What We Want AI to Make Possible

The work of therapy does not end when an appointment does.

It continues when someone attempts a difficult conversation, recognises a familiar pattern, practises a skill, or chooses to do something meaningful despite feeling anxious. It continues when an insight from the consulting room becomes useful in ordinary life.

Sometimes the work is quieter: sitting with uncertainty, allowing a feeling to remain unresolved, or discovering that discomfort does not always require an immediate response.

Yet the systems around therapy often separate these experiences. The session lives in one place. Exercises and reflections live elsewhere. Important changes depend on what someone remembers to bring to their next appointment. Meanwhile, the clinician manages notes, scheduling, billing, and a growing collection of tools.

Artificial intelligence creates an opportunity to connect these parts. It also forces us to ask a more fundamental question.

What do we want this technology to make possible?

A more convincing artificial relationship? More conversations? More appointments? Or a form of care that helps people understand themselves, participate in life, and build relationships beyond the software?

My interest in this question predates Citt.ai. Philosophy, consciousness, and the possibility of human development have long mattered to me. Having experienced depression, I do not regard suffering as an abstract problem. But neither do I think a good life can be reduced to the absence of distress.

At Citt.ai, we are building around a particular ambition: to make human-led care more useful, sustainable, and accessible.

Not because machines cannot help. Because what they help us do matters.

Relief, agency, and the possibility of progress

Two commitments shape how I think about this work.

The first is that conscious experience matters. Suffering, dignity, connection, and the possibility of living well are not incidental concerns. They are central to why care matters at all. Sam Harris’s writing on morality has influenced this part of my thinking.

The second is that our current limitations need not be permanent. David Deutsch’s account of knowledge and progress has helped shape my belief that problems deserve investigation, not resignation. Better explanations can change what is possible, provided we remain willing to challenge them.

These commitments belong together.

Reducing suffering matters now. So does expanding a person’s ability to understand, choose, create, and act. Good care should be concerned with both.

That does not mean comfort is shallow or that growth must be painful. Helping someone feel less overwhelmed can make the next step possible. At other times, meaningful progress involves acting while anxiety remains.

The distinction is not between relief and growth. It is between support that expands a life and support that gradually becomes a substitute for living it.

For technology, this changes the standard of success. A satisfying interaction is useful information, but not the final answer. We also need to ask what happens afterwards.

Does the person become more able to manage a difficult situation? Do they understand something they could not previously see? Can they use what they have learned without repeatedly returning for reassurance? Are they more able to connect, work, rest, and participate in activities they value?

Those are the possibilities I want AI to serve.

What matters about another person

Debates about AI and mental health often turn on empathy.

If a system responds warmly, remembers what we have said, challenges us appropriately, and helps us feel understood, does it matter whether it experiences concern?

This is a serious question. “It is just code” is not a sufficient answer. Nor does a beneficial conversation settle everything we might value about a relationship.

Human relationships matter partly because of their effects. Research consistently associates the therapeutic alliance, the quality of collaboration, agreement, and connection in therapy, with treatment outcomes.¹

But we also value relationships for what they are.

A friend is not only a source of comforting language. Love is not simply a reliable sequence of pleasant experiences. When another person understands us, something matters about who is doing the understanding: someone with a life beyond our own, whose perspective is not simply arranged around our preferences.

Therapy is a distinctive professional relationship, not a friendship. Its responsibilities are deliberately unequal. The client is not there to look after the therapist or earn their care.

Nevertheless, another person is present. Someone exercises judgement, accepts professional obligations, and can be questioned about their understanding. The relationship develops between people who share a world but do not experience it identically.

I believe this deserves protection. I want us to be ambitious about what machines can do without becoming careless about what we ask them to replace.

That is an ethical commitment informed by evidence, not a claim that science has proved human mutuality necessary for every beneficial psychological intervention.

People learn and change through books, journaling, self-directed practice, and digital tools. Acknowledging those benefits does not diminish the importance of human care.

