A Case for Optimism About AI in Mental Health

I am optimistic about what AI could make possible in mental health care. The prospect of helping a therapist develop a richer understanding of someone over time, and use that understanding to provide better care, seems to me a worthwhile future to work towards.

What interests me is the possibility of clinical knowledge becoming more cumulative. A useful observation from six months ago could become easier to recognise when it matters again. A treatment decision could be revisited in light of what subsequently happened. More of the understanding developed through therapy could remain available to support the next conversation.

At the centre of that future is the therapeutic alliance. Being heard and understood by another person is among the most human experiences we have. Research gives us further reason to protect it: a meta-analysis covering 295 studies and more than 30,000 patients found a consistent positive association between the therapeutic alliance and psychotherapy outcomes. I want technology to support the conditions in which that relationship develops, while leaving the relationship itself between the patient and their therapist.

I feel that these positive possibilities deserve a more serious hearing than they often receive. Clinicians encounter accounts of harmful chatbot conversations alongside predictions about machines threatening humanity. Those concerns differ substantially, yet they can merge into a general sense that bringing AI anywhere near mental health care is an unreasonable risk.

We should be able to examine those concerns carefully while remaining ambitious about what we could improve.

Why I think optimism is rational

David Deutsch’s writing has influenced how I think about progress. What I find compelling is his emphasis on our ability to create knowledge and correct errors. In his account of knowledge and human problem-solving, our present limitations need not define what will always be possible.

That gives optimism a practical meaning. When we encounter a problem, we have a reason to investigate it rather than assume it is an unavoidable condition of life. We may lack the knowledge to solve it today, and our first attempts may fail. Neither establishes that improvement is beyond us.

The history of printing offers a useful comparison. Making ideas easier to reproduce changed how knowledge could circulate and accumulate. People could encounter explanations they might otherwise never have seen, examine them and contribute something of their own.

Printing also changed the work involved in making books. The Victoria and Albert Museum’s account of medieval writing and printing describes how early printed books still relied on scribes and illuminators for parts of their production, even as printing made copies cheaper and easier to produce. Established skills acquired different roles as the technology developed. The prospect of wider benefits does not make concern about your own profession unreasonable.

Greater access to ideas also meant greater opportunities to circulate falsehoods. The capability created possibilities that people then had to learn how to use, alongside practices for challenging claims and checking their sources.

AI presents a different set of problems, so the comparison has limits. What carries across is the possibility of expanding our capacity to work with knowledge, along with the responsibility to improve how that capacity is used.

I think that warrants genuine enthusiasm. An honest conversation about technology should make room for what we might achieve, with willing participants prepared to examine the benefits as seriously as the harms.

The opportunity is a deeper understanding of the patient

Much of the discussion about AI for therapists begins with documentation. The burden is familiar, and reducing it could make a meaningful difference to a clinician’s working day. My interest extends further into how understanding develops throughout care.

A therapist learns about someone gradually. An early formulation changes as more of the person’s history becomes available. An approach that seemed helpful may need reconsideration after their circumstances change. Something mentioned briefly in one session can acquire significance months later.

The clinician holds much of this understanding, supported by records that may be difficult to review as a coherent whole. Under the ordinary pressures of practice, recovering the reasoning behind an earlier decision can take considerable effort.

At Citt.ai, our purpose is to help clinicians build on what they learn throughout a patient’s care. The ambition extends beyond making records easier to manage, into making the understanding within them more useful. It is central to our view of how AI can strengthen continuity in therapy.

A living patient graph gives that ambition a concrete form: connecting relevant observations across a person’s care while retaining where they came from. A patient’s account of an experience must remain distinguishable from the clinician’s interpretation, with AI-generated suggestions clearly identified for review.

Consider someone whose difficulties return after a period of improvement. Bringing the relevant history together could help their therapist revisit what appeared to help before, and whether the original explanation still fits. It could also draw attention to information that challenges the current formulation.

That could support diagnostic reasoning and treatment review without giving the software authority to determine either. The value would come from helping a clinician ask a better question or reconsider an assumption, with enough context to judge the suggestion.

This is a more substantial ambition than producing additional notes. It concerns how the knowledge developed through care can remain available to improve care.

Strengthening the therapeutic alliance

A patient should never have to become more legible to a machine at the expense of being understood by their therapist. Any clinical representation will be incomplete, and the person must have room to surprise it.

Used thoughtfully, however, better access to relevant context could support the relationship. A therapist may be able to return to something important without asking the patient to reconstruct the whole history. Time spent searching through records could become time available for listening.

