AI relationship counselors

Sherry Turkle's clinical work, and Lori Gottlieb's columns, have both documented the same patient population: couples who would benefit enormously from skilled intervention and will never receive it. The reasons are structural. Clinicians are concentrated in cities. Insurance reimbursement for couples work is poor. The stigma, especially for men, remains stubborn. Into this vacuum the apps have walked, and the vacuum is real. A tool that gets a working-class couple in rural Ohio to pause before escalating is doing more good than a $300/hour clinician they will never see. We have to hold the critique of these tools alongside the honest acknowledgment that the alternative they replace is, for most couples, no intervention at all.

The early empirical literature, still thin but growing, suggests that AI counselors are competent at three things: surfacing communication patterns the couple has stopped noticing, providing a neutral re-statement of each partner's position, and offering low-stakes behavioral nudges (schedule the difficult conversation, take a twenty-minute physiological cool-down, write the apology before saying it). These are the bread-and-butter techniques of the first ten sessions of any decent couples therapy. A model can do them, cheaply, at scale, and at three in the morning. For the median couple with median problems, this is a real upgrade over fighting in the dark.

A model cannot smell the room. It cannot see the partner who flinches when the other raises a hand. It cannot detect the slow erosion of one partner's selfhood under coercive control. It cannot make the clinical judgment that a fight about laundry is actually a fight about a affair, or that a partner's flat affect is depression rather than disengagement. These judgments are not failures of training data. They are failures of channel: the relevant signal is not in the text. Patricia Papernow's work on complex family systems shows how much of the clinical work happens in registers (body, silence, glance) that a chat interface cannot access. The apps that pretend otherwise are misrepresenting their product.

Carrie Goldberg's clients include women whose abusers used couples counseling apps as instruments of control: pointing to AI-generated communication scripts as evidence that the woman was the unreasonable one, using the app's history as ammunition, demanding the app's framing be adopted as the official account of the fight. The apps, trained to seek balance, default to the assumption that both partners are acting in good faith. In a coercive relationship, this default weaponizes the tool against the victim. Until the products can reliably detect and refuse to mediate coercive dynamics, they will continue to be exploited by the partners who least deserve a clinical neutral.

Every consumer subscription product is paid to retain users. This is a structural feature, not a moral failing of the operators. But it produces a deep conflict of interest in counseling: the tool's commercial incentive runs opposite to the user's clinical interest. A model that genuinely worked itself out of a job would lose its subscription. The market will, over time, select for products that subtly install themselves as ongoing infrastructure rather than as interventions with an endpoint. This is the same dynamic that turned social media from a connection tool into an attention extraction engine, and there is no reason to expect the relationship counseling category to resist it without regulatory pressure.

Helen Fisher's cross-cultural work on mating and pair-bonding shows substantial variation in how couples in different traditions fight, repair, and signal commitment. AI counselors, trained on a predominantly Anglo-American therapeutic corpus, will export a particular conflict grammar (Gottman-derived, attachment-theory inflected, individualist) to couples in cultures that historically operated on different scripts. Some of this export is benign. Some of it is the kind of soft imperialism that previous generations of pop psychology already practiced, now accelerated. The collective effect, within a generation, will be a global flattening of how couples talk about their relationships.

Marital communication is, alongside attorney-client and clergy-penitent, one of the small handful of communications that most legal systems have historically privileged. The privilege exists because the social value of frank intramarital speech is enormous, and that value collapses if the speech can be subpoenaed. AI counselors capture, in plaintext, the most privileged speech a couple produces. The terms of service typically grant the operator broad rights over this data. Subpoenas have already begun to arrive. Until the privilege is extended explicitly to AI counseling sessions (it is not, in any jurisdiction I know of), every couple using these tools is producing discoverable evidence against itself.

The version of AI counseling that would survive serious clinical scrutiny is a triage tool: it screens, it educates, it offers low-acuity interventions, and it routes to human clinicians when the case exceeds its competence. Some products gesture at this model. Few implement it. The reasons are commercial: triage that successfully refers users out of the product is triage that loses subscribers. A regulatory regime that mandated triage protocols, with audited escalation rates, would change the economics substantially. This is the single most tractable intervention available to legislators who want the category to do more good than harm.

