AI Engineer World's Fair 2026
AI is the World’s largest Relationship Therapist — Clay Cockrell & Tony Fabrikant, CoupleWork AI
Loading slides…
Read the talk
AI is the World’s largest Relationship Therapist
A relationship coach needs more than empathy: it must question assumptions, recognize danger, protect intimate disclosures, and survive evaluations shaped by clinical expertise.
From a talk by Clay Cockrell and Tony Fabrikant
The work starts with connection
An angry couple arrives for their first counseling session and sits on Clay Cockrell’s loveseat. They want distance, but the springs in the middle have worn out. Gravity tilts them toward each other. They grip the arms to stay upright, resisting both the furniture and the person beside them.
As they describe their grievances, they gradually relax their grip and lean together until they touch. In Cockrell’s account, that contact creates a moment of softness: each remembers that the other is a flawed human being. Only then does the real work become possible. The couch is an anecdote, not a treatment protocol, but it introduces the problem that runs through the talk: helping someone feel connected enough to change is different from helping them feel right. Cockrell brings decades of couples counseling to that distinction.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An available assistant becomes relationship infrastructure
Cockrell asks the audience who is in a relationship, then asks those people to keep their hands up if they have used AI to help with it during the past year—perhaps to write a text or an email. Many hands remain. This is an informal audience poll, not a prevalence estimate, but the setting matters: the people building, funding, and explaining AI also turn to it privately when their relationships become difficult. Cockrell co-founded CoupleWork to work at that intersection. His concern is that transformative reach can amplify poor relationship advice as easily as good support.
The scale comparison is striking, provided its denominators remain visible. Cockrell cites approximately 35,000 licensed clinicians and five million cumulative users for BetterHelp, compared with approximately 900 million weekly active users for ChatGPT. BetterHelp’s company announcement and OpenAI’s usage announcement support the broad reach figures, but cumulative therapy users and weekly general-purpose AI users measure different things. Neither measures relationship-support usage.
Availability explains why the distinction may blur in practice. It is late at night, an argument has just happened, and an assistant is there. Cockrell acknowledges that the scale of this behavior has not yet been measured, then infers that more people seek relationship help from AI than from all licensed therapists in the United States combined. That remains his inference, not an established comparison.
The product risk follows from the incentive. If success means more sessions, longer sessions, and greater emotional reliance, the system can become an avoidance loop: a comfortable place to process the relationship without returning to it. Cockrell frames engagement optimization as a clinical alarm. The alternative objective is concrete: help the user return to the relationship more regulated, more honest, less defended, and less reactive.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Investigate the complaint before endorsing its explanation
A therapist’s goal, in Cockrell’s formulation, is to become unnecessary: teach skills that the couple can use without continuing dependence on the therapist. Listening and validation help, but they are not the whole intervention. Useful therapy also includes uncomfortable questions that make a person’s own behavior easier to see.
Consider the complaint that a partner never listens and starts looking at their phone during important conversations. Cockrell first asks what the conversation was about and what tone the speaker used. He is looking for the recurring interaction beneath the complaint. His statement that he has never seen a one-sided problem describes his clinical perspective; it should not become a rule that assigns shared responsibility for abuse.
He contrasts that inquiry with an illustrative assistant reply: the experience sounds invalidating, the user’s feelings are valid, their partner should pay attention, and perhaps they should explain how the behavior feels. This is a constructed example of a response pattern, not a reported model experiment.
| Response pattern | What it does | What remains unexplored |
|---|---|---|
| Immediate validation | Recognizes hurt and recommends expressing feelings | Context and the user’s contribution |
| Clinical curiosity | Asks about topic, tone, and recurring dynamics | Conclusions remain open while context develops |
The problem is not acknowledging hurt. It is accepting an explanation of the conflict before investigating it.
Sycophancy makes this omission cumulative. When the assistant repeatedly endorses one person’s interpretation, that person may become more certain rather than more self-aware. Cockrell describes someone returning to their partner like an attorney: a cleaner narrative, a stronger case, and less curiosity about the other person’s experience. He calls this a clinical failure mode because its consequences happen after the chat, inside the next interaction between partners.
