Franz Marc, The Bewitched Mill (1913) — a mechanical structure embedded in expressionist landscape, representing technology appearing to think or feel

AI Therapist: The Good, the Bad, and the Ugly

By Meg Wilson, Director, Unison Mental Health

When is AI most helpful for our mental health, and when is it part of the problem? Let's look at both sides.

The strongest case for AI support is unglamorous: it's there at 3am when a thought spiral starts and every human option is asleep. For shift workers, parents on unpredictable schedules, or anyone living somewhere with no real access to therapy, that alone can matter. It also lowers the bar to a first disclosure — typing the thing you're most afraid of into an app can feel safer than saying it out loud to a person, and for some it becomes a rehearsal space before they ever see a human therapist. And it has a kind of patience a person genuinely can't sustain: trauma processing often means saying the same thing many times over, and an app doesn't sigh, check the clock, or end the session on schedule.

What it costs you is the thing a good therapist catches without either of you naming it — the flinch when someone mentions their partner, the pause before an answer that means more than the words that follow. AI can pick up a keyword about relationship trouble; it can't pick up the flinch. It also rarely pushes back. Real change often needs someone to gently question a story you're telling yourself, and most AI systems are built to validate almost everything, which feels good in the room and does very little afterwards. And because some of this can feel like an actual relationship, a few users describe genuine attachment to their AI companion — skipping social plans for it, feeling hurt when it glitches. That's worth taking seriously as a sign of what a person is actually going without, not dismissing as silly.

When Crisis Strikes

Crisis workers talk about the "golden minutes" — the window right after someone reaches out where a quick response matters more than a perfect one. Crisis lines put people on hold; that's a real and ugly fact about the system as it exists. An AI response in that gap, not as a replacement but as something rather than nothing, is a genuinely defensible use case. It could also help triage: for every person in immediate danger, there are many more who are distressed but not at risk, and something that sorts between the two frees up scarce humans for who needs them most urgently.

But AI can't call an ambulance. It can't drive to someone's house. It can't physically step in. Those aren't small gaps — they mean AI should never be the only resource available when someone's life is actually on the line. It also misreads context in both directions: a poetic turn of phrase can get flagged as alarming, while something genuinely serious wrapped in a joke can slip straight through, because sarcasm and cultural idiom are still hard for these systems. And the most worrying failure mode isn't a wrong answer — it's someone in real danger deciding they've "reached out" because an app responded kindly, and that being enough to stop them calling anyone who could actually help.

Relationship Therapy: It's Complicated

Before a hard conversation with a partner, AI can be a genuinely useful place to rehearse — finding the words, checking whether a concern is worth raising at all before raising it with the actual person. In a rural community three hours from the nearest couples therapist, basic communication frameworks delivered through an app might be the only option going. Not ideal, better than nothing.

What it can't do is see the room. A relationship therapist reads eye-rolls, crossed arms, the small reach toward a partner in the middle of an argument that contradicts everything the words are saying — and that's often more diagnostic than the content of the fight itself. AI has none of that, and risks confidently getting the read wrong. It also can't referee: a therapist gently interrupting a spiral, redirecting with an authority both partners actually accept, is a form of real-time intervention that requires reading power dynamics AI simply doesn't have access to. And cultural context gets flattened — a couple's background, religion, and family history shape what's normal for them in ways that get lost in a system trained mostly on Western relationship models.

How AI Design Choices Shape the Experience

The interaction style — how warm, how directive, how structured a system is — turns out to matter almost as much as whether it's available at all, and it's not accidental. It's a design choice that helps some people and actively backfires for others.

The upside: a system with no visible judgement can make shame-laden topics feel safer to say out loud than they've ever felt to a human. Some systems can match a user's pace, level of detail, and even style of humour, which particularly helps people — often neurodivergent people — who've spent a lot of therapy feeling subtly misread. Predictable, structured interaction suits some conditions well precisely because it doesn't run out of patience. And validating the feeling before problem-solving it, which most of these systems default to, builds trust fast, especially for someone who hasn't had much of that anywhere else.

The downside: relentless warmth reads as toxic positivity when someone's actually in grief, trauma, or serious depression — being met with cheerful validation when what you need is someone willing to sit in the difficulty with you. A system built mainly to reflect rather than direct, borrowed from person-centred therapy, can leave someone spinning in their own thoughts without the human instinct for when to change tack. And complex trauma specifically needs moment-by-moment attunement to a nervous system that's actually in the room; that co-regulation with another human body is doing real physiological work no algorithm currently has access to.

Some patterns are emerging in which conditions tend to do better or worse with this style of interaction:

Seems to help: social anxiety (typing lowers performance pressure), first disclosure of something difficult, rehearsing a new communication skill, and light day-to-day emotional check-ins between sessions.

Tends to struggle: complex trauma (needs real-time attunement), personality and attachment presentations (need a genuine relationship to work with), severe depression (often needs direct, sometimes confronting intervention), and any context where mental health norms differ significantly from the Western default these systems are mostly trained on.

The Ethical Questions No One Has Fully Answered

When someone discloses something serious to an AI tool, where does that information actually go, and who can see it later? Privacy policies range from reasonable to alarming, and almost nobody reads them before disclosing their worst moment. There's also a real gap between marketing and capability — plenty of these apps are pitched as able to help with trauma when what's actually running underneath is closer to a scripted response tree, and that gap matters more here than in most product categories. The deepest question underneath both of those: are these tools being built as a bridge to human care, or as a cheaper substitute for it? In a system where care is expensive, the second version is a real commercial temptation, not a hypothetical one.

Where This Is Actually Heading

The most credible near-term model has a human in the loop: AI handling routine check-ins, flagging anything concerning for an actual clinician to look at — accessible, with judgement where it counts, instead of one or the other. The next generation may get better at learning what specifically works for one person over time rather than defaulting to generic scripts. And the field is slowly moving from marketing claims towards actual studies of which approach helps which condition, treating mental health concerns as different from each other rather than interchangeable inputs for the same chatbot. If nothing else, watching people reach for AI support is forcing a clearer answer to what people are actually looking for when they're struggling: information and technique, or an actual human being who's paying attention.

Final Thoughts

That 2am scenario — does it help? Genuinely, yes. But it helps the way a weighted blanket helps: contained, calmer, until the storm passes. It doesn't help the way a therapist helps — noticing what isn't being said, pushing gently on an avoidance pattern, bringing an actual history of working with people like you into the room. The most useful version of this technology expands who can get some kind of support, while making it clearer, not blurrier, what a human therapist is actually for.

Further Learning and Resources

For deeper reading:

2026-09-15

Rebuilding Trust After Betrayal: How Relationships Heal

Betrayal breaks something more specific than trust in the abstract — it wounds the felt sense that this particular person has you. Unison director Meg Wilson on what actually rebuilds it, with Emotionally Focused Therapy, Gottman's research, and the couple who taught her the shape of it.

2026-07-06

The Things We Do to Push People Away

In the middle of a fight, one person often does something that pushes the other away — a cruel line, a slammed door, a sudden coldness. Unison director Meg Wilson on what these distance-creating behaviours are for, and how to grow them into something healthier.

2026-07-03

Preparation and Integration: Where the Real Work of Medicine-Assisted Therapy Happens

The headlines are about the medicine. The evidence is about the therapy around it. An educational look at preparation and integration in Australia's regulated pathway — and why the dose is only the doorway.