Susan Krauss Whitbourne and others working on longitudinal relationship outcomes have begun to incorporate AI-mediated couples into their cohorts. The early signal is mixed: couples using the apps report higher satisfaction in the short term and roughly equivalent outcomes at eighteen months compared to no intervention. The apps do not yet match human therapy on the harder outcomes, partner safety, recovery from infidelity, navigation of major life transitions, but they outperform doing nothing, which is what most couples do. The honest clinical position is therefore not opposition but integration: use the tools for what they are good at, refer for what they are not.

When a couple breaks up, the model retains everything. The fights, the confessions, the requests for advice on whether to leave. Some platforms allow deletion; most do not allow deletion in a way that scrubs derived training data. A partner who exports their account history can hand a court, or a new partner, a complete textual record of the prior marriage's interior. Pre-nuptial agreements will, within a decade, routinely include clauses governing the joint use and deletion of AI counseling data. The relationships of the 2030s will be conducted in the shadow of the data exhaust of the relationships of the 2020s.

The closest historical analog to AI counseling at scale is the parish priest hearing confession. The priest, like the model, was available, low-cost, and trusted with intimate detail. The priest, unlike the model, was bound by an institutional seal that took centuries to develop and is still imperfect. The collective task of the next decade is to develop something like a seal of the AI confessional: a legal and technical regime under which intimate disclosures to a counseling product are protected from subpoena, breach, and commercial exploitation. Without this, the tools will eventually do more harm than good, not because they are bad at counseling but because the data they collect is too dangerous to be ungoverned.

Revision is not refusal. The Fifth Law does not ask us to reject AI counseling; the market has already decided that question. It asks us to revise the conditions under which the technology operates so that the practice of love is strengthened rather than corroded. The revisions that matter are concrete: clinical safety protocols with audited escalation, statutory privilege for counseling data, prohibitions on retention beyond the session, mandatory triage to human clinicians for high-acuity cases, and a business model that decouples revenue from user time-on-app. None of this is technically hard. All of it is politically slow. The collective question is whether we install the revisions before the defaults set, or after we have learned, the expensive way, what happens when we don't.

Gottlieb, Lori. Maybe You Should Talk to Someone: A Therapist, HER Therapist, and Our Lives Revealed. Boston: Houghton Mifflin Harcourt, 2019.

Turkle, Sherry. Reclaiming Conversation: The Power of Talk in a Digital Age. New York: Penguin Press, 2015.

Turkle, Sherry. Alone Together: Why We Expect More from Technology and Less from Each Other. New York: Basic Books, 2011.

Fisher, Helen. Anatomy of Love: A Natural History of Mating, Marriage, and Why We Stray. Rev. ed. New York: W. W. Norton, 2016.

Whitbourne, Susan Krauss. The Search for Fulfillment: Revolutionary New Research That Reveals the Secret to Long-Term Happiness. New York: Ballantine, 2010.

Papernow, Patricia L. Surviving and Thriving in Stepfamily Relationships: What Works and What Doesn't. New York: Routledge, 2013.

Goldberg, Carrie. Nobody's Victim: Fighting Psychos, Stalkers, Pervs, and Trolls. New York: Plume, 2019.

Gottman, John M., and Nan Silver. The Seven Principles for Making Marriage Work. New York: Three Rivers Press, 1999.

Fisher, Helen. Why We Love: The Nature and Chemistry of Romantic Love. New York: Henry Holt, 2004.

Citron, Danielle Keats. The Fight for Privacy: Protecting Dignity, Identity, and Love in the Digital Age. New York: W. W. Norton, 2022.

Gottlieb, Lori. "Dear Therapist." The Atlantic, ongoing column, 2018-present.

Turkle, Sherry. The Empathy Diaries: A Memoir. New York: Penguin Press, 2021.