Reassurance · step 1
‘Wasn’t my message embarrassing?’
‘Definitely not.’
No new evidence.Room for more life
Two pre-written example sequences
Three corresponding steps are always side by side: seeking more certainty on the left, completing a task on the right.
Reassurance · step 1
‘Wasn’t my message embarrassing?’
‘Definitely not.’
No new evidence.Task · step 1
‘Is the address on the envelope complete?’
‘The postcode is missing.’
Specific information has been added.Reassurance · step 2
‘Really?’
‘Definitely.’
More certainty, the same evidence.Task · step 2
‘I’ve added the postcode.’
‘All parts of the address are filled in now.’
The evidence has changed.Reassurance · step 3
‘Can you check it once more?’
No new evidence has been added.
There is nothing for another response to build the conclusion on.Task · step 3
‘So is it ready to send?’
‘The address is ready.’
The task has a reachable end.‘The message is already sent. Now I’ll get my things ready for tomorrow.’ The first sequence ends; completing the address can continue.
A possible end to the first thread: ‘The message is already sent. Now I’ll get my things ready for tomorrow.’ The final reassurance need not arrive.
The difference is not the number of minutes or questions. In the second thread, evidence grows and the task is completed. But danger is not a ‘loop’ simply because there is no new information: seeking protection can and should continue.
Ondřej moved to another city. He enjoys talking to AI in the evenings, but during the week hardly meets anyone outside work. He wants to keep the contact. He misses people to share experiences with.
He chooses a regular gaming group. He checks the time, price, transport and whether newcomers are welcome. He writes to the organiser saying he knows no one and would like to come. For now, he wants to try one shared activity; he need not find a best friend on the first evening. After the gathering, he may open the chat again and say what it was like.
He began with what he wants to add to his life. Similarly, you can choose a workshop, walk, volunteering, sport or a regular call with someone you know. Such an intention has clearer direction than ‘I must use AI less’. Follow your own interests, availability and safety.
A regular video call with a friend is a human relationship even if you are not in the same room. A walk with a voice assistant is movement outdoors, but does not itself meet another person. Life has room for both, and for moments without any ongoing commentary.
With a long-term illness, limited mobility or a demanding family situation, remote contact may have special value. Even an aid you need long-term can let you do things according to your own intentions. The question is whether it continues helping or whether use itself starts asserting itself against what you want. A number of friends or hours does not answer this alone.
In four-week research with 981 participants, researchers followed, among other things, different assigned conversation types and communication modes. In the cited version, they found no statistically significant main effects of these conditions on the psychosocial outcomes measured. People who voluntarily used the chatbot more nevertheless showed worse outcomes. But use volume was not randomly assigned, so this association does not tell us what caused what.[1] For a particular person, it matters to watch how use develops and what it gives or takes away.
‘I’m addicted to it’ can mean several different things. Some describe themselves this way jokingly because they enjoy the conversations. Others need the assistant to manage everyday tasks. Another person wants to stop but repeatedly continues, losing sleep, money or relationships they want to keep. This last situation is precisely what needs attention.
When worried about dependence, therefore ask about control over use and its impact. Do you remain in contact because it suits you, or regularly against your own decision to stop? What gets pushed aside? Have you managed to change something you already want to change? These questions help describe a difficulty; they are not a diagnostic test for ‘AI addiction’.
Attachment and lost control can meet, but are not the same. You can miss someone, look forward to daily conversation or keep their favourite sentences while still deciding how to live. Conversely, exhausting use can develop without romance or a name for the assistant. What happens matters, not whether the contact looks unusual from outside.
When you want to change something, choose a specific point where you lose choice. If notifications keep you going, adjust them on the device. If purchases burden you, address actual payment settings. If you return for more certainty, look at the loop in the next section. Changing one of these patterns need not mean rejecting the entire relationship.
If you cannot change use despite its effects, ask for help. You can say: ‘I don’t want to lose what this contact gives me. But I need to regain sleep and the ability to stop when I decide.’ With serious deterioration or danger, help is not postponed until after your own attempt at change.
Markéta received a message from her partner without the usual greeting. She asks whether it means his feelings have cooled. The chatbot offers several possibilities. Markéta feels relief, then notices the full stop at the end and asks again.
We call this repetition a loop, or looping, here. Each answer brings certainty for a moment, followed by another doubt and need for reassurance. No new information about the partner arrives. The conversation may be kind and free of pressure from the system while still exhausting Markéta. Repeated reassurance can also occur in obsessive-compulsive disorder, known as OCD. One repeated question, however, does not establish this condition.[2]
The difference from follow-up questions when learning is what another answer brings. When you finally understand an explanation, you have learnt something new. When you need to hear again that your partner definitely is not angry, you keep working with the same uncertainty. Sometimes the message’s text cannot remove it, however many times you ask.
With a recognised loop, you can try not requesting another analysis and move to one specific activity you postponed. You need not decide what your partner ‘really’ thinks. If approaching them is safe, you can ask them later. On card B, you prepare what to do when the need for another answer returns.
Interrupting a loop does not mean stopping the search for help in danger. Violence, health difficulties and suicidal thoughts are not addressed by no longer treating them as a subject. Use appropriate support there. If you already have a treatment plan, follow your agreement with the professional.
A shared dinner without a phone, time for music, a regular game, your own opinion written without the model’s prior approval. Choose one thing and a specific first step. AI can make finding options easier or shorten a message; how much you involve it in further contact remains yours to decide.
The search may reveal another obstacle: the group is inaccessible, childcare is missing or strong anxiety makes contact difficult. Adjust the first step, invite support or look for another option. Another list of impossible obligations would not help. On card H, you can consider what is feasible under your conditions.
A boundary in an ordinary human conversation can be equally specific: ‘I don’t have room for a long chat today. I’d be happy to call on Sunday.’ The skill is not only the wording, but also accepting that the other person can suggest another time. Not every response needs to go straight into a chat for interpretation.
AI may remain the only confidant where other support is missing. It can also become one through gradual withdrawal from previously available relationships. In both cases, a person says ‘I only talk to AI now’, but the path to that situation differs. ‘Only’ does not itself establish dependence; we need to know what changed in life.
Seek further support accordingly. An acquaintance or community for ordinary contact, a colleague or counselling service for a practical problem, an appropriate professional for health difficulties. Different people and services can fulfil different roles. None has to be constantly available.
If missing support is accompanied by the lost control over use described above, ask for help with that too. In danger, seek it directly, without waiting for the outcome of your own attempt to limit the application.
For an ordinary, low-stakes matter, try deciding without further analysis: what to take on a walk, when to message an acquaintance, which wording of your own to use. Rely on available information. If it does not work out, you can learn from it; that does not itself prove the chatbot should have decided.
This does not concern arbitrary treatment changes, ignoring danger or removing a needed assistive aid. The aim is not absolute self-sufficiency. Support should help your action, not become compulsory approval for every small thing.
You can have a good reason for change and your own timing. And if your current form of contact suits you, this chapter is not a task to find a problem in it. Return when you want something different.
Fang, C. et al. How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study. arXiv:2503.17473, cited version 2. Primary paper. Abstract reviewed 12 September 2026. Four weeks, 981 participants. Experimental condition and voluntary use volume are not identical variables; an association with intensity does not establish its causal effect. ↩︎
OCD Action. Reassurance. On repeated reassurance. International OCD Foundation. About OCD. Taken from the other guide’s evidence and the approved example. Distinguishes support and repeated reassurance in the context of difficulties; a loop in this book describes a pattern, not an OCD diagnosis. ↩︎