Domestic Abuse Prevention using AI and Emerging Technologies

A survivor-first blueprint that uses AI, live streaming, beacons, charity links, safe homes, and emergency services—without triggering retaliation or defaulting to knee-jerk police reporting. The north star: choice, control, and long-term empowerment for the victim.

1) Principles (non-negotiables)

  • Survivor choice & informed consent: Every feature is opt-in with clear, plain-language settings. Nothing auto-reports to police unless an imminent-harm threshold the user chose is crossed.

  • Do no harm: Design to de-escalate and avoid retaliation (stealth UX, delayed notifications, decoy screens).

  • Privacy & minimization: Process as much as possible on-device; store the minimum; end-to-end encrypt anything leaving the device.

  • Trauma-informed: No surprise alerts or shaming. Default to quiet modes, grounding tools, and human support options.

  • Human in the loop: Advocacy services (IDVAs, charities) triage first unless the user pre-selects otherwise.

2) System Overview (modular)

A. Safe App (Stealth Mode)

  • Cover app appearance (e.g., “notes” or “weather”): PIN or gesture reveals the real interface.

  • Discreet triggers:

    • Tap pattern on volume buttons, lock-screen widget disguised as a timer, or code word to a chatbot.

    • Bluetooth beacons (tiny stickers/keys) that, when pressed, silently trigger predefined flows even if the phone is across the room.

  • Grounding panel: Breathing coach, “I am safe” scripts, quick exit button that snaps back to the decoy screen.

B. Edge AI for Risk Signals (On-device)

  • Escalation classifier (opt-in): uses on-device audio features (voice stress, overlapping speech, shouting proxies) and device telemetry (sudden motion, late-night usage spikes) to suggest actions.

  • No hot mic by default. The model runs only during explicit safety sessions (e.g., user opens the app or triggers beacon).

  • Adaptive thresholds: user picks sensitivity; the system learns their patterns (not generic ones).

C. Live Streaming That Doesn’t Escalate

  • Pixelated/blur-first live stream to a pre-chosen advocate group (charity hotline, trusted friend, safe-home hub).

  • Audio-only “muffled” stream option.

  • “Shadow stream” records to an encrypted vault without showing any on-screen cues; later share is controlled by the survivor or advocate.

D. Secure Evidence Vault

  • End-to-end encrypted, auto-timestamped clips, photos of injuries, threats (screenshots), and a running incident journal.

  • Release control slider:

    • “Private only” → “Share with advocate” → “Share with police” (with or without identity).

  • Delayed release (e.g., if not canceled within X hours), to reduce in-the-moment coercion.

E. Connection Fabric

  • Charity/advocate routing: One-tap to an IDVA or charity chat; warm transfer to shelters/safe homes; silent text if voice is unsafe.

  • Safe homes network: Geofenced directory; request bed availability discreetly; travel planning that hides routes/history.

  • Emergency services (last mile): Only auto-dial if imminent-harm rule the survivor chose is crossed; otherwise, route to advocates.

F. Economic Safety & Long-Term Support

  • Coercive-control checkups: Optional budgeting assistant that flags unusual cash withdrawals, blocked access to salary, or coerced debt (all local/on-device unless user shares).

  • Micro-goals: saving pots, job search nudges, CV tools, discreet learning modules.

  • Document pack: auto-compiled “leave plan” (IDs, bank details, meds, contacts, school info) with a one-tap export to an advocate.

3) Flows that Reduce Retaliation

  • “Quiet Log” Mode: Logs incidents (time, notes, photos) but never alerts anyone unless user later elevates.

  • “Cool-Down” Safe Check: After a flag, system waits a user-set window (e.g., 30–90 mins) before nudging next steps, to avoid confrontation during peak tension.

  • Decoy Responses: If the abuser grabs the device, the app can instantly show a harmless screen (recipes, calculator).

  • Silent Handover: Instead of “we’ll post on Reddit”-style visibility moves that can provoke backlash, advocates seed neutral, helpful local resources into community channels—no identifiable details.

4) Avoiding “Cry Wolf” & Over-Reporting

  • Multi-signal thresholds: Don’t trigger big actions from a single data point. Combine user-initiated trigger + escalation cues + time-of-day + prior risk pattern.

  • Confidence badges: Show “low/medium/high” risk suggestions—not hard commands.

  • Human triage first: Default route to charities/advocates who can assess context before police involvement.

  • User-set rules: e.g., “Never auto-contact police,” “Contact police only if I say CODE RED,” or “Auto-share vault if I miss two check-ins.”

5) Partnerships & Data Governance

  • MOUs with local charities & shelters: 24/7 warm lines, triage SLAs, safe-home bed API, travel vouchers.

  • Telcos & device makers: zero-rated data for the app; “stealth notifications” that don’t ping or banner.

  • Banks/fintech (opt-in): pattern flags for financial abuse; fast-track to freeze/reopen accounts safely.

  • Governance: survivor advisory council, red-team for misuse (e.g., abusers installing the app on the victim’s phone).

  • Security: hardware keystore; denylist cloud backups; “panic wipe” that only clears the app while leaving the decoy intact.

6) Safety-By-Design UX Details

  • Install & update camouflage (e.g., “System Toolkit”).

  • Iconography neutrality: no “shield/alert” imagery.

  • No push alerts by default; notifications arrive as mundane reminders the user chooses.

  • Accessibility: large-text stealth, haptics, voice phrases that sound like normal dictation.

7) MVP → Scale

Phase 1 (MVP):

  • Stealth app + beacon triggers

  • On-device evidence vault + delayed release

  • Advocate chat/voice with blur-stream

  • Safety plan builder + check-in scheduler

Phase 2:

  • Edge AI escalation suggestions

  • Safe homes availability directory

  • Economic safety assistant

Phase 3:

  • Broader integrations (banks, telcos)

  • Community education content & survivor peer groups (moderated, anonymous)

8) Measuring Empowerment (not just “reports made”)

  • Time-to-support (user → advocate connection).

  • Repeat-incident reduction over 3–6 months (self-reported + advocate logs).

  • Successful exits: housing secured, income stabilization, legal protections (if chosen).

  • User-retained control: % of cases where the survivor—not the system—decided the escalation path.

  • Financial recovery: access to personal funds restored, debt remediation started.

9) Risk & Misuse Mitigations

  • Anti-stalkerware checks: app self-audits for MDM profiles, location sharing, unknown Bluetooth pairings, and prompts safe removal steps.

  • Dual-device protocol: option to mirror to a trusted friend device for redundancy.

  • No location sharing unless explicitly needed (e.g., during an advocate-supervised extraction).

  • Legal & ethical reviews per region; content is informational, not legal advice.

10) What not to do

  • Don’t auto-notify police on a single trigger.

  • Don’t store raw always-on audio/video in the cloud.

  • Don’t send “are you safe?” banners that can be seen by an abuser.

  • Don’t gamify disclosures or pressure survivors into action.

Domestic Abuse, ImpactFrancesca Tabor