Google says it removed 700,000+ apps from the Play Store last year for policy violations, up 70% YoY, credits new machine learning techniques for the increase
Google today shared details of Google Play's efforts to protect Android users with its teams of engineers, policy experts …
Context & Ripple Effects
This report lands mid-arc in Google's shift of Play Store policing from human review to machine learning. Weeks earlier, the Android Security 2017 Year in Review showed 60.3% of potentially harmful apps were already caught by ML models scanning through Play Protect's daily reviews, and by December Google reported its ML-based system stripping millions of fake reviews in a single week.
The 700,000+ removals — up 70% YoY — are the first headline number where Google explicitly credits new ML techniques rather than policy changes alone, and the trajectory it starts is long-lived: submissions rejected rose another 55%+ the following year as tighter policies and broader automated protections rolled out, eventually reaching millions of apps annually in later enforcement reports.
First-order effects
- Developers shipping policy-violating apps face a much shorter window on the Play Store: with ML catching violations at submission and post-publication, bad actors lose the latency gap that manual review used to give them.
- Google's review teams get leverage — automation absorbs the volume increase so enforcement scales without proportional headcount growth.
Second-order effects
- Malicious developers adapt by shifting toward subtler violations that evade classifiers — the same pattern behind Google's later crackdowns on over a million blocked app publications and hundreds of thousands of banned developer accounts, where abuse migrated from obvious malware to policy-edge behavior.
- Legitimate developers absorb stricter false-positive risk, pushing the ecosystem toward cleaner SDK usage and more conservative app behavior to stay clear of automated flags.
Third-order effects
- App-store trust becomes an ML arms race rather than a curation problem: enforcement coverage expands until the binding constraint is classifier precision, not reviewer capacity — a structure visible in how Google's annual rejection counts kept climbing into the millions across subsequent years.
- As automated enforcement matures, platform gatekeeping power concentrates further with Google, since only operators with massive training data can police a catalog this size at this speed.
The trend: Mobile app-store governance is being rebuilt around machine-learning enforcement, with each year's removal statistics serving as the public scoreboard for how far automated policing has displaced human review.