SafeTime AIAI that helps people disconnect when they choose to.
SafeTime is exploring whether AI-enabled mobile devices can help people establish digital boundaries before moments of distraction, temptation, vulnerability, or compulsive internet use occur.
“Consider extending SafeTime tonight — this window has been protected on several recent nights.”
Why am I seeing this recommendation?
SafeTime AI is a public-interest research initiative exploring how privacy-preserving AI, behavioral science, and mobile operating-system design can give people greater agency over their digital connectivity through voluntary pre-commitment.
Connectivity Is Effortless. Disconnection Isn't.
Smartphones have become extraordinarily good at keeping us connected. SafeTime asks a different question: can technology also become better at helping us disconnect when that is what we genuinely want?
Modern smartphones provide constant access to the internet. For most people this connectivity is useful, but some users experience periods when immediate impulses conflict with decisions or goals they made earlier.
Examples people themselves report include:
- Late-night internet use
- Unwanted pornography consumption
- Doomscrolling
- Social media overuse
- Online gambling
- Excessive gaming
- Study distraction
- Sleep disruption
- Repetitive browsing
- Other self-identified unwanted internet behaviors
The intention-behavior gap
A person may genuinely intend during the day to avoid a behavior later at night.
When that moment arrives, fatigue, stress, boredom, loneliness, environmental cues, or immediate reward may influence behavior differently.
Existing technology often expects the user to exercise willpower precisely when that willpower may be under the greatest pressure.
SafeTime does not assume that all pornography use is harmful. Our initial research includes people who themselves identify certain forms of internet use as unwanted, compulsive, inconsistent with their values or goals, or difficult to control.
Why Today's Digital Controls Are Not Enough
Each existing category solves part of the problem. Most share the same structural limitation: they ask the user to hold the line at the hardest moment.
Purpose
Block categories of websites.
Limitation
May depend on blacklists, filtering rules, browser configuration, permissions, or administrator settings.
Purpose
Restrict access through a particular browser.
Limitation
May not protect other browsers or applications.
Purpose
Reduce ads and pop-ups.
Limitation
Not designed to stop intentional access to unwanted websites.
Purpose
Restrict websites or apps.
Limitation
May rely on permissions available to third-party applications and may be removed or reconfigured depending on OS controls.
The missing layer may not be another blacklist. It may be a system for voluntary digital pre-commitment.
The SafeTime Difference: Decide Before the Vulnerable Moment.
At 6:00 PM a person may decide: “I don't want unrestricted internet access between 11 PM and 5 AM.” SafeTime allows that earlier decision to remain active during the later period.
- Decide
- Schedule
- SafeTime Activates
- Protected Period
- Connectivity Restored
SafeTime is not intended to punish users. It gives people a mechanism to voluntarily establish boundaries ahead of time — with emergency and essential access preserved throughout.
AI assists. Humans decide.
Basic SafeTime can function without AI through manually scheduled protection. AI becomes useful because static schedules may not reflect changing behavior.
- AI may help personalize SafeTime while preserving user choice.
- SafeTime AI should not automatically judge content or classify people as addicted.
- Every recommendation must be explainable and refusable.
“Your SafeTime sessions are most frequently activated between 11:30 PM and 1:30 AM. Would you like to make this a recurring protected period?”
Why am I seeing this recommendation? Because SafeTime was activated in this window on several recent nights. Explanations are always available, and recommendations can be turned off.
Protective AI, Not Surveillance AI.
Whenever technically possible, raw activity data stays on the device, browsing content is never uploaded, and learned patterns can be deleted by the user.
- 01Device Activity Signals
- 02On-Device Pattern Analysis
- 03SafePredict
- 04User Recommendation
- 05User Decision
- 06SafeTime Schedule
Data minimization
The smallest useful signal set — timing patterns, not content.
User inspection
Users can see which signals are collected, which patterns are inferred, and delete both.
Optional intelligence
AI features can be disabled entirely; manual schedules keep working.
“Individuals may be better able to regulate unwanted digital behaviors when they can establish protective boundaries before entering periods of vulnerability rather than being required to exercise the same level of self-control during those periods.”
SafeTime is investigating whether privacy-preserving AI and operating-system-level controls can make these boundaries adaptive, personalized, and resistant to impulsive changes while preserving autonomy, emergency access, privacy, and user safety.
“How can AI-enabled mobile devices help people exercise meaningful control over their own connectivity during periods when their immediate impulses may conflict with their previously stated intentions?”
Six workstreams
Each workstream produces evidence the next one depends on. Later phases proceed only if earlier findings justify them.
01 · Digital Vulnerability
Understand when and why individuals experience difficulty maintaining intended digital boundaries.
02 · Existing Digital Controls
Study blockers, filters, screen-time systems, parental controls, and other interventions.
