Apple is preparing to launch a new opt-in system that will use your on-device data to improve Apple Intelligence, the company’s AI platform.
This approach relies on Differential Privacy, a method designed to collect data without revealing personal details. While Apple says user data never leaves your device, the system still analyzes it to refine how Apple Intelligence responds to common prompts.
The new feature, set to debut in iOS 18.5, will gather analytics from users who choose to participate. It uses techniques that resemble Apple’s scrapped CSAM (Child Sexual Abuse Material) detection system—an earlier attempt at on-device analysis—but this time, Apple emphasizes that no actual user content is uploaded or reviewed.
How Apple collects data without taking it
Differential Privacy works by injecting random noise into the data collected. This prevents Apple from identifying individual users while still allowing pattern recognition at scale.
For example, Apple may submit a popular prompt like “dinosaur in a cowboy hat” and check how often similar fragments appear in your usage data. Results from these checks are anonymized and fragmented across devices.
Apple says the system will improve features such as Genmoji, Writing Tools, and Image Playground by identifying frequently used inputs and refining responses. Long-form text generation, such as email summaries, will use a more complex method.
Apple generates synthetic content and compares it to anonymized data samples to find patterns—without actually accessing your emails.
The process doesn’t involve uploading or storing personal messages. Instead, your device compares locally generated samples with synthetic inputs. The results help Apple tune its models while keeping raw content private. According to Apple, all matching occurs on-device, and no personal identifiers are linked to the analysis.
Opting out stays in your hands
Despite privacy safeguards, some users may be uncomfortable sharing any data—even anonymized—to train AI. Apple acknowledges this by making the system strictly opt-in. You can turn it off in Settings > Privacy & Security > Analytics & Improvements by toggling off “Share iPhone & Watch Analytics.”
The techniques may remind users of Apple’s abandoned CSAM detection effort, which used a different system to scan photos for harmful content. That initiative faced heavy backlash over concerns that it could be misused by governments or exploited by attackers.
In contrast, Apple Intelligence training focuses on anonymous prompt data and avoids content scanning entirely.
As previously reported by Bloomberg, Apple aims to improve AI model accuracy by comparing synthetic data against local real-world patterns, all without pulling personal information. The company confirmed that the system will be included in upcoming beta versions of iOS, iPadOS, and macOS.
Apple’s AI group has seen leadership changes and product delays in recent months, and this shift toward on-device training reflects a broader push to catch up with competitors like OpenAI and Google—without compromising on privacy.
The company plans to announce more Apple Intelligence upgrades in June. For now, Apple insists your data stays on your device, and the choice to share it remains yours.