It is late. Your partner has gone to bed, the washing-up is still waiting and you ask a chatbot whether you were unreasonable. Its reply is warm, organised and wonderfully certain. For a moment, being understood feels very close to being right.
Does AI therapy work? Some purpose-built tools have shown benefits in research, but that does not make a general chatbot a therapist or a reliable judge of your relationship. If you are considering an AI couples coach, start with what it has actually been tested to do. A helpful-sounding response and evidence of safe, effective care are different things.
Start by checking what “therapy” means here
An online appointment with a qualified human therapist, an automated wellbeing app and a general-purpose chatbot are different services. Before paying or sharing anything personal, find out who or what will respond. A friendly avatar does not settle that question.
The American Psychological Association advises against relying on consumer chatbots and wellness apps to replace qualified mental health care. Its advisory distinguishes general-purpose systems from tools developed for specific mental health uses; evidence for one category cannot simply be transferred to another.
Look for a clear description of the service: everyday reflection, communication practice or clinical treatment. If the advertising promises therapy but the small print calls it entertainment, pause. You should be able to understand what you are buying.
Source:American Psychological Association: advisory on AI and mental health
What the encouraging research actually tested
In a 2025 randomised trial, Dartmouth researchers compared a purpose-built chatbot called Therabot with a waiting-list group. The study involved adults with depression, anxiety or eating-disorder concerns. It reported improvements in symptom measures, with assessments at four and eight weeks.
Therabot used a specially developed training set, and researchers reviewed conversations and were equipped to intervene if safety concerns arose. The comparison was not with human therapy, and the study did not test couples counselling. Its findings do not establish that any consumer chatbot offers the same benefit or safeguards.
When a provider says “research-backed”, ask for the study of the actual tool, the people studied, what changed and how safety was monitored. A citation about a recognised therapy approach is not automatically an evaluation of an app that mentions it.
An agreeable answer is not an independent verdict
The APA warns that chatbots can validate users’ views and produce biased or inaccurate advice. Agreement may feel supportive while leaving an interpretation unchallenged. That matters when you are asking a system to explain another person’s behaviour.
Imagine typing, “My partner ignored me all evening.” You have described your experience, but the sentence leaves out what happened before dinner, what “ignored” means and what your partner would say. A fluent response does not fill those gaps with facts.
Instead of bringing a generated verdict into the kitchen, separate what you observed from what you concluded: “I tried to talk twice while you were on your phone. I felt brushed aside. What was happening for you?” You can listen without withdrawing your request for attention or excusing repeated disregard.
Neither person should have to argue against a chatbot to be heard. Keep decisions about boundaries, responsibility and the relationship with the people affected, using qualified support where needed.
Source:American Psychological Association: advisory on AI and mental health
Check privacy before telling the whole story
The FTC cautions that health apps may use or share sensitive information beyond providing the service. It recommends comparing privacy notices and checking protective settings. A private-feeling conversation is not proof of confidentiality.
Read the current policy and settings before entering intimate details. The questions below are an editorial checklist; if the answers remain unclear, keep sensitive information out.
Your partner’s messages, medical history and sexual disclosures also concern their privacy. Do not upload them without consent. A fictional example or broad description is usually enough for a wording task; removing a name alone does not remove every identifying detail.
- What information is stored, and for how long?
- Can people review conversations or use them to train models?
- Who receives the data, and for what purpose?
- Which sharing controls exist, and what does deletion actually remove?
- Who can access my account, device or notification previews?
Choose a task small enough to check yourself
If you choose to use a writing tool, give it a bounded job you can assess. “Help me organise questions for an appointment” is easier to check than “Tell me whether my partner is emotionally abusive.” A tool cannot safely settle that decision for you.
Here is an original preparation example. You want to discuss how evenings disappear into separate screens. Write three notes yourself: what happened, what you felt and what you want to ask. You might have: “We barely spoke after dinner. I missed you. Could we agree a short phone-free catch-up?”
If you use AI to shorten those notes, reject anything that invents motives, adds a diagnosis or promises an outcome. Read the result aloud and keep only words you mean. Paper and a pen can do the same job.
Then ask whether your partner wants to talk and whether now is a good time. Review the real conversation rather than the quality of the generated reply. Did you explain the concern accurately? Did both people have room to respond? Did any agreement respect each person’s choice? This is editorial preparation, not a treatment exercise.
Use these questions before choosing an AI couples coach
Availability and price matter, especially when evenings are busy or professional appointments are hard to obtain. They do not answer whether a service fits your need. Choose a specific job first, then compare the offer against it.
For everyday practice, look for prompts you can question, skip or stop. For treatment, seek appropriately qualified care and ask about relevant experience. If you already have a therapist, discuss any AI advice you are using rather than quietly replacing the agreed plan.
A provider should explain the limits plainly. You can ask these questions before subscribing, without sharing your relationship history:
- What evidence evaluates this particular product for my intended use?
- Is a qualified human involved, and how can I reach them?
- What happens when a user reports distress, danger or a harmful response?
- What are the full cost, cancellation terms and data controls?
Source:American Psychological Association: advisory on AI and mental health
When the next step needs a person
If distress persists or you are worried about your mental health, contact a qualified professional. NIMH describes primary care as one route to an initial assessment and referral. Tell your practitioner if chatbot use is affecting your care; do not change treatment on a bot’s advice.
Abuse, coercion or fear of retaliation need confidential specialist support. The Hotline explains that joint counselling can expose an abused partner to further harm. An AI-mediated discussion is not a substitute for a safety plan. Seek a local domestic-abuse service or a live Hotline advocate in the US, using a safer device if yours may be monitored.
For immediate danger, use local emergency services. In the US, call or text 988 for suicidal crisis or emotional distress; call 911 for a life-threatening emergency. Elsewhere, use your local crisis service. Do not wait for a chatbot to assess the situation.
You deserve support that can respond to the need in front of you. The warm reply on your phone does not have to be the last word.
Source:NIMH: finding professional care and immediate support;The Hotline: why joint counselling can be unsafe with abuse
If a relationship feels unsafe, seek confidential specialist help rather than a joint app exercise.
