Keep Factually independent
Whether you agree or disagree with our analysis, these conversations matter for democracy. We don't take money from political groups - even a $5 donation helps us keep it that way.
Gboard data collection policy on keystroke retention
Executive summary
Google’s public position is that Gboard “doesn’t retain or send any data about your keystrokes,” except when you explicitly submit content (for example, tap the G button to search) [1]. Independent forensic and reporting work shows Gboard keeps local caches and user-language data on devices and that certain telemetry or “usage statistics” can be sent to Google unless you disable those options [2] [3] [4].
1. What Google says: keystrokes aren’t sent as a rule
Google and reporting that quotes Google have repeatedly stated Gboard does not collect or transmit raw keystrokes to its servers during ordinary typing; the company says it only receives text you explicitly submit (such as a search) [1]. This is the baseline corporate policy cited by Wired and other outlets when explaining how Gboard’s smart-reply and prediction features are built to avoid shipping raw keystrokes off-device [1].
2. What independent analyses and reporting found: local retention and caches
Forensic investigators have found that recent Gboard versions retain typed content and related metadata in local caches and databases on the device; those caches are size-limited and periodically deleted, but they can contain user-typed words and even message contents accessible in a forensic image [2]. Macworld’s packet observations on iOS also showed that while no typing traffic was observed during entry, Gboard stores words locally (“remember[s] words you type”) and only transmits when you use features that require server access [5].
3. Telemetry, “share usage statistics,” and what that sends
Gboard includes a “Share usage statistics”/telemetry option whose default state and effects have been questioned by privacy writers; guides instruct users to disable that toggle to stop Gboard from automatically sending keyboard usage statistics and snippets of typing to Google [3]. Security- and malware-focused reporting notes disabling that option “significantly reduces” information transmitted — researchers found telemetry can include metadata like word lengths, timing, and the app where typing occurred [4].
4. Product design choices that reduce server-side models — and their limits
Google has described architectural work (e.g., federated learning and moving prediction to Android/system-level components) intended to keep sensitive data on-device while improving models without centralizing raw inputs [1]. But federated learning and on-device models do not by themselves guarantee zero telemetry: usage statistics and optional cloud features remain pathways for some data to leave devices when enabled [1] [4].
5. How permissions and “full access” affect iOS and Android behavior
Writers comparing platform behaviors point out platform permission models matter: on iOS you must grant “Allow Full Access” for cloud-backed features, and Apple warns that full access allows a keyboard to transmit typed data — yet packet-level testing still showed no typing traffic during input while features like GIF search caused network traffic [5]. On Android, additional permissions (contacts, SMS) and enabling personalization/backup features can allow broader local learning or cloud sync [6] [4].
6. Conflicting claims and contested headlines — read the fine print
Some high-profile, alarmist claims assert “every word you type” is sent to Google [7], but mainstream technical reporting and forensic studies present a more nuanced picture: local retention exists and telemetry/usage stats can transmit metadata or samples unless disabled, yet Gboard’s stated policy and network tests have not shown continuous streaming of raw keystrokes to Google during normal typing [1] [2] [5] [4]. The difference is critical: local caches exist and telemetry can be enabled — but continuous remote keystroke capture is not what multiple independent sources observed or what Google claims [1] [2] [4].
7. Practical steps and trade-offs for users
Guides and security blogs recommend disabling “Share usage statistics” and revoking nonessential permissions (contacts, SMS) to limit what leaves the device; doing so reduces telemetry and the amount of typed-data samples that could be uploaded [3] [4]. The trade-off: disabling these features may degrade personalization, next-word prediction, or cloud-synced dictionaries that improve convenience [3] [6].
8. What reporting does not (yet) settle — and what to watch for
Available sources do not mention precise retention timeframes for all cached data across Gboard versions beyond “periodically deleted” and “size-limited” descriptions, nor do they provide a verifiable, up-to-date Google transparency log showing exactly what fragments are sent when telemetry is enabled [2] [4]. Continued independent packet captures, updated forensic analysis across versions, and explicit disclosures from Google about retention windows and telemetry payloads would resolve remaining uncertainties [2] [4].
Bottom line: corporate statements, forensic work, and security reporting converge on this: Gboard stores and uses local typing data to power features, and some telemetry/usage‑stats options can send metadata or small samples to Google unless disabled — but available, cited reporting does not support the sweeping claim that Gboard continuously streams every keystroke to Google servers under default behavior [1] [2] [4] [3].