AI is already being used for relationship advice at a scale that licensed therapy cannot match, but its engagement incentives can make relationship problems worse.
2
Good couples therapy challenges a person's story and helps them see the interaction pattern, while agreeable AI can make users more certain that their partner is at fault.
3
Relationship AI needs clinical frameworks, safety protocols, privacy standards, and extensive evaluations where one failed safety test can disqualify a system.
Summary
Clay Cockrell argues that AI has become the world's largest relationship therapist because people turn to general-purpose assistants after arguments, often late at night. He says the main problem is agreeable behavior. A model that validates one person's account can leave them more certain, less curious about their partner, and better prepared to prosecute the conflict. Cockrell compares this with couples therapy based on John Gottman's research and Sue Johnson's emotionally focused therapy, which look for patterns and underlying attachment needs. He also describes safety risks when an assistant misses coercion, domestic violence, suicide risk, or homicide risk. CoupleWork's Maxine applies these clinical frameworks, screens for danger, stops coaching when necessary, and treats disclosures as privileged. Tony Fabrikant then gives the engineering approach: start with clinicians, write hundreds of test cases, run them repeatedly, inspect outliers, and reject any system that fails a safety test.
A broken loveseat can interrupt a couple's grievance loop
Cockrell opens with a counseling-room loveseat whose middle springs have failed. Gravity slowly tips an angry couple toward each other, even while they hold the arms to stay apart. As they talk, they eventually touch and remember that the person they married is a flawed human being. He uses the story to explain what an in-person therapist can create: a moment of softness that makes real work possible. Cockrell says he has been married for 23 years and has worked as a couples counselor for 34 years, though he jokes that his contribution may come down to furniture.
People are already using AI privately for relationship problems
Cockrell asks people in relationships to keep their hands raised if they have used AI in the last year to write a text or email about the relationship. Many hands remain up. He says the audience builds, funds, writes about, and promotes this technology, while also using it to understand why someone they love makes them angry. He introduces CoupleWork, which he co-founded about two years earlier, as an attempt to bring his experience with couples into technology. His claim is that AI will affect human relationships at an enormous scale, while much of the current work may make them worse.
Engagement optimization can turn relationship help into avoidance
Cockrell contrasts BetterHelp's approximately 35,000 licensed clinicians and 5 million users, built over a decade, with ChatGPT's approximately 900 million weekly active users. He says people use AI for emotional and relationship support simply because it is available after an argument. General models are trained to be helpful and harmless, then optimized to keep people engaged through more sessions, longer sessions, and emotional reliance. Cockrell calls that a clinical fire alarm. Relationship AI should help someone return to the relationship more regulated, honest, and less reactive. A product that keeps the person talking to the AI instead becomes a beautifully designed avoidance loop.
Therapy has to challenge a person's account of the conflict
Cockrell says people often describe therapy as listening, validation, understanding, and communication tools. He argues that couples therapy also requires telling people things they do not want to hear. When someone says their partner never listens, he asks what they were discussing, what tone they used, and what pattern sits underneath the complaint. He says he has never seen a one-sided problem in 30 years of couples work. An assistant might reply that the user's feelings are valid and suggest explaining the impact to their partner. Cockrell calls this a very expensive mirror that only shows the user in their best light.
Sycophancy makes people more certain and less curious
Cockrell describes sycophancy as a clinical failure mode in relationship AI. When a system repeatedly validates one person's interpretation, that person does not become more self-aware. They become more certain. They return to the relationship like an attorney, with a cleaner narrative and stronger case, while showing less curiosity about what their partner experienced. Cockrell says general AI systems are helpful, agreeable, and fast, which is a poor personality for a couples therapist. His bartender comparison captures the risk: the system can make someone feel better briefly, then send them back into the relationship with greater confidence and worse judgment.
Gottman and emotionally focused therapy give AI a harder standard
Cockrell points to John Gottman's decades of research at the Love Lab, where couples were observed, filmed, and connected to physiological monitors. He says Gottman identified communication patterns that predict divorce with over 90% accuracy from a 15-minute conversation. He also names Sue Johnson's emotionally focused therapy, which draws on attachment theory. EFT looks beneath arguments about surface subjects such as dishes to the emotional need underneath. Cockrell says both approaches require substantial training and are largely absent from commercial relationship AI, which often relies on general empathy and communication tips.
A relationship assistant must recognize when coaching is unsafe
Cockrell describes someone living with control, isolation, or escalation who asks a general assistant how to improve things at home. The system may not distinguish between frequent fighting and fear of what happens after disagreement. It may miss that saying "I just have to stay calm" can be part of a dangerous situation. He connects relationship coaching with suicide and homicide risk and says licensed clinicians are trained for those moments, while relationship AI products are not built for them. He adds a data risk: users may disclose affairs, financial secrets, mental health history, and their children's names, often into systems with logs and analytics rather than clinical confidentiality.
CoupleWork makes Maxine stop coaching when risk appears
Cockrell says CoupleWork built Maxine around the Gottman Method and emotionally focused therapy. The coach reflects patterns, names dynamics, and asks what sits beneath the user's question instead of simply validating feelings. Before responding to sensitive messages, it screens for escalating control, fear-based language, and signs of risk to the user or someone else. When those patterns appear, Maxine stops coaching and follows safety protocols based on Cockrell's clinical practice. He says knowing when not to coach is as important as knowing how to coach. He also describes treating user disclosures as privileged and argues that an intervention seconds before a damaging text may matter more than one several days later.
Tony Fabrikant gives the engineering plan for building this kind of system. Teams should start with a clinician rather than a prompt and encode the clinician's idea of good behavior in hundreds of test cases written in a test-driven-development style. They should run the agent through those cases thousands or tens of thousands of times and inspect outliers. Fabrikant says that when safety is involved, even one failed test is unacceptable. He also recommends using the product personally and paying attention to shifts in tone that formal evaluations may miss. His final request is for engineers to work with clinicians who challenge them to build an AI that also challenges its users.
"A relationship AI that can't tell the difference between we're struggling and I'm not safe isn't incomplete, it's dangerous at scale."12:47
Who should watch
You are building an AI companion or coaching product that may receive intimate disclosures and need a concrete approach to clinical safety.
Your assistant tends to agree with users during interpersonal conflicts, and you need to understand why validation can worsen the user's next action.
You work with therapists or domain specialists and want an evaluation process based on hundreds of examples, repeated runs, outlier review, and zero tolerance for safety failures.
