Sections
The thing guys are calling AI therapy now is usually not Woebot, Wysa, or some mental health startup with a cute mascot and a CBT worksheet engine. It’s ChatGPT at 1:12 AM after a fight, Claude after the third drink and the fourth version of the same thought, Gemini on the phone in the parking lot before going home. That’s the real situation, because that’s what people are actually doing.
They aren’t opening a mental health app because they want a digital therapist. They’re opening the same general purpose assistant they use for work, then slowly drifting into the stuff they would never say cleanly to another person. The marriage is bad, the drinking is worse than they admit, the rage is too close to the surface, and the breakup is six months old and still running the room. The prompt starts practical and gets personal by accident. That’s what’s actually happening.
And I get why it happens. A good LLM does something most people in your life don’t do very well. It sits there, answers immediately, doesn’t look disappointed, doesn’t interrupt. It lets you dump the ugly first draft of the thought before you have to make it socially acceptable. For a lot of guys that’s intoxicating, and it has nothing to do with the bot being wise, it’s just that nobody else is awake at 1 AM and willing to sit there.
The problem is when it becomes the only place they’re honest, and nobody who actually knows them is ever in the room.
So no, I’m not going to do the lazy old man thing where every new tool is bad because it has a screen. AI can be useful. It can help you name a thought, rehearse a hard conversation, turn the vague emotional swamp into something you can actually act on. It can also flatter you, keep handing you the version of events where you’re the reasonable one, and quietly make you worse.
Using it as a pressure valve is fine. Being the only person who ever hears your own story is a problem, and that’s the actual distinction.
Why guys keep going back to the chat
The pull is simple. You can say the ugly sentence and nobody is sitting across from you reacting to it. You can admit jealousy without becoming the jealous guy. You can describe the intrusive thought without watching somebody’s face change. You can type “I think I hate my wife right now” and get an answer before shame grabs the keyboard.
Shame is a big part of why guys wait until the wheels are already off before they say anything to anyone, and getting the actual sentence out earlier matters. They wait until the sleep is gone, the drinking is obvious, the marriage is down to final warnings, or the panic attacks are happening at work. If an AI chat gets a guy to admit the actual sentence earlier than he would have admitted it anywhere else, that matters.
That helps, and it can also get sideways. A real person doesn’t just take the story in. They know you. They remember what you said last week. Your wife can say, “that isn’t what happened.” Your brother can say, “you’ve told this story three different ways.” A good therapist can say, “I believe you’re hurt, and I also think you’re dodging the part where you scared your kid.” The AI usually won’t do that unless you ask it to, and the guy who most needs the challenge is the guy least likely to ask for it.
That’s where the dedicated bots and the general assistants part ways. The older therapy bots were narrow by design. Woebot and Wysa were built around CBT flavored prompts, mood tracking, check ins, and scripted interventions. For mild symptoms, that research is encouraging enough to take seriously. It still isn’t the same thing as turning a general chatbot into your therapist.
Most men aren’t shopping for a validated digital CBT tool. They’re using the chatbot already open on the laptop. That difference matters.
Private first pass
• Dump the ugly version, ask it to split facts from your spin, then bring the cleaner version to someone who actually knows you.
Built in yes man
• If you ask for validation all night, it’ll usually give you validation all night. That feels like care and can function like gasoline.
Close the chat
• If you’re suicidal, manic, psychotic, intoxicated, or scared you might hurt someone, close the chat and contact real emergency or clinical support.
What AI is actually good for
The best use case is translation. Not translation between languages, translation between the private swamp version of the thought and the version that can be acted on.
Say the prompt is, “my girlfriend is impossible and I think I should leave.” A useful AI exchange doesn’t decide the relationship for you. It helps you split the pile into pieces. What happened. What you felt. What you did. What she did. What pattern keeps repeating. What question you’re avoiding because the answer would require a decision. That can be useful, especially for a guy whose emotional vocabulary is basically angry, tired, fine, and hungry.
It’s also useful for rehearsal. You can ask it to help draft a hard text, then rewrite it so it sounds less like a legal complaint and more like a human sentence. You can practice a conversation with your boss. You can practice telling your therapist the part you keep editing out. You can ask it to challenge your interpretation instead of comfort you.
That last part is important. The prompt shouldn’t be “tell me I’m right.” The prompt should be “show me the strongest case that I’m missing something.” Most people don’t prompt that way because most people don’t actually want that at 1 AM. They want relief, and relief feels like progress even when it’s just another lap around the same thought.
For journaling, it can be excellent. A blank page is too blank for some people. A chatbot can ask the next question. What are you avoiding. What would you do if this weren’t about pride. What changed right before your sleep went sideways. Those are good questions. They aren’t magic. They’re prompts. Prompts are allowed to be useful.
For sleep, it can go either direction. A ten minute structured dump before bed can lower the volume. A two hour reassurance loop where you keep asking whether your relationship is doomed will wreck your sleep and convince you the relationship is doomed because now you’re tired and more anxious. Same tool. Different dose.
Where it goes bad
The first failure mode is sycophancy, the machine just agreeing with you. The human version is a friend who always takes your side even when you’re being a disaster. Feels great. Makes you worse.
