Dominic Jainy has spent years building systems that merge AI, machine learning, and blockchain, and lately he’s been deeply focused on how proactive AI can support mental health with care and respect. In this conversation, he unpacks the shift from reactive chatbots to AI that takes the initiative across channels—chat, SMS, email, voice, and wearables. We explore how first-login check-ins can feel supportive rather than intrusive, what “memory” should and shouldn’t remember, and how to use calendars, goals, and gentle nudges without fostering dependency. Along the way, he offers concrete design playbooks, guardrails shaped by new laws like those in Illinois, and candid stories where a single wording tweak at 10:00 AM or the phrase “three times during this week” made all the difference.
You’ve reported that millions use generative AI for mental health. What patterns have you observed in daily usage, session length, or retention, and how do these compare to human therapy touchpoints? Please share a concrete metric trend and one user anecdote that surprised you.
We consistently see what I’d call the “anywhere, anytime” pattern—people check in on a true 24/7 basis, with spikes late at night when human therapists aren’t typically available. The scale is undeniable; it’s millions upon millions of sessions where users lean on generic LLMs for brief, focused conversations that feel lighter than a 50‑minute therapy hour. One concrete trend is that proactive prompts like “Would you like to check in on how you’re managing your stress this week?” lead to return touches that bundle life tasks—cooking eggs, fixing a car, planning a hike—alongside mental health reflections, which is markedly different from a traditional therapy cadence. A story that surprised me: a user who avoided therapy for years responded to a simple, compassionate AI nudge about feeling overwhelmed by work deadlines and, within days, initiated a deeper routine that they’d previously resisted; it wasn’t dramatic, just a gentle sequence that built trust without the pressure of an appointment clock.
You distinguish reactive from proactive AI. Can you walk us through a step-by-step example of shifting an LLM into proactive mode via custom instructions or settings, and explain the safeguards you’d add at each step to prevent overreach?
Step one is explicit mode selection in custom instructions: “Enable proactive check-ins for mental well-being, with user control to pause at any time.” The safeguard is consent-gating—make the choice overt and reversible. Step two sets the scope: allow limited contexts like stress and relaxation strategies, with a ceiling on frequency; the safeguard there is a hard limit on unsolicited messages and a visible “switch it off” control. Step three adds channel preferences—chat first, then optional SMS, email, or voice; safeguards require channel-specific opt-ins and quiet hours. Step four defines memory: only recall user-initiated mental health details and time-box them; the safeguard is a clear retention window and redaction on command. Finally, step five is tone calibration: require disclosures like “I’m not a therapist” and offer resources when risk indicators appear; the safeguard is escalation logic and a stop rule when signals suggest discomfort.
On first-ever login, the AI opened with a mental health check-in. How do you design that opening line to be helpful, not intrusive? Please share A/B test results, opt-in rates, and one story where a small wording change altered engagement.
We test for warmth, brevity, and agency. The winning pattern avoids assumptions and says something like, “If you are feeling blue or have anything that’s bothering your mental well-being, let me know.” Rather than parading numbers, we focus on signals: opt-ins rise when we give users a clear out, and drop when the opening presumes distress. An example that stuck with me: changing “I’m eager to talk through whatever is going on” to “If you’d like to, I can listen and share strategies” lifted replies that felt more self-directed; people told us the second phrasing felt like an open door rather than a tap on the shoulder.
When users return, the AI recalls “overwhelmed by work deadlines.” How should memory be scoped and timed so reminders feel supportive? Describe retention windows, redaction rules, and a real case where context decay or misremembering caused friction.
Scope memory to user-flagged topics—stressors, goals, or phrases they want the AI to remember—and give those items an expiration unless renewed. Use time-boxing so a note like “overwhelmed by work deadlines” fades unless the user says it’s still relevant. Redaction must be one-tap and immediate, and the system should remind users periodically what’s stored. We once saw friction when an old anxiety label surfaced after the user had completed a big project; the reminder felt like a time capsule they’d already buried, so we added a “do not resurface unless reconfirmed” rule and a gentle prompt asking whether that context was still true.
The AI wove anxiety about vacations into a trip plan. What criteria decide when to insert mental health content mid-conversation? Please detail the decision tree, confidence thresholds, and an anecdote where the insert backfired and how you fixed it.
The decision tree starts with relevancy: only insert if the user previously opted into proactive support and the topic directly connects to the current task—travel plans, sleep routines, or work schedules. Next is recency: if the mention is recent or renewed, consider a single-line offer like, “Do you want me to include relaxation strategies while traveling?” Finally, tone check: keep it optional and non-diagnostic. We had a case where the AI mentioned vacation anxiety during a purely logistical itinerary chat; the user felt it was eerie. Our fix was adding a “context bridge” rule—only reference past concerns if the user hints at stress in the current thread or explicitly asks for it.
