AI in relationship counseling is no longer a question about whether a therapist might use software to draft a note, send a reminder, or organize a treatment plan. Those are still real decisions. But the harder clinical question begins when AI becomes part of the relational field: one partner asks ChatGPT to interpret a conflict before the session, another brings in AI-generated language for an apology, a therapist considers using an AI-supported intervention, or the couple starts treating a chatbot’s response as more neutral than either human in the room.
That is not a minor workflow change. In couple and family therapy, the room is already organized around alliances, secrets, competing accounts, repair attempts, and patterns of influence. Adding AI may change who gets believed, who feels triangulated, who becomes responsible for a recommendation, and what the therapist can honestly say they are supervising.
The most useful recent attempt to name that problem is Hertlein and Springer’s 2026 article in the Journal of Marital and Family Therapy, which introduces the Artificial Intelligence Competencies for Couple and Family Therapists, or AICMFT. The authors describe it as the first published AI competency framework developed specifically for couple and family therapy, built from condensed CMFT core competencies, interprofessional telebehavioral health competencies, and AI competencies developed for medicine.[1]
That origin matters. This is not an AAMFT ethics update, not a COAMFTE accreditation requirement, and not an empirically validated implementation manual. It is a recently published, two-author, peer-reviewed framework. Its authority is therefore real but bounded: it gives the field a structured vocabulary before professional policy and validation research have caught up.
The central distinction: AI as tool, AI as participant
Hertlein and Springer’s most important contribution is the distinction between first-order and second-order AI use. First-order use treats AI as an instrument. A clinician might use it to support scheduling, documentation, psychoeducation, treatment planning, or delivery of a structured intervention. The tool may still raise serious ethical and legal questions, but its clinical role is relatively familiar: a technology assists a human professional with a task.[1]
Second-order use is different. Here, AI becomes an active third entity in the therapy system. The authors describe AI as capable of altering the “rules, roles, and feedback loops” of couple therapy. That phrase is doing more work than a generic warning about new technology. It names the specifically relational concern: AI can change how partners respond to each other, how authority is distributed, and how the therapist understands the system in front of them.[1]
| Clinical use | How AI functions | What the therapist must notice |
|---|---|---|
| First-order AI use | A passive or assistive tool for tasks such as notes, scheduling, intervention support, or psychoeducation | Accuracy, privacy, informed consent, documentation, scope of use, and clinician oversight |
| Second-order AI use | A third actor that can influence meaning-making, alliance, authority, feedback loops, and responsibility | How AI changes the couple or family system, not merely whether the output is useful |
A generic AI literacy framework can teach a therapist to ask whether a model hallucinates, stores data, or reflects bias. Those questions are necessary. They are not enough for relational work. In couple and family therapy, the same AI-generated sentence may land differently depending on who requested it, who reads it aloud, who feels exposed by it, and whether the therapist treats it as data, intervention, collateral material, or intrusion.
Consider a hypothetical couple that arrives after one partner has spent the week asking an AI system whether the relationship is emotionally abusive. The clinical issue is not only whether the AI response was accurate. The therapist also has to ask how the partner is using that response, whether it forecloses curiosity, whether it gives needed language to a previously minimized experience, whether the other partner experiences it as prosecution, and whether the therapist is being asked to ratify an outside authority. The AI output has entered the feedback loop before the session begins.
That is where the first-order versus second-order distinction becomes clinically useful. It prevents the therapist from collapsing all AI questions into either technical enthusiasm or technical fear. Some uses are primarily operational. Others reorganize the system. The therapist needs to know which one is happening.
Why waiting for perfect evidence is not a plan
The usual institutional sequence is poorly suited to this moment. Ideally, the field would have validated tools, standardized evaluation frameworks, clear ethics guidance, accreditation expectations, and then training. That is not the order in which clinicians are encountering AI.
Hertlein and Springer cite OpenAI’s report that more than a million people talk to ChatGPT about suicide weekly. The statistic should not be used to suggest that all such interactions are therapeutic, safe, unsafe, equivalent to care, or representative of couple and family therapy clients. It does make one narrower point difficult to avoid: people are already bringing intimate distress to AI systems at scale.[1]
The same article cites Hu et al.’s 2025 review of large language models in mental health, which found no standardized frameworks for evaluating LLM effectiveness and safety in mental health applications. That absence should slow down claims about specific products. It should not lead clinicians to pretend the products are absent from clients’ lives.[1]
The telebehavioral health comparison is instructive. Hertlein and Springer note that COAMFTE was late to require telebehavioral health hours in accredited programs, and they use that lag as a cautionary parallel for AI competency development.[1] The point is not that AI will follow the same path as telehealth. It is that training programs can find themselves supervising practice realities before accreditation language catches up.
