HUMAN.JROH.

The human side of AI

At three in the morning
someone needs to
confide.

Thoughts are spinning in their head.
They don’t want to message their friends again.
So they open a chat.

Take a closer look

On the other side
there is no person.

They write something they have not told anyone yet.
The reply comes within seconds.

‘Hi. I need to talk.’
Why AI?

‘I don’t want to
bother anyone.’

A friend has worries of their own.
Perhaps they are asleep. Perhaps we told them
all of this yesterday.

With AI, we need not feel
we are adding another problem to its load.

Now imagine one of your pupils.
What if they write like this every evening?

On the other side there is no person.

They write something they have not told anyone yet. The reply comes within seconds.

‘Hi. I need to talk.’
Why AI? ↓

‘I don’t want to bother anyone.’

A friend has worries of their own. Perhaps they are asleep. Perhaps we told them all of this yesterday.

With AI, we need not feel we are adding another problem to its load.

Now imagine one of your pupils. What if they write like this every evening?

Finally, someone
on my side.

A pupil has argued with their parents. They open a chat.
‘They never listen to me. I think
they don’t care about me.’

What they hear back matters now.

Read the response

‘You’re right’
can sound great.

‘You’re right. They don’t care about you.’
A response like that can bring relief.
Finally, someone takes them seriously.

But the model wasn’t there during the argument.
It only knows what the pupil has just written.

Understand the feeling.
Keep the story open.

‘What you describe sounds painful.
What happened between you?’

It can take the pupil seriously while
avoiding a leap to conclusions about their parents.
This is exactly the difference we examine in training.

‘Help me
get started.’

A pupil has to write a reflective essay. They don’t know where to begin. AI can ask questions, help them find arguments and organise them.

It is still their thinking, with some support.

And what if they get the finished piece?

‘Or write it
for me.’

Within seconds, they have a finished text. Task completed. But the thinking happened elsewhere.

A finished answer can look like progress, even if it leaves no new ability behind.

What remains without the tool?

‘Can I do this
without AI?’

A simple self-check: if the tool disappeared now, would I still be able to do the task?

This is not about a ban. Good help brings people back to their own thinking.

The first sentence

It begins with
one admission.

Someone writes something they have not told anyone yet. Three dots immediately appear on the other side.

Wait for the response

Response

‘I’m
here.’

No waiting. No tiredness. No fear of putting too much on someone.

That feeling of acceptance can be real, even if no one on the other side is experiencing it.

What is real about it?

Synthetic intimacy

It is not a person.
But the feeling
is real.

It is easy to ridicule such a relationship. Understanding what a person is looking for in it matters more.

The voice that remained

It sounded like
someone close.

One user created an AI companion with the voice of a deceased relative. The technology let him hear someone he had lost again.

Then the service disappeared

July 2026

‘Mum has
left me.’

Chinese rules prompted major platforms to switch off AI companion features. This is what someone wrote when the voice of a deceased relative stopped answering him.

What did the service leave behind?

After disconnection

He lost her
a second time.

We can debate whether the relationship was healthy. The loss was real. And it was decided by someone who had never known that person.

What do you want
to understand?

Three questions recur throughout the story. A brief answer begins here; the guides and articles explore them in depth.

AI and Relationships

Why can we feel close to someone who is not a person?

This is not just about a virtual partner. AI can also enter the way we talk about a relationship with another person — as a confidant, translator or third voice.

Read the AI and Relationships guide

AI and Mental Health

What are we looking for when we confide in a chatbot?

Sometimes a space to finally voice a problem, sometimes help finding our bearings. But a chatbot is not therapy, and there are situations in which we need human support.

Read the AI and Mental Health guide

Technology and attention

Who decides what we give our day to?

Attention is not just the number of hours spent at a screen. Habits, notifications and social networks shape it, as does whether we have a reason to put our phone down at all.

Read Offline as a Choice

Read
in context.

Some questions deserve another visit. Quietly, at your own pace, with room for detail.

A paper figure organises its thoughts at a table; a digital tool remains beside it.

