Your Phone's AI Features: What Actually Stays On-Device and What Gets Sent to the Cloud

Every major smartphone now has AI features built in. Apple has Apple Intelligence. Google has Gemini Nano on Pixel. Samsung has Galaxy AI. All three use phrases like "on-device processing" and "privacy-first AI" in their marketing — but the reality is more nuanced, and the differences between platforms matter.

Here is what actually happens to your data across each platform, based on their published technical documentation and independent analysis.

The Three Categories

Before going platform by platform, it helps to understand the categories that exist across all of them:

Fully on-device: The model runs locally on the phone's Neural Processing Unit (NPU) or GPU. No data is sent anywhere. No network connection required. This is the genuinely private option.

On-device with cloud telemetry: The inference runs locally, but metadata (model performance, crash logs, usage frequency) may be sent to the company. Your actual content is not transmitted, but usage patterns may be.

Cloud-processed: Your content — the actual text, image, or audio — is sent to the company's servers for processing. Results come back from the cloud. This is fast and capable, but your data leaves the device.

Most AI phone features are a mix of all three depending on the specific task.

Apple: Apple Intelligence

Runs fully on-device (3B parameter model):

• Notification summaries
• Smart Reply suggestions in Messages
• Writing Tools (basic rewrite/proofread for short text)
• Photo descriptions in Photos search
• Priority notifications sorting

Goes to Private Cloud Compute (Apple's servers, no data retention claimed):

• Complex writing tasks that exceed the on-device model's capability
• Longer document summarisation
• More sophisticated Siri requests

Goes to OpenAI (optional, requires explicit consent each session):

• General knowledge queries when ChatGPT integration is enabled
• Creative writing tasks beyond Apple's model capability

Apple's privacy claim: Private Cloud Compute processes requests on Apple Silicon servers that Apple says do not log or retain request content. The PCC code is auditable by security researchers via Apple's Security Research Device programme.

What to turn off if privacy is a concern: Settings → Apple Intelligence and Siri → ChatGPT (disable entirely). You can also disable Apple Intelligence Features altogether in the same menu.

Google Pixel: Gemini and Gemini Nano

Runs fully on-device (Gemini Nano via AICore):

• Call screening and live transcription
• Recorder app summarisation (Pixel 8+)
• Gboard reply suggestions
• Pixel Screenshots summarisation (Pixel 9)

Goes to Google's cloud (Gemini Pro/Ultra):

• The Gemini app (all queries unless you specifically use Gemini Nano via developer API)
• Google Assistant queries beyond simple device commands
• Summarise webpage in Chrome
• Magic Eraser and Magic Editor in Google Photos

Google's privacy note: On-device Gemini Nano processes locally. Cloud Gemini queries are subject to Google's standard data policies — by default, conversations may be reviewed by human reviewers to improve the model. You can opt out in Gemini settings.

What to turn off: Google Account → Data and Privacy → Web and App Activity → Gemini Apps Activity (pause this to stop conversation history being used for model training).

Samsung: Galaxy AI

Samsung's Galaxy AI is the least transparent of the three platforms about where processing happens, but here is what is documented:

Claimed on-device (varies by feature and device generation):

• Live Translate (S24+, on-device for common languages)
• Chat Assist basic suggestions
• Interpreter mode transcription

Goes to Google's cloud:

• Circle to Search queries
• Gemini-powered features on Galaxy

Goes to Samsung's servers:

• Note Assist (summarisation)
• Transcript Assist
• Browsing Assist
• Writing Assist for complex tasks

What to check: Settings → Advanced Features → Galaxy AI → Processing Mode. On S24+ you can switch some features between "On-device only" and "Device and cloud." When on-device only is selected, features may be slower or limited.

The Honest Summary

| Platform | On-device capability | Cloud transparency | Opt-out ease |

|---|---|---|---|

| Apple | Strong (3B model, broad task coverage) | High (PCC auditable) | Easy |

| Google Pixel | Strong for Nano tasks, falls back to cloud Gemini | Medium (standard Google policies) | Medium |

| Samsung | Varies, cloud used more broadly | Lower | Medium |

What You Should Actually Do

If you want the most private AI phone experience: iPhone with Apple Intelligence ChatGPT integration disabled. Apple's on-device model is the most capable of the three for common tasks, meaning fewer requests escalate to the cloud.

If you are a developer building privacy-sensitive apps: Gemini Nano via Android AI Edge SDK gives you direct programmatic access to on-device inference, which Apple does not currently offer third parties.

If you are on Samsung: Check your Galaxy AI Processing Mode setting. The default is not always on-device only, and many users do not realise their note summaries are being processed on Samsung's servers.

For everyone: The AI features you opt into explicitly (asking Siri to write an essay, using the Gemini app) are cloud-processed by design. The passive features (notification sorting, keyboard suggestions) are more likely to be on-device. The marketing language tends to blur this distinction.