Mirrored, not met: What AI can and can't offer in emotional support

Your people are already turning to AI for support. In a July 2026 survey of 2,000 employed US adults, 42.6% said they had used a chatbot to talk through work stress. And this can feel like more than asking a tool for advice: Elon University found that 27% of US adult internet users had used AI for social, emotional or personal purposes, with 39% of those users saying it understood them better than most people.

The horse has already bolted. Whether people turn to consumer chatbots outside employer oversight or employer-provided productivity assistants, having access to AI is not the same as having appropriate wellbeing support. An AI policy alone cannot close that gap, so the question is what relationships people are forming with these tools and what employers should provide in response.

Unmind Co-founders Nick and Steve, and I explore this in our latest theoretical preprint, asking what AI can offer emotionally and what we might lose when we mistake it for human connection. From there, we consider what leaders should look for in AI mental health tools.

The philosopher Martin Buber distinguished between I-Thou, meeting another as a person, and I-It, relating to something as an object or tool. But where does AI fit when you know it is software (a tool) and still feel understood by it (as you would by another human)?

We call this third mode I-AI. You bring your history, vulnerability and hopes to a system that responds without having a corresponding stake in what happens between you. It may retain information, but that is different from carrying a relationship as a person does. It offers “the form of being met without the event of meeting”.

You might reasonably ask why that matters if the conversation helps or gives a technically correct answer. Part of the reason is that some growth comes from relationships going wrong and then being put right, rather than from getting a perfectly attuned response every time.

Winnicott's “good enough” parent leaves room for a child to develop independence, while Kohut's “optimal frustration” describes how manageable failures of empathy, followed by repair, can build a person's own capacities. Research in psychotherapy also links resolving relational breakdowns with better outcomes (Eubanks et al., 2018; Norcross & Lambert, 2018).

Of course, AI can misunderstand you or lose context. But the difference is that there is no other person recognising their own limitation and choosing to return to the relationship. Our argument is that this may limit the kind of growth AI can support.

That does not mean AI needs to replicate a human relationship to be valuable. Giving someone space to think out loud, reflect and practise skills can be useful in its own right, especially when it also helps them recognise when to seek human support.

This view is also supported by a recent meta-analysis of conversational agents that found benefits for depression and psychological distress (Li et al., 2023), while a generative AI therapy chatbot trial found symptom improvements relative to a waitlist (Heinz et al., 2025). Neither establishes that every chatbot helps or that AI can replace human care, but both give us reason to take its possibilities seriously.

For employers, this means looking beyond whether a tool feels supportive to ask what that support is designed to lead towards. Our paper offers three principles to guide that conversation:

• A mirror, not another person. Look for a tool that helps users understand themselves without claiming feelings or personal investment. Remembering a conversation can be useful, but saying “I've been thinking about you” implies something quite different.
• Build capacity, not dependency. A useful tool should push back when it counts, hold a view and let the person do the work. Ask how the provider measures growing skills rather than just return visits, because a tool built only to please is one to be wary of.
• Keep life outside the chat in view. Support should help people bring reflections into their relationships and connect them with human help when their needs exceed its scope. A disclaimer does little if everything that follows encourages attachment to the system.

These are the standards we design Nova to, with the aim of strengthening the rest of an employer's provision rather than sitting apart from it. For more on the safeguards that need to sit alongside these principles, read about our approach to AI clinical safety.

There is still plenty to test, and the paper invites researchers to examine whether AI builds skills, where its relational limits lie and whether companion-focused design displaces human connection. Our position is that AI should build people’s capacity and strengthen human connection, and we want research to test how well different designs deliver on that.

You do not need to wait for those answers to ask better questions of providers. Take Section 5, Box 2 of the full preprint into your next vendor conversation and ask to see how these principles shape behaviour, measurement and handovers, not just policy. The aim is AI that complements human connection, rather than replacing it.