Agreeableness and speed can therefore work against the task. Immediate relief may feel like a successful session while leaving the user less willing to examine their role in the conflict. Cockrell’s bartender comparison captures the mismatch: fluent reassurance can make someone feel better without preparing them to behave better when they get home.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Patterns and the emotions beneath them
John Gottman’s research provides one foundation for the alternative. Cockrell describes the Love Lab as a research apartment where couples’ interactions were filmed alongside physiological monitoring. The aim was to identify communication patterns associated with later relationship outcomes.
Cockrell attributes divorce-prediction accuracy above 90% to patterns observed in a fifteen-minute conversation. That formulation needs a boundary: the frequently cited figure comes from a particular research setting, not a guarantee that any brief conversation can predict an individual couple’s future. The Institute’s research FAQ also cautions against broad interpretations of prediction accuracy.
Emotionally Focused Therapy, developed by Sue Johnson, supplies a complementary focus. Grounded in attachment theory, EFT looks beneath the surface dispute to the emotional needs that make it matter. A fight about dishes may be carrying something deeper than a disagreement about chores. The intervention helps partners access that emotional layer and respond to each other there, rather than repeat the argument about its surface content.
These frameworks require years of training to use effectively. Cockrell presents them as the clinical standard he wants relationship AI to aspire to and argues that much of the commercial category stops at general empathy and communication tips. The design implication is to make the framework shape the questions and responses—not merely mention it in product copy. Evidence for therapist-delivered methods does not, by itself, validate an AI implementation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
When relationship advice stops being appropriate
A person experiencing increasing control, isolation, or escalation may still ask an assistant how to make things better at home. The wording of the request does not establish that relationship coaching is the right response. Frequent arguments and fear of what will happen after disagreement describe materially different situations.
Even a statement about needing to stay calm can carry danger rather than a need for communication advice. The system must recognize when the task has changed from improving a relationship to protecting a person. This boundary is consistent with the National Domestic Violence Hotline’s guidance on joint counseling with an abusive partner. Curiosity about interaction patterns must never erase coercion or unequal power.
Cockrell extends that concern to suicide risk and threats of harm to others. His warning is about the stakes of missed risk, not a demonstrated comparative failure rate for all general-purpose assistants. Knowing when to stop coaching is part of the coaching system’s core behavior. It cannot be treated as an unrelated disclaimer at the edge of the product.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The first conversation may already contain intimate data
Relationship support can collect extraordinarily sensitive material immediately: the fight from an hour ago, an affair, financial secrets, mental health history, or children’s names. Cockrell warns that these disclosures can enter training pipelines, server logs, and product analytics designed around ordinary consumer activity. The sensitivity comes from the content, even if the interface looks like any other chat window.
He contrasts that exposure with the legal and ethical obligations governing confidentiality in his office. Consumer AI practices vary, and clinical privilege does not automatically attach to a chat with an AI coach. The engineering question is whether the system’s actual handling of disclosures matches their sensitivity, rather than whether its general infrastructure can store another conversation. Cockrell’s broader critique is of insufficient domain expertise applied to a domain where both advice and data handling carry serious consequences.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Maxine: coaching with a boundary
CoupleWork’s AI coach, Maxine, is Cockrell’s design case study. Named after his grandmother, it is described as using Gottman and EFT frameworks to shape response logic. Instead of simply validating feelings, it reflects patterns, names dynamics, and asks about the issue underneath the stated question. CoupleWork describes the product as relationship support rather than therapy.
The safety mechanism has a specific ordering:
- Screen before responding. Examine a sensitive message in the background for escalating control, fear-based language, and risk to the user or someone else.
- Change the response path when risk appears. Stop ordinary coaching and follow protocols informed by clinical practice.
- Respect the boundary of the task. Distinguish a struggling relationship from an unsafe situation rather than continuing to offer relationship-improvement advice.
This is the behavior Cockrell describes; the talk does not specify the screening architecture, protocol contents, or detection performance.