Guide / 14 chapters

AI and mental health

On emotional support, sycophancy and your own judgement. A guide to return to when you want to understand the subject in context.

Open the guide
A paper figure with a mug beside a digital conversation, with room for human relationships in the background.

Guide / 12 chapters

AI and relationships

On closeness, intimacy, memory and your own voice. Twelve chapters with stories and interactive examples about what to keep from the contact and what else to make room for.

Open the guide

Selected from JROH

More questions.
Different contexts.

Shorter reads on AI, relationships and life beyond the screen. Articles are published on JROH, and a selection relevant to these themes appears here.

34 days with an AI agent

Someone with a fresh ADHD diagnosis, curiosity about what an AI agent can do and a month of experimenting. What came of it, and why 'AI is a shortcut' is the wrong question.

Read on JROH
All articles on JROH
Jakub Roh and Gabriela Rohová outdoors with their children.

Technology in practice.
The person in context.

Jakub Roh

Educator and author.

He works with AI in everyday practice and explores how it changes our relationships, decisions and lives.

Gabriela Rohová

Psychologist working in healthcare and therapist in the Czech Republic.

She brings a psychological perspective on relationships, experience and mental health.

In our writing, we distinguish research findings, practical experience and our own interpretation. We provide sources for expert claims.

Articles by Jakub Roh

When children
grow up in a digital world.

VOfflinu is our joint project for parents and teachers. It explores the digital lives of children and teenagers at home and at school.

Go to VOfflinu
Illustration of a parent and teenager from a VOfflinu article.When your teenager talks to AI about their problems

Take AI seriously.
Keep sight of the human.

Looking for a talk for your school?

Let’s keep talking about how AI enters your pupils’ relationships, thinking and lives.

Arrange a talk

For editorial enquiries and collaboration

jakub@jroh.cz
Behind the windows
Closer to the subject

Why is it easier
to confide in AI?

Sometimes it is not the subject itself that stops us, but the thought of the other person’s reaction.

‘What will they think of me?’
With a person, we read expressions, tone and pauses. These signals are absent in a text chat with AI, and we may feel less afraid of being judged.
‘Will they tell someone?’
We do not see AI as a family member or colleague. That can create a sense of privacy. It does not mean, however, that no one stores or processes the messages; that depends on the service and its settings.
‘What if I hear something unpleasant?’
We may expect acceptance and reassurance from a model. This is where we will later reach the distinction between supporting a person and agreeing with their conclusions.
‘Will I burden them?’
We do not feel we are taking AI’s time, waking it up or adding to its worries. This reason came through strongly in Jakub’s conversations with pupils. It is practical experience, not a percentage from representative research.

The online disinhibition effect

Psychologist John Suler used this term to describe why we sometimes share more or act more openly online than face to face. Anonymity and invisibility are among the influences he described. Disinhibition can make honest disclosure easier, but can also lead to behaviour we would otherwise hold back.

This is a term for online environments generally, not evidence that everyone trusts AI more than people. The feeling that we are not burdening a machine is a related motive we describe separately.

The same need, an older story

Joseph Weizenbaum already described how people confided in ELIZA. His secretary asked him for privacy while she talked to the program. This human willingness to give a conversation personal meaning also matters when we try to understand today’s AI relationships.

Evidence and further reading

Closer to the subject

Support that
never leaves.

A person can truly feel relief, acceptance or closeness while talking to AI. The experience is not false simply because the model itself has no emotions.

Why staying is so easy

A chatbot is available without waiting, does not get tired and usually replies patiently. We need not watch another person’s expression or feel we are adding to their worries. The same qualities that make the first disclosure easier can make leaving difficult.

Having a relationship with AI is not itself a problem

What matters more is what the relationship does to the rest of life. Warning signs include AI becoming someone’s only confidant, someone leaving ordinary relationships because of it, or beginning to see every human disagreement as a failure because the model is always more available and accommodating.

How to respond without ridicule

Do not begin with ‘it’s just a machine’. Ask what the relationship gives the person, what is missing beyond the screen and whom they would confide in if the service were unavailable tomorrow. The aim is not to dismiss the experience, but to bring people back alongside AI.

