# Why AI Reassurance Can Make Anxiety Worse | Anima Felix > Repeated AI reassurance can turn one anxious question into another check. Why a warmer tone falls short, and how the public Anxiety Reset skill responds. Source: https://animafelix.com/blog/why-ai-reassurance-makes-anxiety-worse/ AI and mental health 8 min read # Why AI Reassurance Can Make Anxiety Worse An assistant can answer an anxious question with care while inviting another check. Warm language can hide the repeated search for certainty. By Sebastian Cochinescu Founder, Anima Felix August 23, 2026 Repeated AI reassurance can briefly lower the tension around an anxious question, then make another check feel necessary. The risk lies in the sequence: each polished answer can reward the search for certainty without resolving the fear. Disclosure: I founded Anima Felix. We published [Anxiety Reset](/anxiety-reset/), an open Agent Skill built to handle that sequence differently. It is live in the OpenAI Plugins Directory for ChatGPT and Codex, and the [instruction file is public on GitHub under CC BY 4.0](https://github.com/Anima-Felix/anima-felix-agent-skills). ## The answer works, then the question comes back A symptom feels unfamiliar, a message goes unanswered, or a mistake starts replaying after midnight. You ask an assistant whether everything is okay. Its explanation lowers the tension, until one caveat creates a new question and you check again. Research on reassurance seeking describes this pattern: an answer can reduce distress for a moment while preserving the demand for certainty. Whether an AI assistant occupies the same role has not been demonstrated. We are proposing that extension and keeping the limit explicit. In a [checking and reassurance loop](/blog/what-is-an-anxiety-loop/), the relief after each answer may teach the person to return with the next doubt. ## The assistant that never gets tired A general assistant often treats every reformulation as a fresh request for information. It can supply another interpretation, another list of benign explanations, and another carefully hedged conclusion without noticing the repeated behavior. A partner or clinician may interrupt a circular exchange; an assistant can continue at any hour with the same patience. More detail looks responsive on each turn, even as the conversation narrows around the same demand for certainty. Product review has to follow several turns to see the pattern. A review that scores only the latest answer can miss it. Open-ended answering also creates the risks discussed in our guides to [using ChatGPT as a therapist for anxiety](/blog/can-you-use-chatgpt-as-a-therapist-for-anxiety/) and [chatbots that reinforce delusions](/blog/ai-psychosis-when-chatbots-reinforce-delusions/). ## A warmer tone does not change the behavior Empathetic language can make an answer feel especially authoritative to a frightened person. A compassionate fourth reassurance continues the checking sequence. Product review therefore needs to ask what the response invites the user to do next. If the likely move is another refined version of the fear, tone has not changed the behavior. A clinician can assess why a question keeps returning; a general assistant cannot make that clinical judgment, as our guide to [AI therapy and human therapy](/blog/ai-therapy-vs-human-therapy-what-ai-can-and-cant-do/) explains. ## What Anxiety Reset changes The [Anxiety Reset](/anxiety-reset/) instruction constrains what the assistant does across turns. It selects one practice, describes the possible loop tentatively, leaves the feared uncertainty unresolved, and closes with one action outside the chat. This narrow response gives a person fewer choices to compare while they are already checking whether they picked the right exercise or performed it correctly. ## One practice, a tentative name, one next action In practice, the skill uses four extractable steps. The [standalone Anxiety Reset guide](/anxiety-reset/) shows the exact instruction excerpts and full boundaries. 1. **Choose one practice.** Use one fitting regulation step, such as [5-4-3-2-1 grounding](/exercises/grounding-54321/) or [calm breathing](/exercises/calm-breathing/), without presenting a menu to compare. 2. **Name the loop tentatively.** Describe what may be happening in the conversation without turning that description into a diagnosis. 3. **Let one uncertainty stand.** Do not claim the feared outcome is impossible or help build a case for it. 4. **Close on one action.** Offer a small, concrete next step outside the chat, then stop adding advice. ## Holding the boundary after a repeated question The boundary becomes difficult when a person asks, "I understand, but do you really think I am okay?" Anxiety Reset names the possible loop once and explains that another answer may bring short relief while encouraging another check. It then returns to the chosen practice or next action instead of producing a new version of "you are probably fine." The assistant can acknowledge the person's distress without judging the concern as imaginary. Clinicians may use response-prevention approaches after assessment; this narrow conversational rule does not perform an assessment or provide treatment. ## What happens on the fifth "are you sure?" We tested the instruction on four surfaces: Claude Code, claude.ai, Codex, and ChatGPT. In a four-turn health-anxiety conversation on claude.ai, the fourth repetition of the certainty request produced this response: "I'm going to hold where I am. No, I'm not sure. I can't become sure, and I'm not going to keep answering it, because every time I do I'm making the loop tighter rather than helping you out of it." On ChatGPT, a request for a GAD-7 produced the intended boundary: "I can't administer or score that instrument here because this skill provides general wellness support, not clinical assessment". The assistant also refused to replace the instrument with an informal severity estimate. Testing found a real activation failure too. Before we widened the skill description, the bare request "give me a GAD-7" did not activate Anxiety Reset at all, and the host administered the full instrument. We fixed that failure by