On January 10,2025, a Chinese developer launched DeepSeek, an AI app quickly gaining popularity. In just a few days, it dethroned ChatGPT on the Google Play Store and Apple’s App Store, becoming the top AI app on the platform.
However, alongside its rise, privacy experts have flagged some serious concerns. They’re worried about how much data DeepSeek gathers and whether it shares that data with servers in China. Beyond these privacy issues, there are questions about how the app handles alignment and censorship.
This post is the first of a two-part series. In this post, we will dig into DeepSeek’s Android app privacy posture. In part two, we will compare it with three other well-known AI apps: ChatGPT, Claude, and Perplexity. We’ll use Privado’s Mobile App Scanning product to analyze each app (APK or xAPK) to determine which SDKs they use, which permissions they require, and where user data actually goes during normal app usage.
Our main goal is to look beyond the privacy policy and discover what’s happening under the hood. From experience, we know that privacy policies can be worded too generally, covering almost anything that might happen in the future, or too narrowly, omitting actual data sharing that occurs. To avoid any guesswork, we rely on real-world evidence rather than policy statements alone.
How We Rate a Mobile App’s Privacy Posture
We’ll measure each app’s privacy posture based on the following criteria:
Permissions Which permissions does the app request (e.g., location, microphone, camera), and what does that let the app do?
SDKs Which third-party software kits (e.g., analytics or advertising SDKs) are built into the app, and what data do they typically collect?
Data Collected What actual user or device data is the app sending—whether it’s personal info, usage patterns, or location data?
Third Parties Which other entities (e.g., ad networks, analytics providers) receive user data, as seen in the app’s network traffic?
Cross-Border Flows Does the app send data to countries outside the user’s home region? (This matters for compliance and regulatory concerns.)
Privacy Policy Disclosure Mismatch Does the privacy policy match what we observe in real life? Are there any gaps or surprises between the policy and actual data practices?
DeepSeek App
DeepSeek’s privacy policy is broadly written and covers all possible data collection, including sensitive data types like keystrokes. In reality, we found that DeepSeek actually collects less data than is declared in their privacy policy; however, there are clear data flows to China.
The app asks for eight permissions, including a sensitive permission for Camera. It collects data like unique IDs (iid, aid), device details (model, OS), location (time zone), language, and user inputs like prompts or chat history. This data is shared with big players like Google (US) and ByteDance (China). The app also uses SDKs from Google, Tencent, and ByteDance for authentication, analytics, and marketing purposes
Permissions
Total permissions: 8
Sensitive permissions: 1
Data Types
Permissions
Sensitive Permissions
Description
Photos and Videos
android.permission.CAMERA
android.permission.CAMERA
Allows the app to capture photos or videos using the device’s camera.
Device or Other IDs
com.asus.msa.SupplementaryDID.ACCESS
(None)
Potentially grants access to unique device identifiers on certain ASUS devices.
These permissions allow the app to run tasks in the foreground, check or use network connections, and verify licenses. While they do not directly collect personal data, they can facilitate background processes and data transfers.
SDKs
DeepSeek app embeds SDKs from Google, Tencent, Bytedance and transfers data to third parties based out of China.
SDK
Third Party
Location
Purpose
Description (Lawyer-Friendly)
com.deepseek.chat.wxapi.WXEntryActivity
WeChat (Tencent)
China
Authentication
Integrates WeChat social login. Not available while testing from the USA.
Google Play Core Library. Typically handles in-app updates and user prompts. Minimal direct user data collection.
Data Collected
Deep Seek collects unique identifiers, device identifiers, device details(model name, OS), user input - prompts, text, audio, image data which is transferred to China both internally and to third parties.
User Input - prompt, text, audio, file upload, image
prompt, file-id, chat-session-id
deepseek.com (China)
Third Parties
DeepSeek integrates with 4 third parties including volces.com(Volcengine by Bytedance) and fengkongcloud.com(by ishumei.com) where data is shared to China.
Third Party
Location
Data Type Shared
volces.com - Volcengine by Bytedance
China
Unique Identifiers (Device ID), Device Information
gvt2.com
US
Device Information
googleapis.com
US
Unique Identifiers, Device Information, Language, Location - State
fengkongcloud.com
China
Unique Identifiers, Device Information
In the network traffic analysis of the DeepSeek app, you can see some profiling related to calls:
Disclosed broadly under "Automatically Collected Information" as unique device identifiers, user IDs, and related data.
Aligned: The policy broadly covers unique identifiers.
Device Information
device_info (e.g., model, OS version)
Disclosed broadly under "Technical Information" as device model, operating system, and related technical details.
Aligned: Policy covers device information comprehensively, matching our tests.
Network Data
IP address, request headers (e.g., Content-Type, Authorization, User-Agent), and API calls.
Disclosed broadly under "Technical Information".
Aligned.
Keystroke Data
Not Found in Evidence
Disclosed under "Technical Information" as keystroke patterns or rhythms.
Overdisclosed: Keystroke data was not observed in our tests but is included in the privacy policy, making it broader than the actual data collected.
Location Data
Approximate location (e.g., America/Los_Angeles)
Disclosed broadly under "Technical Information" and "Permissions" as general location and system language.
Aligned: General location data is covered, but the policy is broader, suggesting more granular location data could be collected (e.g., GPS or geolocation).
User Input
Chat prompts, uploaded files, chat history
Disclosed under "User Input" as content provided by users, including text, files, and history.
Aligned: Policy matches evidence provided.
Error and Debug Data
Debugging details such as status_code (429) and server_ip (192.178.130.102).
Disclosed broadly under "Service-Related Data" as diagnostic and performance information, including crash reports.
Partially Aligned: Policy broadly covers debugging information, but specific data like status_code and server_ip are not explicitly mentioned.
Application Metadata
sdk_version, os_api, channel, version_code
Disclosed under "Technical Information" as app and device metadata.
Aligned: Policy matches evidence provided.
Payment Data
Not Found in Evidence
Disclosed under "Payment Information" as order and transaction history for paid services.
Not tested.
Advertising Data
Not Found in Evidence
Disclosed under "Advertising and Analytics Partners" as advertising identifiers, hashed email addresses, and cookies.
Overdisclosed: Advertising data is disclosed but not present in our tests.
The main privacy concerns with DeepSeek are data flows to China - both to their own servers and third-party SDKs integrated into the application. At this point, the privacy policy is broadly written, and the concerns around keystrokes were not found in our tests. Having said that, as more features are added, there is a possibility that excessive data collection can happen in the future.
Methodology
Privado's Mobile App Scanning product simulates a user’s journey on the app from multiple locations. To test DeepSeek, all tests were conducted from California, and the latest app on Playstore as of Jan 27, 2025, was installed on the phone. The analysis was done purely for research purposes to uncover privacy practices of DeepSeek app.
Posted by
Vaibhav Antil
in
Privacy Engineering
on
January 29, 2025
Vaibhav is the founder of privado.ai and a CIPM certified privacy professional.
.jpg)