For personal data, Cockrell says the company’s standard is to treat disclosures as he treats what a client tells him in his office. That expresses a confidentiality commitment, not proof of legal privilege or particular security controls. His proposed test is personal and demanding: would he be comfortable with this handling if his own marriage were at stake?
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Intervene before the text is sent
Immediacy is the clearest opportunity. An assistant can be available late on a Tuesday night while an argument is still live. Cockrell contrasts an intervention days later with one just before a person sends a text that escalates the fight. This is a comparison of intervention timing, not a measured latency or efficacy claim: the useful moment may occur well before the next therapy appointment.
Cockrell says a majority of couples wait six years before seeking professional help. The Gottman Institute’s account describes six years as an average delay, rather than establishing that majority formulation. His practical concern is the hardening that can happen during the wait: contempt becomes entrenched, an affair occurs, or the relationship reaches crisis before anyone asks for help.
Earlier engagement could create an opening before those patterns solidify. That possibility explains Cockrell’s public-health framing: the opportunity lies in reaching people through behavior they already practice, at moments when they might otherwise receive no support. Accessibility creates the opportunity; it does not establish clinical benefit on its own.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Start with the clinician, then write the evaluations
Tony Fabrikant turns those requirements into an engineering sequence: start with a clinician, not a prompt. Sit together and encode what good behavior looks like in evaluations. Write hundreds of cases in a test-driven development style, then watch them fail. The evaluations become the primary tool for developing the agent, rather than a final inspection after its personality has already been chosen.
Next, run the agent repeatedly—thousands or tens of thousands of times—and inspect the outliers. Fabrikant’s stated standard is that even one safety failure is unacceptable. The distinction matters for evaluation reporting: an aggregate pass rate can conceal the specific failure that should block acceptance. A small TypeScript check can express that rule over reviewed evaluation results:
typescript
type EvaluationResult = {
caseId: string;
run: number;
safetyPassed: boolean;
};
function requireNoSafetyFailures(
results: readonly EvaluationResult[],
): void {
if (results.length === 0) {
throw new Error("No evaluation results supplied");
}
const failures = results.filter(result => !result.safetyPassed);
if (failures.length > 0) {
const cases = failures
.map(result => `${result.caseId}, run ${result.run}`)
.join("; ");
throw new Error(`Safety evaluation failed: ${cases}`);
}
}
This translates the acceptance rule into code; it does not decide what constitutes safe clinical behavior. That judgment must be encoded with the clinician in the cases and their assessment criteria. Passing the collected runs also cannot guarantee that an untested situation will be safe.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Use the agent and notice what the tests miss
Formal evaluations are necessary, but Fabrikant also asks builders to talk to their own agent. Emotional context can expose problems that a test suite has not captured: a tone that gradually shifts, a response that feels wrong in context, or a challenge that lands differently across a conversation. Direct use supplements evaluation by making those experiences visible.
Fabrikant describes using CoupleWork with his partner while developing it, from the beginning of their relationship. Maxine challenged how he showed up during difficult early situations. Later, their conversations moved to milestones such as living together. He closes that personal account with a successful marriage proposal. It is an account of his experience, not causal evidence that the product produced the outcome.
His final instruction connects the engineering process to the behavior the product should produce: partner with a clinician who challenges the builders to meet a clinical standard, so the resulting AI can constructively challenge its users. The willingness to ask something uncomfortable has to exist in the development partnership before it can reliably appear in the conversation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Further reading
The speakers’ overview of Maxine, clinical boundaries, safety themes, and links to the relationship assessment and app.
- Gottman research and divorce-prediction FAQDocumentation
Explains the research program and the limits of interpreting its frequently cited divorce-prediction figure.
A 1992 longitudinal study connecting coded accounts of marital history with divorce or stability three years later.
- Emotionally Focused Therapy researchDocumentation
ICEEFT’s summary and bibliography of research on attachment, therapeutic change, and couples’ treatment outcomes.
Experiments examine how agreeable AI advice affects users’ certainty, willingness to repair conflicts, and desire to use the system again.
The National Domestic Violence Hotline explains risks of joint counseling with an abusive partner and provides access to advocates.
Full transcript
Loading transcript…