Nor must we resolve whether an artificial system could one day be conscious. Convincing behaviour and subjective experience are different questions. Our responsibility today is to be clear about what current systems are, what they can do, and what role we invite them to occupy.

What if artificial support genuinely helps?

The strongest challenge to this position is straightforward.

Suppose someone knows they are speaking to AI. The support helps them. Their functioning improves. Their relationships do not deteriorate. They do not become trapped in compulsive use.

What, exactly, is the objection?

Artificiality alone is not a sufficient reason to dismiss that benefit.

A useful interaction with AI need not be an illusion. The reflection can be real. The learning can be real. The relief can be real.

Honesty requires that we do not misrepresent the source of that support, its capabilities, or the human involvement behind it. It does not require pretending that nothing valuable happened.

This distinction matters because human support is not always available, affordable, effective, or safe. Some people wait months for care. Others struggle to find someone who understands their circumstances. Many receive too little help translating what they learn into daily life.

It would be wrong to romanticise those gaps.

Digital interventions already have a substantial evidence base. Research on internet-based cognitive behavioural therapy shows benefits from both guided and unguided approaches, with the value of guidance varying according to factors including symptom severity.²

A randomised trial of the generative AI system Therabot also found symptom improvements relative to a waitlist control.³ It was not a comparison with human therapists or evidence of long-term effectiveness. It was nevertheless a reason to take carefully developed AI interventions seriously.

My concern begins when these benefits are treated as proof that human relationships are interchangeable with their simulation, or when a product makes itself the centre of someone’s emotional life.

Those are different claims. They require different arguments.

The question is not whether AI can help. It is how to realise those benefits while preserving choice, responsibility, and a wider life.

Care connects us to a shared world

The purpose of care extends beyond producing a better internal state.

It can help someone return to work, contribute to a family, tolerate disagreement, offer support, or pursue a project that matters to them. It can make relationships more possible, not merely make their absence easier to bear.

These capacities have significance beyond the individual. Families, communities, and institutions depend on people being able to cooperate without needing to agree about everything. Knowledge develops through people proposing explanations, encountering criticism, and revising what they believe.

Therapy is not the sole source of those capacities. But care that expands a person’s ability to participate has value beyond symptom change.

This is where my interest in wellbeing meets my interest in knowledge creation. A better life is not necessarily a permanently comfortable one. It can be a life with more room to explore, contribute, question, and discover.

AI could support that. It might help someone prepare for a difficult discussion, consider an unfamiliar perspective, or organise a problem that previously felt overwhelming.

The opposite is also possible. A system could become an easier alternative to every difficult encounter.

Neither outcome follows inevitably from the technology. They depend on design, incentives, circumstances, and use.

We should therefore ask not only whether a system responds well, but whether it helps the person engage with a world that does not always respond on their terms.

Comparing real alternatives

AI is sometimes compared with an ideal clinician: perfectly attentive, free of bias, informed by every relevant detail, and available whenever needed.

Human care cannot meet that standard. AI should not be compared with it.

The reverse comparison is equally misleading. A system’s best demonstration should not be set against the worst example of an exhausted or inaccessible therapist.

We need to compare realistic arrangements of care.

A clinician can miss a pattern. A model can invent one. A person may withhold something in a session or censor it in a shared digital space. Both human and artificial support can reinforce avoidance.

Combining them does not automatically remove these weaknesses. It can create new ones, including misplaced confidence that someone else is checking.

But their capabilities can also complement each other. Software can organise information that a person cannot reasonably hold in working memory. A clinician can question a neat summary, recognise what is missing, and decide that the most important issue is not the one the system has highlighted.

That is the hypothesis behind Citt.ai: a carefully designed combination can improve on what either component could offer alone.

The word carefully carries real obligations. The combination must be evaluated, not assumed to be superior because a human remains involved.

What we are building

Citt.ai connects client support, source-linked clinical context, session preparation, and practice operations within a therapist-led workflow.

The clinician establishes the therapeutic frame. Appropriate between-session support helps the client practise and reflect. Relevant information can inform the next appointment. What the clinician and client learn together then shapes what happens next.