03 · Behavioral Pre-Commitment
Study whether decisions made before vulnerability are more durable than decisions made during it.
04 · AI Personalization
Explore how privacy-preserving AI can recommend useful boundaries without invasive data collection.
05 · Ethics & Human Agency
Study consent, autonomy, coercion, privacy, accessibility, bias, and potential misuse.
06 · Responsible OS Integration
Study technical requirements for implementing voluntary connectivity controls at the mobile OS level.
Grounded in published work — and honest about its limits.
We organize existing literature into six themes and record each source with its key finding and its limitations.
Always-On Connectivity
Smartphones have made internet access immediate, continuous, mobile, private, and highly personalized. Understanding this baseline matters before proposing any intervention.
Problematic Internet Behaviors
Existing research examines problematic internet use, problematic pornography use, social media overuse, gaming, gambling, sleep disruption, and compulsive digital behavior. Prevalence estimates and definitions vary substantially between studies.
The Intention-Behavior Gap
Behavioral research explores how people form long-term intentions yet make different decisions when exposed to immediate rewards or high-risk contexts.
Pre-Commitment
People may voluntarily restrict future options so that later behavior stays consistent with earlier intentions. Examples outside SafeTime include savings commitments, gambling limits, study schedules, productivity tools, and device limits.
Digital Wellbeing
Research examining screen time, digital self-control, digital wellbeing tools, and intentional technology use.
AI and Human Agency
An emerging question: can AI support users in achieving goals they themselves choose, instead of primarily optimizing engagement?
SafeTime distinguishes between established research findings and hypotheses currently under investigation. References presented on this website should not be interpreted as proof that SafeTime itself is effective. The SafeTime intervention requires independent evaluation.
SafeTime Digital Agency Study 2026
Understanding how individuals currently manage unwanted online behaviors, and whether voluntary pre-commitment could improve digital agency.
- Internet habits
- Unwanted digital behavior
- Times of vulnerability
- Existing blocker use
- Why blockers are disabled
- User attitudes toward pre-commitment
- Preferred restriction levels
- Cooldown preferences
- Trust in AI
- Privacy concerns
- Emergency exceptions
- Attitudes toward OS integration
A research prototype, clearly labelled as one.
The prototype makes the research question tangible: scheduling, protection modes, cool-off delays, essential access, and AI-suggested periods.
The current SafeTime interface is a research prototype. True device-wide connectivity enforcement would require deeper operating-system integration.
“Consider extending SafeTime tonight — this window has been protected on several recent nights.”
Why am I seeing this recommendation?
SafeTime Principles
AI should strengthen users' ability to make decisions.
Collect only what is necessary, and keep it on-device where possible.
Restrictions should reflect voluntary prior choices.
Do not classify people morally or score their character.
Recommendations should always have understandable reasons.
AI recommends; humans decide.
Emergency access remains possible at all times.
Effectiveness should be independently evaluated.
The smallest useful signal set, and nothing more.
People affected help design the tools intended to help them.
External researchers should test and challenge our claims.
Deeper enforcement demands stronger safeguards, not fewer.
From always-on connectivity to digital agency
- 01Always-On Connectivity
- 02Periods of Vulnerability
- 03Intention-Behavior Gap
- 04Voluntary Pre-Commitment
- 05AI-Assisted Personalization
- 06User-Selected Boundary
- 07Possible OS-Level Enforcement
- 08Greater Digital Agency
- 09More Intentional Technology Use
“If individuals can define protective boundaries before vulnerable moments, and AI can help personalize those boundaries without undermining privacy or autonomy, mobile technology may better support intentional digital behavior.”
Seven phases
We are looking for collaborators.
AI researchers, behavioral scientists, digital wellbeing researchers, privacy and ethics experts, universities, civil society, mobile engineers, and smartphone manufacturers.
Why this may eventually need the operating system
Third-party apps operate within permissions granted by mobile operating systems. A deeper implementation could enable more reliable, consistent, and safer boundaries.
SafeTime has not yet established the technical, ethical, or regulatory feasibility of full OS integration. This is a research and partnership objective.
AI for Human Agency
Much of modern AI is used to increase engagement, predict what people will click, and keep users interacting with digital platforms. SafeTime explores another possibility: can AI help people protect decisions they have already made for themselves?
SafeTime investigates AI as infrastructure for agency rather than attention capture.
Agency
Strengthening choices people have already made.
Privacy
On-device inference and minimal signals.
Ethics
Consent, explainability, anti-coercion.
Platforms
Responsible mobile OS collaboration.
What if our devices became better at respecting the boundaries we set for ourselves?
SafeTime AI is exploring a future where artificial intelligence does not simply learn how to capture our attention — it also learns how to help us protect it.
SafeTime AI — Technology that respects the decisions we make for our future selves.