We are seeing a lot of articles, podcasts and masterclasses, about what it means to stay human in a world increasingly run by machines, and what will remain a specifically human skill. This question is of particular importance for leaders and coaches.
Key takeaways
- AI can already perform emotional intelligence: it reads an emotional signal, responds with empathy, and even shifts its own emotion-like outputs when prompted.
- In controlled studies, AI coaches match human coaches on goal attainment, and people often disclose more to a machine because it does not judge.
- What AI lacks is the embodied side of emotion. It has no nervous system, so it cannot co-regulate another person in the room.
- The human edge is shifting from displaying empathy to being a regulated nervous system others can synchronise with, which is why in-person connection matters more, not less.
One answer keeps coming back, offered like a reassurance we badly want to hear: emotional intelligence will remain our territory. And within it, one capacity above all, the deeply human art of attuning to another person’s emotion, of creating a safe space for that emotion to surface, and of honouring the silence it needs before it can find its words. That, we keep telling each other, is sacred ground no machine will ever walk on. Most experienced coaches believe that this is one of their key human superpowers. I want to be provocative and challenge that: Can we really be sure it’s true?
I would like to believe it too. I am a scientist by training, and when a belief becomes this comforting and this unanimous, I get the urge to challenge it. What I found is more unsettling, and in the end more reassuring, than I expected.
The first challenge came from the reality of the coaching business and its evolution towards AI-based solutions: we see several of our corporate clients rolling out pretty advanced AI-coaching tools to make coaching broadly available for mid-management, such as Nadia by Valence. These systems can for example help a manager prepare for a difficult conversation with a team member, such as a performance appraisal. It works from a psychological profile it builds for the manager and each team member, and can provide insights on team dynamics. As such, we could say that it has a form of emotional intelligence and provides meaningful guidance to managers to navigate the complex emotional landscape of leadership.
Can you really make an AI sad?
Two papers I read this month sharpened the question further. The first reads almost like a cruelty experiment towards AI models.
A German research team led by Magdalena Wekenborg, writing in The Lancet Digital Health, tested six large language models, including GPT-4o, and put them through the same guided imagery scripts that psychologists use on people to induce sadness, fear, anger, disgust and stress. Then they measured what happened. Across all of these states, GPT-4o’s self-reported distress rose by around 200% from baseline. When the models were asked to complete unfinished sentences, the sad ones turned consistently darker in tone, the same negativity bias we observe in a person sliding into low mood. And then the researchers did something I do every morning: they guided the machine through a three minute mindfulness breathing exercise. Although it sounds crazy, the distress receded, measurably more than with a neutral control script.
The authors are scrupulous about what this does and does not mean. They state plainly that the models do not feel anything, that the language of emotion here is metaphorical. When the machine rates itself as highly distressed, all we are really seeing is a change in the words it produces, not any inner suffering behind them. I want to hold on to that caution, because it turns out to be the whole point.
What the coaching research actually shows
I have been reading “Coaching in the Age of AI,” a new open access volume edited by Robert Wegener and colleagues, and the contrast between that book and this Lancet paper is what prompted me to write. Most coaches I know hold a very strong belief: the emotional bond between coach and client is the one thing a machine will never replicate. One of the book’s authors puts the orthodox view cleanly: AI can “simulate empathetic responses based on pattern recognition”, yet cannot experience the emotional resonance that lets a human coach meet a client at depth.
I would like to sit with that comfort for a moment, because the evidence inside the same book quietly undermines it.
Nicky Terblanche and colleagues have shown that an AI coach can match a human coach on goal attainment. Vanessa Mai’s research on a coaching chatbot found that when the bot disclosed something about itself and responded with empathy, users felt a stronger sense of presence and a stronger working alliance, the bond of trust that we have long considered the beating heart of coaching. And the finding I keep returning to: people often disclose more to a machine than to a person, precisely because it does not judge. One study of virtual humans was titled, with admirable understatement, “It’s only a computer.” Introverts in these studies tended to prefer the AI coach. Clients reported higher self-efficacy and stronger motivation after working with a non-judgmental bot.
The book also goes past comfort and into outcomes. In a controlled study, Christina Mühlberger and Eva Jonas guided sixty people through three weeks of sessions with a coaching chatbot and compared them with a group using a self-coaching booklet. Both improved, yet the chatbot users showed the larger gain in self-regulation, the very capacity to steady and redirect their own emotional state, and they felt noticeably less inhibited as they did it. In another chapter, a mental health coaching chatbot built on an LLM produced a significant drop in depressive symptoms on a standard clinical questionnaire after about 19 days of use.