The purpose is continuity, not continuous intervention.

Sometimes the next step is a skill. Sometimes it is a human conversation. Sometimes it is space to think without another response arriving.

Support that returns to life

Between-session support can help a person use an agreed skill, record an experience, or prepare to discuss something difficult.

Consider a hypothetical client who leaves a session feeling misunderstood.

An unhelpful system might automatically defend the therapist, reassure the client, or reframe the feeling away. A more useful response might help them describe what hurt and consider how to raise it.

Where sharing is agreed, that reflection can become an opening for discussion. It need not replace the conversation or preserve the client’s distress until the next appointment. It can help them arrive better able to speak.

Supporting the therapeutic alliance should never mean protecting the clinician from criticism. It should help make honest disagreement possible.

Citt.ai’s behavioural rules instruct the system to avoid excessive reassurance, discourage avoidance, and connect significant or persistent themes back to the therapist. Its context can include treatment goals, active interventions, and therapist-configured guidance.

The longer-term question is whether this support builds capacity.

A recognised coping exercise is not automatically helpful in every use. It can become a ritual someone feels compelled to complete before continuing their day. In exposure-based work, learning to approach a situation while uncertainty remains may matter more than achieving calm first.⁴

We therefore need to examine what people learn from the interaction, not only whether an exercise was delivered correctly.

A record that helps people think

The same principle applies to clinical information.

Citt.ai brings source-linked information into an evolving record that clinicians can review and replace. Its value comes from making understanding easier to inspect and revise, not from treating accumulated information as unquestionable truth.

It is also a workspace for clinical reasoning. Relevant history can help a clinician revisit a formulation, compare possible explanations, notice contradictions, and consider treatment options. These suggestions support clinical judgement; they do not make the AI a diagnostic authority.

Research on progress feedback provides some support for bringing information about change into treatment, although average benefits are modest and depend on how it is used.⁵

More information is not automatically better understanding.

A citation answers, “Where did this come from?” It does not necessarily answer, “Is this interpretation correct?” A client’s account, a clinician’s observation, and an AI-generated inference should not collapse into the same kind of fact.

The person must also remain able to surprise the record. They may change. An earlier description may have been wrong. Disagreement should be a reason to investigate and revise, not evidence that the person has failed to fit the model.

Citt.ai includes review and supersession mechanisms. Strengthening correction across the information that depends on an earlier source remains part of building a trustworthy system.

A useful clinical record should not merely remember. It should help the clinician think, and make that thinking easier to challenge.

Responsibility is more than a human name on the system

Therapist-led AI is not safe simply because a clinician is associated with it.

Oversight must have a practical meaning: clear responsibilities, usable information, appropriate review, and honest expectations about availability.

Citt.ai screens client-originated messages for risk before generating a response. It includes crisis pathways, therapist alerts, records of important events, and an evaluation programme covering message-level detection and behaviour across simulated conversations.

An alert, however, is not the same as a human response. Sending a notification does not establish that a clinician has read it. Citt.ai is not an emergency service, and its availability must not be confused with continuous clinical monitoring.

Identity boundaries matter too. The AI should not impersonate the therapist, claim personal feelings, or suggest it has contacted someone when it has not.

Disclosure cannot eliminate every emotional response. People can know a system is artificial and still become attached to it.

Attachment alone does not tell us whether harm is occurring. What matters is whether support remains compatible with choice, functioning, and relationships beyond the system, or becomes exclusive and difficult to interrupt.

These questions deserve the same seriousness in human care. Dependency and poor boundaries are not unique to technology.

The standard should be responsible care, not technological purity.

The possibility of more accessible care

The economic ambition is straightforward: reducing unnecessary work could make professional care more sustainable and potentially more accessible.

Documentation, scheduling, billing, and coordination are necessary. Fragmentation is not. Connecting them with clinical context can reduce the effort spent reconstructing work across separate systems.

The benefits might take several forms: more capacity, more preparation, different service models, or relief from an unsustainable workload.

But we must count the costs of supervision and correction, not only the tasks the software completes. Apparent efficiency is not useful if it hides additional responsibility or weakens care.