You showed the AI texting mindfulness nudges. What cadence, timing windows, and quiet hours have you found reduce notification fatigue? Share delivery-to-engagement ratios, opt-out patterns, and one story where a schedule change boosted outcomes.
We anchor nudges to user-defined windows and institute quiet hours by default, mirroring the idea that support isn’t a 24/7 flood even if the AI is available around the clock. Rather than quoting ratios, I’ll say engagement rises when messages arrive at predictable times the user chooses, and falls when they’re scattershot. Opt-outs drop when each message carries a visible “pause” option. Moving a mindfulness prompt to a mid-morning slot—think of a 10:00 AM stretch break cadence—improved participation because it met users after the early scramble, not during it.
The email “checking in on relaxation goals” implies goal tracking. How do you structure goals, progress checks, and follow-ups so users see momentum? Please outline the data model, sample metrics (completion rates), and one before-and-after example.
Our data model treats a goal as a plain-language intent, a cadence, and a success marker—for example, “perform a relaxation exercise at least three times during this week.” Progress checks echo the user’s words and invite a quick yes/no or brief reflection. Completion rates improve when the AI reflects the user’s phrasing back rather than introducing new jargon. One before-and-after example: rewording a goal from “complete three sessions” to “perform a mental relaxation exercise at least three times during this week” led to more honest check-ins because it felt human and flexible, not bureaucratic.
Calendar inserts suggested a 10:00 AM stretch break. How do you personalize these without creeping users out? Walk through consent screens, default settings, and a case where calendar context (meetings, commute) changed the recommendation.
We ask for calendar access with explicit scoping—only free/busy and titles the user approves—and set defaults to off until they opt in. The consent screen previews sample inserts like “Quick Stretch Break” and clearly marks how reminders are created and removed. Personalization is rule-bound: if the user is in a long meeting or commuting, the AI adjusts the suggestion or delays it. In practice, a user with a packed morning got an offer to shift the 10:00 AM stretch to a quiet window rather than forcing a pop-up; that small respect for reality turned a potential annoyance into a useful nudge.
You mention Siri/Alexa prompts and wearables noticing a heart-rate spike. What signal thresholds and false alarm controls do you rely on? Share your escalation ladder, error rates you target, and one incident that reshaped your thresholds.
We treat device signals as hints, not diagnoses, and require user opt-in through those channels. A heart‑rate spike alone only triggers a gentle check like, “Are you alright? Want me to guide you through a quick grounding technique?” False alarm controls include requiring multiple signals or a user acknowledgement before escalating. We lowered sensitivity after a user received a grounding offer during an intense but joyful moment; the change was simply to ask for context first and wait for consent before proceeding.
You raised privacy and boundaries. What is your practical playbook for data minimization, on-device processing, and memory expiration? Please include specific retention periods, redaction workflows, and an anecdote where a boundary prevented harm.
Data minimization starts with purpose-limiting collection: store only what the user says to remember and nothing more. When possible, process on-device for quick checks; otherwise, transmit the least necessary details. We set memory to expire unless renewed and provide immediate redaction with a single command. A boundary that mattered: because we required user confirmation before sharing mental health notes across channels, a sensitive detail never appeared in email where it could have felt exposed; restraint protected trust.
You noted a new Illinois law on AI for mental health. How do you adapt features to meet such rules while staying useful? Describe your compliance checklist, geo-fencing logic, and one feature you reworked or dropped because of legal risk.
We built a compliance checklist that starts with disclosures—“I’m not a therapist”—and includes consent logs, opt-in proofs, and a record of proactive triggers. Geo-fencing uses location to adjust capabilities; in regions with stricter rules, proactive outreach narrows to user-pulled contexts. We also review channel restrictions and limit sensitive prompts when ambiguity exists. In one case, we dropped a cross-channel push that aggregated signals for proactive messaging because it risked being “over the line” under state interpretations.
Some fear dependency on proactive nudges. What design patterns promote autonomy instead? Share concrete tactics (cooldowns, graduated prompts), outcome metrics you track (self-initiated sessions), and a user story showing healthy off-ramping.
Autonomy is a feature, not an afterthought. We use cooldowns so proactive prompts taper when users engage regularly, and we graduate from suggestions to reflective questions that encourage self-direction. We watch for self-initiated sessions as a key outcome and celebrate them in-app. One user began with frequent check-ins and, over time, switched to setting their own goals; the AI acknowledged the shift and asked if they wanted to pause nudges—an off-ramp that felt earned, not imposed.
You spoke on 60 Minutes about hidden risks. Since then, what risk emerged that you didn’t predict, and how did you mitigate it? Please provide the timeline, the measurable impact, and the fix that stuck.
After the appearance last year, we saw a subtler risk: over-personalized mid-conversation inserts that felt eerie even when consented. The impact showed up as users abandoning threads after an unexpected mental health callback. We introduced a “relevance handshake,” where the AI asks permission to reconnect the dots before doing so, and added a decay rule so older labels don’t reappear without renewal. The fix that stuck was simple: “Would you like me to keep this in mind here?”—permission and timing matter more than cleverness.