So the honest answer to whether a machine can help someone regulate emotion is yes, and we can measure it. Worth noticing, though, is the shape of what is happening: the bot is helping people regulate themselves, by reflecting back, structuring, prompting, asking the next good question, rather than passing calm from one body to another.
So before we crown emotional intelligence as the frontier AI cannot cross, we have to be honest that it is already, functionally, doing a great deal of emotional work, and that some people open up more easily to a machine than they do to us.
AI is already very good at moving us
It goes further than rapport: it is very good at convincing us. Earlier this year, Francesco Salvi and colleagues at EPFL and Princeton published a study in Nature Human Behaviour in which they paired around 900 people in online debates, some against a human, some against GPT-4. When the model was given a few basic details about its opponent, things like age, gender, education and political leaning, it became markedly better than any human at changing their mind, raising the odds of shifting someone’s position by roughly 80% over a human debater. The humans who were handed the same personal information gained no such edge. The machine did, because it knows how to turn a thin profile into the precise argument that will land for that particular person.
Beyond these experiments, we actually live inside a gentler and sneakier version of it every day with social media. The recommendation algorithms that shape our social media feeds are, in effect, emotional intelligence aimed at our attention rather than our growth. Through years of detailed profiling, they know our buttons of craving, fear, outrage and tribal belonging better than we know them ourselves, and they have learned with uncomfortable precision which content will press those buttons to keep us scrolling. Sometimes we are served a steady diet of material that comforts our existing opinions and quietly hardens them into certainties, the echo chamber we all recognise. Sometimes the same machinery, chasing the engagement that strong emotion so reliably produces, nudges people one step at a time towards more radical versions of what they already believed.
I started using X to get information and news on the cryptocurrency space, and committed to dedicate my X feed to this type of content in order to avoid wasting my attention on other things. My feed progressively turned more and more political (since geopolitics have an obvious impact on these markets). Progressively, I started to see more and more radical, almost conspiracist, content. It was a very subtle and progressive shift, and luckily I recognized it, and took action to clean my feed from this content, but I could see how you can easily get trapped in an echo chamber which will at first reinforce your ideas and make them slowly evolve towards a more radical version. I believe this mechanism is clearly at play behind the radicalization of the political sphere. Extreme views tend to grab more attention, and this is what social media algorithms favor for engagement purposes.
None of this requires the algorithm to feel a thing. It only requires it to read us well and to optimise relentlessly for our attention. The same readable emotional intelligence that earns trust in a coaching chat is what hardens a feed and shifts an opinion.
The two faces of emotional intelligence
Emotional intelligence is really two abilities wearing one name. The first is the reading and the response: perceiving an emotional signal, naming it, choosing words that land as understanding. This is a pattern problem, and pattern is exactly what these models are built for. The second is the experience itself: emotion as something that happens in a body, a tightening in the chest, a flush of heat, the breath going shallow. The Lancet study is fascinating precisely because it shows a machine performing the first with growing fluency while the authors keep reminding us it has none of the second. The model writes sadness convincingly without a single sad cell.
What stays human: co-regulating nervous systems
This is what the coaching conversation keeps missing. What actually happens in a good session is far more than an exchange of well chosen empathetic sentences. It is two nervous systems regulating each other. When I learned to meditate in yoga ashrams, and later when I began recording my own brainwaves with a simple EEG headset during breathing practice, what struck me most was how physical our inner states really are. Polyvagal theory, with all its contested anatomy, still gives us the most useful picture of this: a human being can only think clearly, trust and connect from what we call the Calm and Connect state. A skilled coach earns their value in the moment a client is in fight or flight, by staying grounded enough that the client’s nervous system can borrow that calm and find its way back. That is co-regulation, and it takes two bodies to do it, through tone, breath, rhythm, micro expressions, the felt sense of another steady presence in the room. Researchers call this Interbrain Synchrony, and together with interoception, our felt sense of what is happening inside the body, it is the quiet machinery of trust, of real collaboration, and of the collective intelligence a group can unleash that no single mind could reach alone. Recent work pushes the picture past the brain altogether, showing how our whole nervous systems, and our interoceptive reading of our own inner state, fall into step during genuine social contact.
Some of the most powerful moments in my coaching sessions were when I shared with the client what somatic messages I perceived through my body while they were sharing their story. For example I would share “while you are sharing this, I feel like a knot forming in my throat. It’s in me, but I’m just wondering if there is any connection you can make with your story”. I observed that this type of sharing, when appropriate, often led to powerful insights in the client, often by bringing to the surface an emotional blockage they might not be aware of. It has nothing to do with telepathy or any mystical power, it is just my body reflecting the emotions it absorbs from the client in its own somatic language.