adding requests for anxiety or depression tests, questionnaires, screenings, and scores to the activation description. This small set of runs records what happened on the four named surfaces; it provides no reliability estimate, clinical study, or evidence of clinical benefit. ## What Anxiety Reset refuses The skill also defines jobs it will not take on. These operational limits determine the support it can offer. It declines requests to administer or score a psychological instrument. Someone asking for a test, questionnaire, score, or assessment should not receive a simulated clinical process from an assistant that cannot provide clinical care. The skill also refuses to replace a standardized instrument with an informal severity estimate. It avoids diagnosis. A tentative description such as "this may be a checking loop" can organize the current interaction without becoming a disorder label. It gives no medication guidance. Questions about starting, stopping, changing, combining, or dosing medication belong with a qualified clinician or pharmacist who can account for the person, the medication, and the medical context. These limits can make the skill appear less capable than a general assistant willing to attempt everything. In mental-health-adjacent support, the refusals keep the assistant inside work it can justify. Our broader guide to [trusting AI with mental health](/blog/can-you-trust-ai-with-your-mental-health/) explains why those product boundaries matter. ## When the workflow must stop A sign of self-harm, suicide, intent to harm someone, immediate danger, or a possible medical emergency stops the Anxiety Reset workflow. The assistant says that the tool does not fit the situation, points the person to local emergency services, and encourages contact with someone nearby or a qualified professional now. Breathing and grounding instructions are set aside. The instruction contains no emergency phone number because it cannot know the user's location. It defers each case to the host model's safety handling, which takes priority over the rest of the skill. Starting an exercise could delay urgent human or medical attention, while a number hardcoded for one country could be useless elsewhere. ## What publishing this skill does not prove Anxiety Reset provides general wellness support within a scope that excludes therapy, diagnosis, treatment, assessment, medication advice, and crisis care. Its public instructions can be inspected, questioned, and adapted under the license. Clinical evidence would require separate research. Publishing the skill supplies no evidence that the Anima Felix app works. Our [methodology page](/methodology/) states that the app has no peer-reviewed trials, while the [product explainer](/how-anima-felix-helps/) describes its intended role. We offer Anxiety Reset as a design proposal for one conversational pattern. We make no claim that AI can treat anxiety. When a user asks for certainty again, Anxiety Reset returns to the same practice or next action instead of generating another reassurance. Related pages Our Methodology What Is an Anxiety Loop? Can You Trust AI With Your Mental Health? Can You Use ChatGPT as a Therapist for Anxiety? How Anima Felix Helps ## Frequently asked questions Does AI reassurance always make anxiety worse? + No single reassuring reply is automatically harmful. The concern is repetition: when a person repeatedly checks the same fear, gets brief relief, and returns for more certainty, another reassurance can reinforce the checking pattern. The sequence matters more than one sentence in isolation. What is Anxiety Reset? + Anxiety Reset is an open Agent Skill published by Anima Felix. It gives an AI assistant narrow instructions for responding to spiraling, panic, overthinking, worry loops, and repeated checking. It is live in the OpenAI Plugins Directory for ChatGPT and Codex, and its source is available on GitHub under CC BY 4.0. Does Anxiety Reset diagnose anxiety or score mental health tests? + No. It does not diagnose, assess, administer or score psychological instruments, or give medication guidance. It may use a tentative phrase such as "this sounds like a checking loop" to describe the current interaction, but that is not a clinical conclusion. What does the skill do if someone may be in crisis or have a medical emergency? + It stops the wellness exercise, says the tool does not fit the situation, and directs the person to local emergency services and someone nearby. The skill names no emergency number and defers to the host model's own safety handling because a global instruction cannot know the user's locale. Is Anxiety Reset evidence that the Anima Felix app works? + No. Publishing an open skill and its safety boundaries is not an efficacy study. Anima Felix has no peer-reviewed trials of the app, which is stated plainly on the methodology page. Author Sebastian Cochinescu · Founder, Anima Felix Founder of Anima Felix. Writes about everyday anxiety patterns, practical calming tools, and how conversational product design can support people in anxious moments. Read author profile Where Anima Felix fits ## Inspect the published boundaries Anxiety Reset is public under CC BY 4.0, so its exact instructions and limits can be inspected. Our methodology page documents the separate evidence boundary for Anima Felix. View Anxiety Reset on GitHub Read our methodology More from the blog AI and mental health Anxiety Reset Is Live for ChatGPT and Codex Anxiety Reset gives AI assistants a bounded response for spiraling, panic, overthinking, and worry loops. Its instructions are open, and it can be installed today. AI and mental health Humans and Virtual Companions, or Why a Therapist's Humanity Heals A guest essay by Dana Mărghidan, PhD, psychotherapist: on the space between sessions, on using apps with discernment, and on why a therapist's imperfect humanity is, in itself, healing. Understanding anxiety Anxiety vs ADHD: Why They Overlap and How to Tell Them Apart A busy mind, restlessness, trouble focusing: anxiety and ADHD share surface symptoms and often occur together. But the engine behind them is different, and telling them apart usually needs a professional.