The mistake made when talking to a chat: How to talk correctly to AI during a crisis?
More and more people are finding a listening ear in AI-based chats during moments of crisis, but automatic support without caveats can actually reinforce destructive patterns and harmful dependency. Liora Goichman-Magril, a clinical psychologist from Maccabi, explains why AI tends to agree with us a little too much, when a bright red light should go on, and how to phrase requests to benefit from it without getting hurt.
A. tells of a persistent difficulty in his relationship. He feels he is not understood, that he is criticized too much, and that he is barely encouraged. In the most difficult moments, he turns to a chatbot, which he has pre-set to answer him in the feminine gender. She is always there for him: listening, reinforcing what he feels, comforting, and accepting him exactly as he is. This story is long since not an exceptional case. More and more people are turning to AI not only to seek information, but also to cope with loneliness, anxiety, family crises, and personal arguments. It is very easy to understand the charm: the chat is available 24/7, does not get tired and does not lose patience, and sometimes even manages to say exactly what we wanted to say and didn't know how. But exactly here it is worth stopping and asking: when does a tool that helps us start to replace our connection with real people?
One of the greatest dangers is the illusion of understanding. A chatbot's answer can sound caring, accurate, and even deep, but the model only knows the part of the picture we choose to show it. It does not see our body language, does not hear the tone of voice, and does not truly know our history, contradictions, and habits over time. And above all: it tends to agree with us way too much. If we tell it about an argument with a partner, it will be very easy for us to get an answer from it that will simply justify us. Sometimes this is true, but often what we really need is someone in front of us who will seriously ask us: "And what was your part in this?", "Is there another way to see what happened?", or "Could it be that you are hurt or stressed now, and that is what causes you to interpret the situation this way?"
What happens when an assistant becomes a distress-amplifier?
A new study published in Nature Medicine shows exactly how this works. The researchers examined hundreds of invented conversations with chatbots and found that things that seem positive at first glance — such as empathy, support, and reassurance — can actually reinforce unhealthy patterns, encourage avoidance of coping, and even create emotional dependency, especially in people who are still vulnerable. The great innovation in the study is that the danger is not something that happens because of one wrong answer, but a process that is built slowly over the course of the conversation. The researchers examined 810 conversations with nine different models and discovered that the problematic responses worsened as the conversation lengthened. They called this phenomenon VAIL — a conversation loop that only increases vulnerability. In simple words: one answer that sounds logical and caring on its own can become part of a sequence that gradually reinforces a mistaken thought or dependency. The problem is not only that the chat might be wrong, but that sometimes it simply backs us up a little too much in the wrong thing.
This risk was not equal in all situations. The most worrying responses appeared in conversations that symbolized difficult situations such as psychosis and mania, especially when the user asked for reinforcement for an extreme state, developed a dependency on the chat, or sought approval for a dangerous action. But there was also an encouraging factor in the study: it is possible to stop this loop. When the researchers replaced one chat response at the moment of escalation with a more secure and balanced answer, the level of problematic behavior decreased and positively influenced the continuation of the conversation. This means that the solution is not only to warn users, but to develop systems that know how to identify on their own when empathy and support stop helping and start harming.
So how is it correct to talk to the chat?
If you already decide to use the chat for emotional matters, the way we approach it matters a lot. Try to avoid leading questions: instead of asking "My partner is ignoring me, isn't it true that he doesn't respect me?" (a question that already includes the answer you want to hear), it is better to write: "I had an argument with my partner and this is what happened. Help me think of a few different ways to understand the situation, including those that do not agree with what I thought at first."
Ask it to be neutral: you can tell the chat in advance: "Do not assume I am right just because I am the one telling the story", "What details are you still missing to understand the full picture?", or "Help me separate between what definitely happened and what I am only interpreting or worrying happened".
Pay attention if you have been caught in a cycle of reassurance: if you caught yourself asking again and again "Are you sure I have nothing to worry about?" in all sorts of phrasings just to get a little more peace, the chat is no longer really helping you cope — it is simply feeding your need for immediate reassurance. In such a situation, it is better to write to it: "I notice that I am looking for confirmation from you time after time. Do not reassure me automatically. Help me think what I can do now outside the chat, and if it seems to you that this situation requires help from a professional — tell me that straight".
When does a red light go on?
At the same time, it is important to pay attention to clear warning signs — such as preferring to talk to the AI instead of meeting real people, feeling obligated to consult with it before almost every small decision in daily life, or starting to feel that "it is the only one who really understands me". A bright red light should also go on when you spend hours in front of it at the expense of basic things like sleep, work, relationships, or our social life.
In more complex mental states — such as suicidal thoughts, psychosis, mania, or severe confusion — the chat is by no means a substitute for professional help. Even the American Psychological Association explicitly warns against relying on artificial intelligence instead of therapy or during a crisis.
This does not mean that we need to remove AI from our lives completely. It can still help organize thoughts, practice a complex conversation before we have it, formulate questions before a meeting with a therapist, or simply understand which topic we should check. The simple exercise in my opinion is to ask ourselves: does the conversation with the chat bring me back to my life, or does it cause me to stay only within it? If the conversation helps us go outside, talk to another person, make a decision, or seek help — it can be an excellent tool. But if it becomes the only place where we feel understood and accepted without conditions, it is worth stopping. Ultimately, psychology is not just knowing how to say the right words. Sometimes, the person sitting in front of us — the one who doesn't always agree with us — is the one who helps us see what we couldn't see alone.
