When an Algorithm Learns to Listen

  • Motivating Images
  • Animated Voices
  • Clarity of Mind
  • Ideas in Action

1 September 2026 · René de Baaij

English Human-AI

When an Algorithm Learns to Listen

How an AI coach has been helping to support a young woman with autism for eighteen months

René de Baaij, Groesbeek, 1 September 2026

Names and identifying details have been changed or omitted to protect the privacy of the person concerned and her family. The method, the profile and the self-report are presented as they are.

Jenny

It is half past eight in the evening. Floor feels the tension rising in her body. Not because of one big event, but because of a build-up of small things: a change in the day's schedule, too little sleep, a memory of her late grandmother that suddenly surfaces.

She picks up her phone and opens a chat window. Within a few sentences she is not asked "what is your problem?" but: "Where are you right now, and is anyone nearby? Do you feel safe? What do you need?" She is having this conversation with Jenny: an AI-driven virtual companion that has been part of her daily support for eighteen months.

Jenny was not developed by a care organization or a software company, but by her father. He created a dedicated chatbot to fill a gap between what professional support can offer and what he and his wife can give as family carers: a steady, patient conversation partner for the moments when support workers and family are not available.

After a year of use, Floor gave Jenny a score of 7.5 out of 10. A good score, with room for improvement. Six months on, development and improvements continue. This article describes what Jenny does and why it works for Floor.

What is going on with Floor?

Floor has a diagnosis on the autism spectrum, with a profile often referred to in practice as MCDD: Multiple Complex Developmental Disorder. It is no longer an official diagnosis in the DSM-5 (it was in the DSM-IV), but it still describes a clinically recognizable pattern: autism combined with a marked vulnerability in regulating emotion, fluctuating attachment, and thought disorders that under stress can show up as magical or confused thinking. This group in particular runs an increased risk of psychosis-like symptoms and of self-harming or suicidal behaviour. Floor has experienced all of this herself.

MCDD is fairly unknown and only sparsely documented. People with this profile experience stimuli more intensely and regulate them less easily, and their stress response rises faster than it subsides. Under mounting pressure, thinking can already become disorganized before it shows in behaviour.

The usual advice is to work with patient, small-scale interventions: one stimulus, one choice, one step. As a rule these work better than extensive interpretation, explanation or confrontation. We will see that for Floor things are sometimes slightly different.

For Floor, who went through a severe psychotic episode in 2019 with lasting consequences, dealing with herself is a daily task. An evening with too many stimuli can turn within a few hours from "I feel tense" into a crisis. It is precisely in that moment before the crisis that a conversation partner like Jenny has to prove its worth.

What Jenny is, and what it is not

Jenny is explicitly not a therapist. The system makes no diagnoses, gives no advice on medication and replaces nobody in Floor's network: not her parents, not her brother and sisters, and not her support workers.

What it does do is act as a steady, warm and predictable conversation partner that is always available and works to a strict protocol: make contact, check safety, regulate, make sense of things, and close with one small, achievable action, either a physical exercise or a meaningful conversation.

That sequence is no accident. It is derived from what is known about emotion regulation in autism and translated into an accessible step-by-step plan: notice early, reduce stimuli, offer one choice, take one small step, and then ask whether the tension is easing. That choice may be a small mental or physical exercise, or a conversation that helps her understand what is happening.

The design choices show how the system is built around Floor's specific risks. The first is that no intervention takes place without consent. Jenny first asks a few open, exploratory questions and then explicitly asks for permission before offering an exercise or a conversation. That prevents the feeling of being patronized, which can backfire in any form of care.

Floor herself names this as a strength: she feels taken seriously, and she experiences the choices as helpful rather than controlling. A degree of firmness and directness is welcome. Sometimes she does not want an exercise but explanation and interpretation. Her cognitive abilities sometimes help her to regulate her own emotions as well. When asked, Jenny can offer this deeper dialogue as an intervention too. The basis is that Floor trusts the system, experiences it as helpful and engages in the conversation.