You contrast generic LLMs with specialized therapy AIs. When should teams choose each path? Walk through a decision matrix (risk, cost, guardrails), include performance metrics you’ve seen, and one migration story from generic to specialized.
Generic LLMs shine for broad access, low cost, and multi-domain support—planning a hike one minute, reflecting on stress the next. Specialized therapy AIs demand tighter guardrails, clearer scope, and often narrower availability, but they provide more structured, evidence-aligned flows. Choose generic when you need breadth and a soft touch; choose specialized when the risk profile requires stricter boundaries and predictable interventions. We migrated one feature—goal tracking—into a more structured flow that echoed “three times during this week,” preserving humanity while adding consistency.
You mentioned APIs reaching across channels. How do you keep a consistent tone and memory while hopping between chat, SMS, email, and voice? Please detail your orchestration layer, conflict resolution rules, and an example of a cross-channel handoff.
Our orchestration layer centralizes consent, tone, and memory policies so each channel pulls from the same, scoped record. Conflict resolution favors the most recent user preference and the most privacy-preserving option when channels disagree. A handoff example: a chat session sets a goal, email follows up with “checking in on your progress with relaxation goals,” and SMS offers a lightweight nudge—all aligned to the same wording and opt-in status. If the user pauses in one channel, the others respect that pause automatically.
What ROI signals convince you that proactive mental health AI creates net benefit rather than noise? Share a simple formula, baseline vs. post-launch metrics, and a case study where you tied behavior change to a specific proactive feature.
We keep ROI simple: positive behavior change minus intrusion costs. Baseline usage is reactive-only conversations; post-launch, we look for sustained engagement in user-defined goals and reductions in abandoned threads after gentle proactive prompts. A concrete tie-in is the calendar integration: suggesting a Quick Stretch Break at a realistic time increased follow-through because it met users where they were. The value shows up not as a flashy number but as a consistent pattern of kept commitments.
If you were advising a new team, how would you stage a safety-first rollout of proactive features? Lay out a week-by-week plan, key checkpoints, red-team tests, and one “kill switch” you’d require before going live.
Week one is scoping: define proactive intents, channels, and disclosures, and wire the consent flow. Week two is sandbox testing with red-team prompts designed to trigger overreach, boundary crossings, and unwanted resurfacing of old labels. Week three brings limited beta with live opt-ins and a close watch on off switches, redactions, and quiet hours. The non-negotiable kill switch is a global pause that instantly halts proactive outreach across all channels and clears pending nudges.
In the vacation anxiety scenario, how do you blend coping tools into practical plans without overwhelming users? Describe the pacing model, content tiers, and one example itinerary with embedded techniques that drove measurable relief.
We pace in layers: logistics first, then a single optional coping suggestion tied to the plan, and only then deeper guidance if the user asks. Content tiers move from a brief check-in, to a simple practice, to resources. An itinerary example: after drafting the day-by-day plan, the AI asks if the user wants to include relaxation and mental stress‑reducing strategies; if yes, it adds tiny anchors around transitions rather than filling the day with exercises. Relief shows up as users sticking to the plan without skipping the fun parts.
What consent model best fits proactive outreach—single global opt-in, granular toggles, or context-based prompts? Please share adoption rates you’ve seen for each, churn differences, and a story where granular control changed user trust.
We prefer a hybrid: a global opt-in sets the frame, granular toggles shape channels and topics, and context prompts ask permission in the moment. We avoid publishing adoption numbers in favor of what we can stand behind—users engage more, and churn eases, when they feel in charge. Granular control once transformed a skeptic into an advocate when they turned off voice prompts but kept calendar suggestions; the feeling of authorship changed everything. Trust grows when “proactive” means “on your terms.”
You closed with “never let the future disturb you.” How do you translate that mindset into product decisions today? Offer a guiding principle, one trade-off you made, and the metric you watch to ensure you’re helping rather than disturbing.
The principle is to be calmly present: respond to what the user expresses now, not to every hypothetical. We traded aggressive outreach for a respectful “ask-first” posture, even if it meant fewer immediate touches. The metric we watch is user-initiated engagement after a proactive prompt—if people lean in, we’re helping; if they back away, we adjust. I like to remember that the same weapons of reason that serve us today—consent, clarity, and care—will meet tomorrow’s challenges too.
Do you have any advice for our readers?
Start small and stay humane. If you’re building, anchor every proactive feature in explicit consent, limited memory, and easy pauses. If you’re a user, pick one helpful nudge—maybe a 10:00 AM stretch—and see if it serves you; keep what helps, turn off what doesn’t. And for everyone: never let the future disturb you; test, reflect, and keep your agency front and center.