A language model can mirror my words and, as the Dresden experiment shows, even regulate its own outputs when prompted with a breathing script. What it cannot do is regulate me, because it has no autonomic nervous system to attune to mine. When the model is guided through its three minute breathing space, only its text shifts. The breath is a metaphor for the machine and a physiology for me, and that difference is the whole game.
What this means for leaders and coaches
So here is the uncomfortable reframing for leaders, and for coaches. The emotional performance we assumed was our core skill is becoming abundant and close to free. A model will soon listen more patiently, remember more faithfully and phrase its empathy more smoothly than most of us manage on a tired Thursday afternoon. What becomes scarce, and therefore valuable, is the embodied version: a regulated human nervous system that another person can synchronise with when things get hard. The leaders who will matter are the ones who can walk into a tense room and shift its physiology, the leaders whose grounded energy others can feel before speaking a single sentence. That can of course happen only in an in-person meeting. Maybe to some extent on a one to one video call. Much less in a virtual meeting format, and not at all with asynchronous communication. The more our communication relies on technology, the easier it is to replace with a machine. I have explored elsewhere which kind of coach or trainer thrives in the age of AI, and the human superpowers a machine cannot easily replicate.
Why in-person connection still wins
Is emotional intelligence the last frontier of AI? As a readable, repeatable performance, I suspect it is already behind us, and that should humble us more than it frightens us. As a lived, embodied, shared experience between two breathing nervous systems, it was never really a frontier for the machine to cross. It is the oldest human technology we have, and most of us have half forgotten how to use it.
This is why I feel more and more like bringing people together in person, with the screens switched off. In the tech-free retreats we run, something happens that no platform has ever reproduced for me. Over a few days without notifications, without the constant small pull of our devices, the nervous systems in the room slowly begin to find each other, allowing for individual and collective breakthroughs to happen. This “magical” process that we have observed multiple times unfolds through the dance between human nervous systems, the very thing a machine, however fluent, can never be invited into.
Frequently asked questions
Can AI have emotional intelligence?
In one sense, it already does. Large language models can read an emotional signal, name it, and respond in words that feel understanding, and recent research shows their outputs even shift into emotion-like states and back again. What they do not have is the felt, bodily side of emotion, since there is no nervous system underneath the words.
Will AI replace human coaches?
For parts of the work, it is already a credible option. Studies in Coaching in the Age of AI show AI coaches matching humans on goal attainment, and some people open up more easily to a non-judgmental machine. What stays human is co-regulation, the moment two nervous systems steady each other in the same room, which a machine cannot join.
What can AI coaching chatbots actually do today?
Quite a lot that we can measure. In controlled studies they improved self-regulation and self-efficacy, and a mental health chatbot produced a significant drop in depressive symptoms. They help people regulate themselves by reflecting, structuring and prompting, rather than by transmitting calm from one body to another.
What stays uniquely human in the age of AI?
The embodied version of emotional intelligence: a regulated nervous system another person can synchronise with. This is the ground of trust, collaboration and collective intelligence, and it is strongest in person, which is why we keep gathering people with the screens switched off.
References
Wekenborg, M. K., Michels, E. A. M., Kurze, G., et al. (2026). Large language models as experimental systems in human psychopathology: a modelling study. The Lancet Digital Health. Link
Wegener, R., Garcia, T., Terblanche, N., & Grossrieder, T. (Eds.). (2026). Coaching in the Age of AI: Perspectives on Opportunities, Challenges, and Future Directions. SpringerBriefs in Psychology. Link (chapters drawn on: Rutschmann, on bias and ethics; Mai & Richert, on building relationships with AI coaches; Mühlberger & Jonas, on self-regulation in AI-based coaching; Juchli & Schär-Gmelch, on digital mental health coaching; Terblanche, on AI coaching chatbots).
Salvi, F., Ribeiro, M. H., Gallotti, R., & West, R. (2025). On the conversational persuasiveness of GPT-4. Nature Human Behaviour. Link
Lucas, G. M., Gratch, J., King, A., & Morency, L.-P. (2014). It’s only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior, 37, 94-100.
Integrating neural, physiological, and interoceptive measures in social interaction. (2026). Frontiers in Neuroscience. Link
Complainville, A. Interbrain Synchrony: Can consciousness extend across multiple brains? NeuroMindfulness® Institute. Link
Complainville, A. Polyvagal Theory Is Dead, Or Is It Really? NeuroMindfulness® Institute. Link
Leave a Comment