AI Is Now Sitting In On Your Therapy Session, We Should Probably Talk About That

Therapy was supposed to be the last room AI walked into – too human, too nuanced, too dependent on the kind of presence that no model trained on text could replicate.

That assumption is now officially out of date. AI is inside mental health consultations in real time, transcribing sessions, tracking sentiment, flagging risk and nudging clinicians toward specific interventions while the patient is still sitting across from them.

According to a Practitioner Pulse Survey, 29% of mental health practitioners now use AI tools at least monthly in their practice – nearly one in three. And the tools aren’t staying in the background, some implementations are feeding diagnostic nudges directly into clinical decision-making during live sessions. That’s not a pilot programme or a research trial – that’s how care is being delivered right now, in practices where patients have no idea it’s happening.

The debate about whether AI should be in therapy is no longer a debate, it’s here. The debate we need to be having now is about what happens when it gets something catastrophically wrong, and who carries that.

The Gap Between What AI Claims And What It Actually Delivers

Let’s start with what the evidence actually says, because the gap between the pitch and the reality is bigger than the funding announcements suggest.

Research comparing AI-augmented and traditional therapy found that standard human-led therapy achieved roughly 45 to 50 percent reductions on common anxiety and depression scales. AI-driven approaches came in at around 30 to 35 percent. That’s a meaningful shortfall in the outcomes that matter most.

AI improves throughput, reduces documentation burden and can flag patterns across sessions that a time-pressured clinician might miss. Those are meaningful benefits, but the honest version of the AI-in-therapy argument acknowledges that access and efficiency are not the same thing as quality of care. Scaling the former at the expense of the latter is a choice with real consequences for real patients.

The reality is that the people most likely to end up receiving AI-mediated mental health care are the people who can’t afford an alternative, and that should give investors more pause than it currently seems to.

Nobody Is Actually In Charge Here

Here’s the accountability problem, stated plainly: when an AI flags a patient as high risk and that flag shapes a clinical decision that turns out to be wrong, who is responsible? Right now, the honest answer is unclear, and that lack of clarity is a feature of the system.

Most regulators still classify AI as a support tool rather than a co-decision-maker, even as real-world deployment is pushing steadily toward shared decision-making. That classification gap means clinicians are on the hook for outcomes that are increasingly shaped by opaque algorithmic signals they may not have fully understood or been able to override.

The data privacy dimension makes it worse. When the most sensitive conversations of your life are being transcribed and processed by third-party systems with terms of service few people read, patient scepticism isn’t a failure of digital literacy. It’s a reasonable response to a system that hasn’t earned trust yet.

Surveys of mental health professionals identify data privacy concerns and fears of algorithmic bias as the top barriers to broader patient acceptance of AI-assisted care – that’s the sector telling you something. Whether investors and builders are paying attention is less obvious.

The Money Keeps Coming In Anyway, Here’s Why.

The chronic shortage of mental health clinicians is well documented, and wait times stretch into months in most developed economies. The cost structure of high-quality, sustained therapy makes it inaccessible for a huge proportion of the people who need it most.

AI promises to compress documentation time, harden adherence to evidence-based protocols and make each clinical hour go further. For health-tech investors, that problem statement is irresistible, and the capital flows accordingly.

What’s considerably harder to defend is the pace at which deployment is outrunning accountability. The AI upskilling conversation happening across industries hasn’t caught up with what it means to deploy AI in environments where the stakes are this high.

Most sectors can absorb a badly designed AI tool with inconvenience and wasted money, but in mental health, the cost of getting it wrong isn’t a poor quarterly metric. It’s a patient in crisis who received the wrong intervention because an algorithm was overconfident.

If You’re Building In This Space, The Bar Is Higher Than You Think

The pattern in high-stakes AI follows a sadly familiar arc: the technology arrives, the funding floods in, and the ethical standards, liability frameworks and regulatory guidance trail behind by years, with patients carrying the risk in the gap.

For anyone building AI in regulated, sensitive domains, the mental health space is a useful and instructive mirror. The minimum standard isn’t a clever model: it’s AI suggestions that are fully auditable, with every flag and every nudge logged so it can be reconstructed after the fact.

Clinicians need to be able to overrule the algorithm without it counting against them. Patients need consent that’s understandable and not buried in a privacy policy – details about what’s recorded, how it’s used and what they can opt out of. And the founders building these tools need to be willing to hold those standards even when doing so slows the product down.

The real scandal isn’t that AI is in the therapy room – it’s that it arrived before anyone agreed on the rules. This wave of investment will be judged not by how good the models get, but by whether the people building them treat clinical humility as a requirement rather than an afterthought.