Build a calm pause-and-check companion, not a fear-based scam detector

This blueprint draws on 18 deduplicated evidence rows from 18 sources across 11 independent hosts — a mix of official US guidance (FTC, FINRA, CFPB, FBI IC3) and peer-reviewed or emerging research (2023–2026) — plus the FBI IC3 headline figure: complainants aged 60+ reported $7.748 billion in losses in 2025, up 59% from 2024. The evidence scan is directional, not a systematic review.

The safety lifecycle: detect, pause & verify, avoid, report & recover — inside a trusted-human loop

Each stage is sized by the evidence behind it. Click a stage (or a tab below) to see its evidence cards. The trusted circle and cross-cutting principles surround every stage, because research shows warnings alone are insufficient and trusted humans are the strongest guardians.

official guidance peer-reviewed / emerging research numbers in stage boxes = evidence rows per stage

• AI help at the moment of decision nearly tripled the odds of correct scam detection (OR 2.94) and added ~9.7 seconds of useful friction in a 72-participant study — promising, not settled. asi.cs.cityu.edu.hk
• Technical skill does not protect: interviewed victims 60+ self-rated computer savviness 3.73 of 5, yet were defrauded by sophisticated tactics — the tool must carry the burden. link.springer.com
• Older victims adopt protective guardians far less than younger ones (credit-card freeze 25% vs 46%), so the Avoid stage must actively set protections up, not merely suggest them. frontiersin.org
• Help is accepted when it comes from trusted people, not strangers — victims want supportive “togetherness”, which is why the trusted circle is consent-based and surrounds the whole lifecycle. mdpi.com

Evidence-to-feature cards

Recommended MVP, in priority order

1. “Check this” intake — paste text, forward a message, upload a screenshot or URL, or enter call/payment details by hand or voice.
2. Cue-based explanation card — names the exact cue (impersonation, secrecy, urgency, unusual payment method, “move money to protect it”, remote access, credential requests), shows evidence and confidence, never claims certainty.
3. Persistent pause — a large one-tap “Pause — do not pay or share codes yet”, with independent verification through a vetted directory, never through contact details the suspicious message supplied.
4. Trusted-circle button — consent-based, revocable sharing of minimum context with a chosen family member, friend, peer, or advocate; the older adult stays in control.
5. Protective actions — block sender, contact the bank or card issuer, set transaction alerts and trusted contacts, credit-freeze guidance, temporary-hold escalation where institutions permit.
6. Guided reporting & recovery — collect evidence once into a reusable plain-language incident packet; route US users to FTC ReportFraud, FBI IC3 for cyber-enabled fraud, Adult Protective Services or local law enforcement for suspected exploitation, and the bank for transaction response; localize channels by jurisdiction in production.
7. Role-play practice — short rehearsals that teach users to name scam cues and practice explicit refusals, not passive articles or one-time quizzes.

Safety principles and evaluation plan

Safeguards

• AI analysis is advisory and interpretable, with easy escalation to a human.
• No surveillance, covert caregiver access, infantilizing language, or dark patterns; no auto-reporting without informed consent except where a regulated organization has legal duties.
• Preserve autonomy and never shame; explain why every warning appeared.
• Clear, empowering, multilingual language; minimal screen content; voice, large text, high contrast, screen readers, low dexterity supported.
• Prominent emergency mode for money just sent or accounts currently compromised.
• Red-team for false reassurance as well as false alarms; minimize sensitive-data collection.

Evaluation scorecard

• Co-design with diverse older adults (visual, hearing, motor, cognitive, language, digital-literacy differences), scam survivors, caregivers, advocates, banks, APS, law enforcement.
• Measure task success, cue-recognition accuracy, risky-action rate, time-to-pause, successful independent verification, report completion, recovery-action completion.
• Track false reassurance, false-alarm burden, comprehension, trust, autonomy, shame/distress, accessibility, and disparities by subgroup.
• Usability tests with realistic benign and malicious scenarios, then a field pilot; opt-out and deletion controls throughout.

All 18 evidence rows

Stage	Type	Evidence or guidance	Design implication	Date	Source

Show all rows

Method: 18 deduplicated evidence rows from 18 URLs across 11 independent hosts (no URL supplies more than 5.6% of rows), scanned 2023–2026; one row per source, mixing official US guidance with peer-reviewed and preprint research. Stage assignment to the lifecycle is editorial. Headline metric is the FBI IC3 2025 annual report figure for complainants 60+ ($7.748B, +59% vs 2024). Directional evidence scan, not a systematic review; single-country studies do not generalize to all seniors. Evidence text shortened for space; full wording at each source link.