The patient’s ability to disagree matters here. A description that once seemed accurate may no longer fit, or may never have captured their experience properly. That disagreement should change the record rather than be treated as an inconvenience to the system.

The same principle applies between sessions. There may be useful ways to help someone practise skills agreed with their therapist and bring relevant experiences back into the next appointment. The purpose would remain connected to the person’s care and life outside the software. As we have argued in our writing on designing mental health AI around disengagement rather than engagement, encouraging an ongoing emotional dependence on a machine would take us in a direction I do not want to pursue.

The therapeutic alliance is part of what gives this vision its value. It deserves to remain central as the technology develops.

Human and artificial fallibility give us something to work on

Clinicians have limited attention and can become attached to an explanation that once made sense. AI introduces its own weaknesses, including invented connections and confidence unsupported by the underlying information.

Combining them will only help if the workflow makes those errors easier to recognise. A model that repeatedly agrees with a clinician could reinforce an existing mistake. A polished summary could conceal uncertainty that was obvious in the original record.

This is where deliberate design matters. Important suggestions should be traceable to their sources. Conflicting information should remain visible, and a correction should carry through to later summaries.

Over time, a clinician could compare earlier expectations with what actually happened. That could help them recognise when an approach needs revisiting and build a more informed understanding of their own decisions.

That is the compounding learning I find promising: reviewed knowledge accumulating within care, with the ability to change it. It does not imply that private patient information can be freely reused to train models or applied to other patients without an appropriate basis.

There is real potential in designing systems around the fallibility of everyone involved. We can aim for a combination that helps each identify mistakes the other might miss, while evaluating whether that actually happens.

Keeping the AI debate specific

Artificial general intelligence generally refers to broadly general capabilities, although definitions remain contested and predictions about its arrival are uncertain. The possible consequences deserve serious study.

Those uncertainties do not settle whether a particular clinical application is useful or appropriate today. We can evaluate an application without first resolving whether machines could become conscious or what a future AGI might mean for humanity.

Present-day systems can already cause harm, including when they are given authority to act through external tools. The relevant assessment concerns the capabilities and permissions involved, as well as who is exposed to their outputs.

A system executing changes across a network presents a different risk from one offering source-linked information for professional review. Patient-facing support needs a further assessment because of the circumstances in which people may rely on it.

I would like these discussions to remain grounded in identifiable mechanisms of harm and evidence about how well controls work. General concern about the future cannot substitute for that examination, any more than enthusiasm about AI can establish that a product is ready for use.

For clinicians considering a tool, understanding what a mental health AI safety evaluation can actually establish offers a more practical starting point than treating every application as an instance of the same hypothetical future.

Responsible building is part of the positive story

Competition at the frontier creates pressures that deserve scrutiny. Companies need to recover substantial investments, while governments consider the strategic significance of increasingly capable models. Speed can receive a more immediate reward than caution.

Independent evaluation and meaningful accountability matter under those conditions. Researchers need credible ways to raise concerns, and liability should help prevent the consequences of negligent deployment from falling entirely on those affected.

Technical restrictions matter too. Instructions asking a model to behave appropriately cannot replace enforceable limits on what it can access or do.

Application developers share these responsibilities. For Citt.ai, responsibility includes how models are used within the product and the boundaries around their outputs. Confidence in the work means being prepared to demonstrate its value and respond when something needs improving. Clinicians should expect evidence appropriate to the claims being made, with a clear distinction between useful functionality and demonstrated clinical outcomes.

A future worth building towards

The mental health system we have today leaves considerable room for improvement. Clinicians’ time is constrained, and people can struggle to access appropriate care. Those conditions carry consequences of their own.

AI offers possibilities for making available expertise go further and helping the understanding developed through therapy become more useful over time. If we can achieve that while preserving the therapeutic alliance, the contribution could extend well beyond administrative efficiency.

Clinicians have an important role in shaping this future. Meaningful engagement means bringing clinical judgement into the discussion and examining the technology closely enough to influence how it develops. If those who understand therapeutic care step away entirely, more of the decisions will be left to people with less understanding of what needs protecting.

This is why I feel optimistic. We have an opportunity to investigate problems that matter, using capabilities that are still developing. We can involve clinicians and patients in shaping that work, learn from failures and change approaches that do not justify themselves.

I would like the public conversation to make more room for that possibility. A future in which technology helps people provide better care is worth imagining clearly enough that we can begin to build it responsibly.

To explore how this approach informs the product, see Citt.ai’s therapist-led care platform and the features supporting clinical workflows.