for features that need cloud communication, and relevant requests are recorded in the Network Usage Log.

To view it, enable Developer Options by tapping Build Number seven to ten times. Then go to Settings > Security & Privacy > More privacy settings > Android System Intelligence > Network Usage log.

This is not a complete record of every AI operation on an Android phone. It applies to Android System Intelligence and may not cover a third-party AI app, a manufacturer’s separate service or the Gemini app.

Use airplane mode as a supporting test

Force-close the app or feature.
Turn on airplane mode.
Start a new request using a fresh photo, recording or prompt.
Restart the phone and repeat the test if the result could have been cached.
Compare the result after restoring network access.

If the feature produces a result offline, the tested operation can run without an active connection at that time. That is useful evidence for on-device processing. Gemini Nano is designed for supported offline use cases, and Android documentation says local inference eliminates the server call for the inference request.

Offline success still does not prove that the complete feature is local. The model may have been downloaded earlier, a result may have been cached, or only one stage of a hybrid process may work offline. A feature may also require the internet for account checks, online search, remote content or synchronisation even when its core model runs on the phone.

Conversely, failure in airplane mode does not automatically prove that the AI model is in the cloud. The feature may need an online service for a supporting function rather than for the actual inference.

What network monitoring can reveal

Per-app data usage, a firewall, a VPN-based monitor or a router log can show whether an app communicates with remote servers after you submit a prompt, image or recording. A repeatable upload is evidence that some data leaves the phone.

It is harder to prove the opposite. No visible traffic does not prove that processing is local: traffic may be encrypted, delayed, batched or sent through another system component. Background synchronisation, telemetry, advertising, account checks and model downloads may also create traffic unrelated to the AI request.

Use careful wording in your notes: “The test suggests that the feature communicates with a server.” Unless official documentation identifies the remote processing, avoid claiming that network testing alone proves that the model is cloud-based.

Hybrid AI is common

A single feature can include local preprocessing, cloud inference, cloud data retrieval and local post-processing. Google’s Pixel Studio, for example, combines an on-device diffusion model with the cloud-based Imagen model.

Apple describes a process in which simpler Apple Intelligence requests are handled on the device and more complex ones can be sent to Private Cloud Compute. Samsung also documents a mixture of local and cloud-based Galaxy AI functions.

This is why a fast response, an NPU or AI accelerator, or successful operation in airplane mode cannot prove that every AI feature on a phone is local. Such hardware supports local machine-learning tasks, but it does not determine how every app is configured.

A practical verification workflow

Name the exact feature. Do not test the phone’s AI branding as a whole.
Read its documentation. Search for “on-device”, “cloud”, “online processing”, “network connection” and the model name.
Look for a system control or official log. Samsung provides a device-only processing switch; Apple provides the Apple Intelligence Report; Android provides the Android System Intelligence Network Usage Log.
Run a clean airplane-mode test. Use a new input after force-closing the app.
Check network activity. Look for repeatable uploads while using the feature.
Repeat with another input. Restart the phone to reduce the chance of a cached result.
Keep the conclusion feature-specific. Do not generalise from one AI tool to the entire phone.

What Bangladeshi users should consider

For users managing mobile-data costs or dealing with inconsistent connectivity, offline testing is practical. A feature that continues to work after a clean airplane-mode test may be useful when Wi-Fi or mobile data is unavailable, although the result should still be checked against official documentation.

Privacy-sensitive users should prefer features with explicit on-device documentation or a setting that restricts cloud processing. On Samsung phones, enabling the device-only control may reduce functionality. On Apple devices, the Apple Intelligence Report can show whether requests have been sent to Private Cloud Compute.

Availability depends on the phone model, software version, language and region. The existence of a Bangladesh support page should not be taken as confirmation that every Apple Intelligence feature is available in Bangladesh.

Frequently asked questions

Does an NPU prove that AI runs on the phone?

No. An NPU or AI accelerator shows that the handset has hardware suitable for local machine-learning tasks. It does not prove that a particular feature uses that hardware instead of a cloud model.

Does airplane-mode operation prove that a feature is fully on-device?

No. It shows that the tested operation can produce a result without an active connection at that time. Cached results, downloaded models and hybrid workflows can make the result misleading.

Is an internet requirement proof that the AI model is in the cloud?

Not always. The model may run locally while the feature uses the internet for search, authentication, remote data, synchronisation or a model update.

Is the Gemini app the same as Gemini Nano?

No. Gemini Nano is a device-resident model managed through Android AICore for supported tasks. The Gemini app can use broader Gemini services and connected features, many of which may require remote processing. Check each function separately.

Can cloud processing be disabled?

On supported Samsung Galaxy phones, enable Process data only on device. Some functions may then be limited. Other phones may not provide an equivalent master control, so documentation and activity logs become more important.

Conclusion

The dependable answer is feature-specific, not phone-wide. First read the manufacturer’s documentation, then look for a device-only control, an official activity log or a named on-device model. Use airplane mode and network monitoring as supporting tests, not as final proof.

For Samsung users, the device-only processing switch is the clearest practical check. iPhone users can inspect the Apple Intelligence Report for Private Cloud Compute requests, while Android users can consult the Android System Intelligence Network Usage Log. These tools provide stronger evidence than branding, chip names or response speed, but none explains every AI operation on a phone.

Search This Blog

TechBD Daily