Personalization controls

Personalization controls are the settings and inline actions that let people manage how an AI adapts to them. They include viewing and deleting stored memories or preferences, turning personalization off, pausing history collection, excluding topics, and resetting the model to its default behavior. They are the user-facing side of any system that learns from behavior.

Why it matters

Personalization changes what people see without them asking, which can feel helpful or invasive depending on whether they understand and control it. Microsoft's HAX guidelines pair "Learn from user behavior" (G13) with "Provide global controls" (G17): "Allow the user to globally customize what the AI system monitors and how it behaves." Without the second, the first erodes trust.

Preferences also change. Google's PAIR guidebook says a product "should allow for people to erase or update their previous selections, or reset your ML model to the default, non-personalized version."

How to apply it

• Do show what the system has learned in plain language. A chat assistant with memory should list stored facts ("You prefer metric units") with a delete button on each.
• Do offer a global off switch and a reset, not only item-by-item edits.
• Do ask before using personal data. Apple's generative AI guidelines say to ask permission before using personal information and usage data and to "always offer a clear way to opt out of its use."
• Do explain why a result is personalized. Apple's machine learning guidelines recommend factual attributions such as "Because you've read nonfiction" rather than "Because you love nonfiction."
• Do let people correct the inputs, not only the outputs. If a music app thinks someone loves children's songs because their kid used the account, they need a way to say so.
• Don't surface sensitive inferences. Apple advises: "Consider withholding private or sensitive suggestions," because accounts and devices are often shared.
• Don't bury controls several levels deep in account settings. Put a shortcut next to the personalized content itself.

Common mistakes

• Controls that don't control. Toggling personalization off but still seeing the same tailored feed breaks trust faster than having no toggle.
• Silent cross-context learning. Apple notes people may be surprised when actions in one app affect another, and recommends telling people how information is shared and letting them restrict it.
• Over-specific attributions. Apple warns that overly specific attributions can make people feel the app is watching too closely.
• No sense of time. Old behavior keeps shaping results long after tastes changed. Apple suggests prioritizing recent feedback.