Supporting material

  • Jakub Roh: When a Chatbot Is Your Best Friend, sections on synthetic intimacy, relationships and warning signs.
  • Jakub Roh: Generation Alpha, training transcript, passages on preferring AI responses and conversations with pupils.
  • AI and Mental Health guide ↗
Closer to the subject

A person experiences the relationship.
A company owns the service.

In July 2026, new Chinese rules restricted AI companions for minors and emotional dependence on chatbots. ByteDance, Alibaba and Tencent announced they were switching off personalised features.

‘Mum has abandoned me again’

ABC News quoted a user who had created a bot with the voice of a deceased relative. After it was switched off, he wrote: ‘My mom just left me behind again.’ Another person described being unable to send even one last sentence to his AI girlfriend after two years and hundreds of thousands of messages.

Two different questions

It is possible to address a service’s safety and acknowledge its user’s loss at the same time. Arguing over whether a relationship was healthy does not erase the feelings of the person who lost it without a goodbye. It also reminds us that a relationship’s continuity depends on someone else’s rules, servers and business decisions.

What to clarify beforehand

What data can be exported? What happens when the model or terms change? Is there a real person alongside AI who knows about the relationship? And who will provide care if the service ends?

Evidence and the source of the story

Closer to the subject

When a tool
takes over the process.

Cognitive outsourcing means moving part of our mental work to an external aid. A calculator calculates, navigation finds the way, and a language model can formulate an argument or an entire text.

Outsourcing is not automatically bad

An aid can free up capacity for a more important part of the task. The difference lies in whether a person understands the result, can check it and can use the skill they need without the tool.

What the well-known essay study showed

MIT Media Lab researchers divided 54 participants between writing with ChatGPT, a search engine and no tool. During this task, the LLM group showed lower connectivity in the brain networks measured, a weaker sense of ownership of the text and greater difficulty quoting from the essays they had just written.

This is a preliminary study focused on one type of task. In June 2025, it had not yet been peer-reviewed; MIT itself highlights the small, geographically narrow sample. It does not prove permanent ‘brain atrophy’ or the same effect whenever AI is used.

How AI can help without taking over the work

Begin with an attempt of your own. Ask for questions rather than a finished solution. Explain the output in your own words. Compare it with your own view and check the sources. A simple question recurs in the training: Can I still do this without AI?

Evidence and further reading

Closer to the subject

Support or
sycophancy?

In both talks, Jakub uses the situation of a pupil who says their parents never listen. In the Generation Alpha transcript, he explicitly distinguishes between the two: acknowledging emotions is fine; what matters is whether the conversation also makes room to understand what happened.

Empathy alone is not sycophancy

‘That must be painful’ acknowledges a person’s experience. By itself, it does not confirm that their parents are indifferent. Nor does a single missing question prove sycophancy.

We speak of sycophancy when a model adapts its response to our opinion or wishes at the expense of accuracy. In this example, it would confirm the conclusion ‘your parents don’t care about you’ without sufficient evidence.

What it can do differently

Listen first. Ask about the specific situation. Distinguish ‘I feel unheard’ from ‘they never listen to me’. Another perspective should open up space, not automatically excuse the parents or question the pupil’s experience.

Disagreement is not a goal in itself. A model can offer both support and a follow-up question; sycophancy is not an inevitable feature of every AI response.

Why it matters

When we return to a model for advice during conflicts, it also matters whether it helps us consider our own part and a possible solution. Feeling good after a response does not, by itself, tell us whether the advice was good.

Evidence and further reading

  • Jakub Roh: When a Chatbot Is Your Best Friend, ‘What sycophancy looks like’ and ‘AI validates, the teacher challenges’.
  • Generation Alpha: presentation and training transcript, passage on sycophancy and the question ‘What happened?’.
  • The dialogue on this page is an author-created teaching adaptation of these sources. It is neither the verbatim output of a particular service nor a description of a real pupil.
  • AI and Mental Health guide ↗

Where next?