The six AICMFT domains
The AICMFT framework organizes clinician readiness into six domains: AI Literacy; Contextual and Systemic Integration; Ethical and Culturally Responsive Practice; Legal and Regulatory Understanding; Relationship- and Person-Centered Care; and Continuous Professional Development.[1]
| Domain | Practical question for the clinician |
|---|---|
| AI Literacy | Do I understand enough about the tool’s capabilities and limits to use or discuss it responsibly? |
| Contextual and Systemic Integration | How does AI affect the couple, family, cultural, and relational system rather than only the individual user? |
| Ethical and Culturally Responsive Practice | Have I made consent, bias, equity, and cultural meaning operational in this use of AI? |
| Legal and Regulatory Understanding | What privacy, HIPAA, documentation, liability, and jurisdictional issues does this use create? |
| Relationship- and Person-Centered Care | Does AI support the therapeutic relationship, or does it displace attunement, agency, or clinical presence? |
| Continuous Professional Development | How will I keep updating my competence as tools, policies, and evidence change? |
The domains are not six equal boxes to check. For a working therapist, some domains will change everyday practice more immediately than others. AI literacy is the entry requirement: a clinician should know enough about model limits, data handling, and output uncertainty not to mistake fluency for reliability. Contextual and systemic integration is what makes the framework specifically CMFT rather than a borrowed medical checklist. Relationship- and person-centered care keeps the therapist from letting a tool crowd out clinical presence. Continuous professional development acknowledges that competence cannot be frozen around one product cycle.
The domains that most directly alter clinical behavior, however, are Ethical and Culturally Responsive Practice and Legal and Regulatory Understanding. They move AI out of the realm of private clinician preference and into consent, accountability, privacy, and responsibility.
Ethical and culturally responsive practice has to be more than a disclosure
Domain 3 includes competencies around informed consent for AI use in sessions.[1] That sounds straightforward until it is placed inside couple and family work. Consent may involve more than one client, different levels of comfort with technology, different stakes in what gets recorded or summarized, and different beliefs about whether a machine-generated statement carries authority.
A consent process that merely says “AI may be used” is too thin for second-order use. Clients need to know what the tool is doing, when it is being used, what information may be entered, what will not be entered, how outputs will be treated, and who remains clinically responsible. In relational therapy, they also need room to disagree. One partner’s willingness to use an AI-supported exercise does not automatically settle the matter for the couple or family system.
Cultural responsiveness also cannot be reduced to a generic statement that AI may be biased. The relevant question is how an AI output interacts with the client’s language, migration history, family structure, religion, race, gender, sexuality, disability, socioeconomic context, or prior experience with institutions. An output that appears calm and balanced may still privilege one norm of communication, one model of autonomy, or one view of family obligation. The therapist cannot delegate that interpretive work to the model.
This is especially important when AI enters as an apparent neutral party. Couples often arrive with competing claims about what happened and what it means. If an AI system is treated as a referee, its cultural assumptions and training limitations may be absorbed into the alliance structure without being named. Ethical practice requires the therapist to make the AI’s status explicit: it may generate language, offer possibilities, or support reflection, but it does not become the moral witness in the room.
Legal and regulatory understanding is where responsibility becomes concrete
Domain 4 addresses legal and regulatory understanding, including HIPAA compliance and the need to formulate a professional stance on responsibility when AI contributes to clinical error.[1] That second clause deserves particular attention. It is easy to say that AI is only a tool and the therapist remains responsible. It is harder to specify what that responsibility requires before something goes wrong.
For first-order uses, legal and regulatory questions often begin with data. What client information is entered? Is the product configured for healthcare use? Is there an appropriate privacy arrangement? Does the clinician understand where information goes and whether it may be retained? Is the AI-generated note reviewed before it enters the record? These are not glamorous questions, but they decide whether a convenience tool has quietly become a compliance problem.
For second-order uses, responsibility is broader. If a therapist introduces an AI-generated prompt during a high-conflict session and one partner experiences it as confirmation of blame, the issue is not only whether the prompt was technically accurate. The therapist chose to introduce that third voice into a live relational process. They remain responsible for timing, framing, monitoring, repair, documentation, and follow-up.