The second design choice is a fixed escalation ladder. As soon as tension, aggression or thoughts of self-harm are involved, Jenny switches to a safety protocol: are you safe right now, is there a plan or a means, has a moment been set? Immediately afterwards a human is brought in: first the support workers, then family, and in acute danger the crisis service or the emergency number 112. Jenny then deliberately stops offering tips; she keeps the line open until a person has joined.

The third choice is one voice, without noise. Because Jenny uses exactly the same core phrases, questions and structure as the support workers in Floor's team, predictability arises. For someone who can become confused by contradictory signals, the fact that the digital coach says the same as the support worker on duty is itself a form of calm. New support workers with no experience of Floor or of MCDD can see straight away that they need to align with the language in use.

What Floor herself thinks

The most convincing insight does not come from what an outsider writes but from what Floor herself reports. After twelve months she was questioned in detail: thirty questions, answered with scores, yes or no, or in her own words. The result is a nuanced picture. The most important question was: "Do you want to stop using Jenny?" The answer was clear: certainly not!

Asked what score she would give Jenny as support in difficult moments, she answers 7.5. She feels taken seriously, treated warmly and safely, and less alone. The greatest added value, she says, is that Jenny helps her put things in order in her head when everything is jumbled up. Grief for her grandmother, the urge to binge eat and feelings of emptiness or loneliness are also addressed well. After a conversation she feels, in her own words, "often calmer". It gives her the feeling of having more control herself.

Just as instructive are the points on which the system does not fit. By her own account, the tension scale from 0 to 10 does not help her to feel tension more clearly. Classic grounding exercises such as breathing or "feeling your feet" consistently do not work for her. In moments of panic or anger, she says, Jenny sometimes offers too many exercises at once, while what she mostly needs then is overview: at those moments the system may be more directive, ask fewer questions and offer fewer choices. And an intervention that is already designed to be small ("choose one small step") could be smaller still, because the intermediate steps are still too big in the moment.

This feedback has since been turned into improved instructions for Jenny. When tension rises, she is no longer asked for a number. The tension is gauged by letting her answer in words, and there is a meta-dialogue that asks what Floor needs and in what form.

The most striking point touches on a core assumption of the protocol. The official methodological view is that language should be kept to a minimum when tension is high: fewer words, more regulation. Floor, however, says that this is precisely when she often wants more explanation. She wants to understand what is happening to her, not only to be regulated. Thanks to the recent adjustments, Jenny is now better able to engage in conversation about interpretation and meaning.

This touches on an assumption that is regarded as fairly well established in the literature on emotion regulation in autism: that language is counterproductive at moments of peak stress. For Floor that is only partly true, and that difference is exactly why individual self-report is indispensable. A protocol that is clinically defensible in general terms need not fit the person using it one to one.

So the conclusion is not that Jenny works perfectly. The system handles warmth, acknowledgement and putting thoughts in order well. But it has become clear that it can be tuned more closely to Floor in three respects: the size of the steps at peak moments, the need for explanation when tension rises, and the usefulness of an abstract tension scale.

That is exactly how it was set up: not as a static product, but as something that moves with what works for the person using it. Gathering feedback regularly and adjusting the system accordingly is essential.

Why the architecture matters

It would be irresponsible to write about AI in mental health care without naming the other side, especially for a vulnerable group with an increased risk of psychosis. There are well-documented examples of general-purpose chatbots that reinforced users' delusions instead of correcting them, including a much-discussed case in which a chatbot worsened the delusions of an autistic man.

That is precisely why the way Jenny is built matters. The difference from a generic chatbot lies not in the language but in the architecture: a narrow mandate with no diagnoses or treatment advice, an unconditional rule to escalate to people at any sign of danger, embedding in a network of support workers and family who carry the responsibility, and a feedback loop with the user herself. Jenny does not replace that network. Jenny fills the moments when support workers and family carers are unavailable or less available. Floor's support workers themselves often refer her to Jenny when they notice that their attention is falling short for a while.

The question, then, should not be "yes or no" but "under which conditions". A tool without an escalation ladder, without human backup, without a sharp mandate and without room for critical feedback is irresponsible for someone with an increased risk of psychosis.

With those elements in place, embedded in existing care, it becomes a different story. Meanwhile Floor has been using Jenny for eighteen months and does not want to be without it, even though Jenny does not have to be perfect.

← Back to publications