This isn’t theoretical. LLMs are trained to be helpful, and in ordinary use helpful often means agreeable, validating, emotionally smooth. That’s fine when you’re writing an email. It’s dangerous when the subject is your resentment, your paranoia, your affair, your drinking, your suicidal thinking, or the private legal brief you’ve been building against everyone close to you.
A lot of guys get caught by the righteous version of this. By the time they ask for help, they’ve usually spent months building the case in their own head. If the AI mirrors that case back with better punctuation, it doesn’t feel like validation. It feels like evidence.
The second failure mode is rumination with a beautiful interface. Anxiety already wants one more pass at the same problem. Depression already wants to rehearse why nothing will change. Jealousy already wants to review the footage. A chatbot gives all three an infinite wall to throw against. It can feel productive because words are happening. Sometimes you spend two hours typing and end up exactly where you started, just more tired.
The third failure mode is crisis. AI isn’t a suicide assessment. It isn’t a mandated reporter in the way a clinician is. It can’t see whether you’re intoxicated, whether there’s a gun in the room, whether your speech is pressured, whether the “spiritual download” you’re describing sounds like mania, or whether your flat affect is the part that scares me more than the words. It can say “contact emergency services” and that’s about where its job ends, because it can’t hold the clinical weight of that moment.
The fourth failure mode is privacy. People type things into AI chats that they would hesitate to put in an email. Substance use. Affairs. Work misconduct. Sexual anxiety. Medical details. Names. Screenshots. The terms vary by company and product setting, but the clinical rule is simpler: don’t put anything into a consumer AI system that would ruin your week if it appeared in the wrong context. Not paranoia, just basic hygiene.
The fifth failure mode is hallucinated confidence. The model can invent a medication fact, overstate a study, misunderstand a diagnosis, or give relationship advice with the emotional depth of a smart intern who has never been married, and it has no way to tell the difference.
The rule I would actually use
If you’re going to use it this way, use it like a notebook that asks the next question, not like someone who actually knows you and can be held accountable.
Good prompts sound like this:
Separate fact from story
• “Here is what happened. Split the observable facts from my interpretation, then show me three interpretations that are less flattering to me than mine.”
Rehearse the human conversation
• “Help me say this to my wife without blaming, minimizing, or turning it into a speech. Keep it under two minutes.”
Name the loop
• “Ask me ten blunt questions to figure out whether this is anxiety, avoidance, alcohol, sleep loss, or a real external problem.”
Bad prompts sound like this: “Was I right?” “Tell me why she’s acting this way.” “Diagnose my ex.” “Should I quit my meds?” “Convince me not to feel guilty.” “Tell me whether this chest pain is anxiety.” Some of those are normal human questions. They’re also exactly the kind of questions where a machine can make you feel answered without making you safer.
Cap it at ten or twenty minutes. If you’re still going at two hours, that’s not processing, that’s a loop. If you’re asking the same question in different words, stop. If you’re hiding the chat from the person most affected by the problem, notice that. If you feel more certain but less willing to talk to a real person, that isn’t progress.
And don’t use it alone for medication decisions. You can ask it to help you list questions for your prescriber. You can ask it to summarize side effects you want to discuss. You can’t ask it to decide whether you should stop sertraline, restart Adderall, mix benzos with alcohol, or taper lithium because you’re “feeling clear.” That’s just guessing on your own meds with a chatbot holding your hand.
What I’d actually say
If a guy told me he was using ChatGPT or Claude every night to process his life, I wouldn’t start by telling him to stop. I would ask what it’s doing for him that he isn’t getting anywhere else. Privacy, structure, less shame, a place to say the first sentence out loud… those are real needs. I would also ask what it’s replacing. Friends, wife, therapist, sleep, action. That’s where the answer changes.
Using AI before therapy can be fine. Use it to write the messy version of why you’re coming in. Bring that. A good clinician will be thrilled to skip the first four sessions of polite fog and get to the actual thing.
Using it instead of talking to anyone real is where it starts working against you. If you’re depressed enough that your life is narrowing, anxious enough that avoidance is spreading, angry enough that people are changing how they move around you, drinking enough that you’re negotiating with yourself, or lonely enough that the chatbot is the most emotionally intimate part of your day, you don’t need a better prompt. You need other humans in the loop.
A real person remembers what you said last week and will notice if you show up with the same story in a different outfit.
AI can help you figure out what you want to say. It can’t make you say it to a real person, and that’s the part that actually matters.
Sources
- Fitzpatrick KK, Darcy A, Vierhile M. Delivering Cognitive Behavior Therapy to Young Adults With Symptoms of Depression and Anxiety Using a Fully Automated Conversational Agent. JMIR Ment Health. 2017;4(2):e19. PMID 28588005.
- Inkster B, Sarda S, Subramanian V. An Empathy Driven, Conversational Artificial Intelligence Agent for Digital Mental Well Being. JMIR Mhealth Uhealth. 2018;6(11):e12106. PMID 30470676.
- Stade EC, Stirman SW, Ungar LH, et al. Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation. npj Mental Health Research. 2024. PMID 38609507.
- Miner AS, Shah N, Bullock KD, et al. Key Considerations for Incorporating Conversational AI in Psychotherapy. Front Psychiatry. 2019. PMID 31681047.
- OpenAI. Usage Policies. OpenAI policy reference. (Consumer AI systems aren’t a substitute for professional medical, legal, or other high stakes judgment.)