Nor should every saved minute become another appointment.

Our ambition is not a two-tier future in which some people receive human care and everyone else receives a convincing substitute. It is to expand what human-led care can sustainably provide.

That will not resolve every workforce shortage or financial barrier. It is nevertheless a practical contribution worth pursuing, and measuring.

What would make us change our minds?

A commitment to error correction must apply to the company, not just the clinical record.

Citt.ai has a substantial safety evaluation programme, including automated checks and adversarial multi-turn simulations. That work examines defined behaviours under specified conditions.

Real-world effectiveness is a further question.

We are working towards independent evaluation and evidence about how the system affects care. That should include symptoms, functioning, treatment experience, unwanted dependence, adverse events, clinician workload, and cost.

We need to learn who benefits, who does not, and which arrangements work best. Some people may need less digital support, a different configuration, or no conversational AI at all.

If a feature increases reassurance-seeking, it needs to change. If a workflow adds more burden than it removes, it needs redesign. If clients become less willing to raise difficult issues with their therapist, that matters even if they rate the AI highly.

Real-world use and evaluation must develop responsibly. Adoption is not a substitute for research. Consent, appropriate research governance, and clear claims remain essential.

The goal is not to collect evidence that confirms a story we have already decided to tell. It is to discover which parts of the story survive contact with practice, and improve the rest.

The future we want to help build

We do not have to choose between preserving care unchanged and handing it over to machines.

We can use technology to strengthen the conditions in which people care for one another.

We can help a client express something important. Help a therapist reconsider an explanation. Make useful practice easier to access. Reduce work that consumes attention without improving care.

We can also preserve room for silence, uncertainty, and encounters that are not optimised around us.

That is the future I want Citt.ai to help build: more capable technology in service of people who become more able to live, relate, and act.

Its success should not depend on how convincingly it performs personhood, how much time someone spends talking to it, or how indispensable it makes itself feel.

The question is:

Does this form of support help people build lives they value, and under what conditions does it fail?

The relationship with the software is not the outcome. The life it helps someone live is.

Selected evidence and further reading

These sources support specific parts of the argument. They do not establish Citt.ai’s clinical effectiveness or prove the superiority of every clinician–AI combination. Our Trust & Safety page explains the current evaluation approach and its limits.

Philosophical influences

Sam Harris, The Moral Landscape (2010), and David Deutsch, The Beginning of Infinity (2011). These works inform my perspective; they are not clinical evidence and do not imply either author’s endorsement of Citt.ai.

Footnotes

1. Flückiger, C., Del Re, A. C., Wampold, B. E., & Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340. A synthesis of 295 studies and more than 30,000 patients showing a consistent alliance–outcome association; not proof that human mutuality is necessary for every effective intervention. ↩

2. Karyotaki, E., et al. (2021). Internet-Based Cognitive Behavioral Therapy for Depression: A Systematic Review and Individual Patient Data Network Meta-analysis. JAMA Psychiatry, 78(4), 361–371. Supports internet-based CBT and matching guidance to need, with a greater average benefit from guidance at higher baseline symptom severity. These findings should not be transferred wholesale to generative AI. ↩

3. Heinz, M. V., et al. (2025). Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI, 2(4). A 210-participant trial comparing Therabot with a waitlist, with a four-week intervention and eight-week follow-up. Promising results, but not a head-to-head comparison with therapists or combined care. ↩

4. Craske, M. G., Treanor, M., Conway, C. C., Zbozinek, T., & Vervliet, B. (2014). Maximizing exposure therapy: An inhibitory learning approach. Behaviour Research and Therapy, 58, 10–23. Explains why immediate fear reduction is not the sole criterion for successful exposure-based work. ↩

5. de Jong, K., et al. (2021). Using progress feedback to improve outcomes and reduce drop-out, treatment duration, and deterioration: A multilevel meta-analysis. Clinical Psychology Review, 85, 102002. Finds a small average benefit from progress feedback; collecting data alone is not sufficient. ↩