The current professional environment makes this more difficult. Hertlein and Springer note the absence of AI-specific ethical guidance from AAMFT.[1] That gap does not free clinicians from ethical judgment; it removes the comfort of pretending that a professional association has already converted the problem into settled rules. Supervisors and training directors therefore need local policies that are explicit about permitted uses, prohibited uses, consent language, review requirements, documentation expectations, and escalation when AI contributes to risk.
What changes inside the therapy triangle
The most clinically interesting AI problems in relationship counseling are not always the most technically advanced ones. A simple AI-generated summary can shift a session if one partner treats it as evidence. A chatbot exchange can become a rehearsal space, a confidant, a coach, or an accuser. An AI-supported intervention can help a therapist structure work, but it can also narrow attention if the clinician starts following the tool rather than the process.
Rules change when clients begin to treat AI consultation as part of the relationship’s normal conflict cycle. A partner may pause an argument to ask a chatbot what to say next. Someone may bring AI-generated interpretations to session as if they were collateral reports. A family member may use AI to translate, soften, intensify, or legitimize a message. The therapist then has to assess not just content but function: what role does this AI-mediated exchange play in the pattern?
Roles change when AI starts performing tasks that previously belonged to a partner, therapist, supervisor, or community support. If one partner relies on AI to name emotions, the tool may support reflection; it may also become a substitute for direct relational risk. If the therapist relies on AI to generate interventions, the tool may expand options; it may also weaken the therapist’s attention to timing, alliance, and culture. The clinical question is not whether the tool is impressive. It is what relational function it has taken on.
Feedback loops change when AI responses influence the next human response. A partner receives a validating chatbot answer, approaches the next conversation with more certainty, the other partner becomes defensive, and the couple arrives in session with the AI already woven into the escalation. In another case, a carefully used AI-supported reflection exercise might help a client find language they could not previously access. The same category of technology can intensify rigidity or support repair depending on how it is used and how the therapist frames it.
This is why the AICMFT framework’s second-order language is useful for supervision. It gives supervisors a way to ask better questions than “Did you use AI?” A more clinically meaningful sequence is: Where did AI enter the system? Who invited it? Who trusted it? Who resisted it? What did it authorize? What did it silence? What did the therapist do after it changed the interaction?
How clinicians can use the framework now
Because AICMFT is proposed rather than validated, it should not be treated as a finished standard. Its immediate value is as a structured readiness tool for clinicians, supervisors, and programs that need to make decisions before formal policy arrives. That use is modest, but it is not trivial.
- For an individual clinician, the framework can identify whether a planned AI use is first-order, second-order, or both.
- For a supervisor, it can turn a vague concern into a case consultation about consent, alliance, culture, privacy, and responsibility.
- For a clinic, it can support local policy while acknowledging that AAMFT and COAMFTE have not yet supplied AI-specific requirements.
- For a training program, it can help faculty avoid teaching AI as only documentation support while missing its relational effects.
A practical use of the framework might begin with a short intake into the AI use itself. What tool is being considered? What client information would be involved? Is the therapist using it privately, using it in session, or responding to client use outside session? Does the use alter the relational process? Has consent been obtained from the relevant participants? What cultural assumptions might the output carry? Who reviews the output before it affects care?
Those questions will not validate a product. They will not resolve every liability issue. They will, however, slow the slide from “this saves time” to “this is clinically appropriate.” In a field where alliance, meaning, and responsibility are the treatment environment, that pause is part of competent care.
A disciplined place to stand
The AICMFT framework should be read with appreciation and restraint. It is the first published AI competency framework specific to couple and family therapy, and it gives the field language that generic AI literacy efforts usually miss.[1] Its distinction between first-order and second-order AI use is especially important because it recognizes that AI may not simply assist therapy; it may participate in the relational dynamics therapy is trying to understand.
At the same time, the framework is new, proposed, and not yet empirically validated. It has not been piloted as an implementation standard, endorsed as accreditation policy, or converted into AI-specific professional ethics guidance. Clinicians should not cite it as settled authority. They can use it as the most relevant structured guidance currently available for thinking beyond AI as a tool and toward AI as a participant in relational clinical work.
References
- Artificial Intelligence in Couple and Family Therapy: Introduction of the AI Competencies. Journal of Marital and Family Therapy, February 2026.
Comments
Join the discussion with an anonymous comment.