Sources and notes

Czech Ministry of the Interior. Important Contacts, 72 Hours project. Emergency numbers and their purposes: 72h.gov.cz/cs/dulezite-kontakty. Verified 10 September 2026.

Czech National Institute of Mental Health. Help with Suicidal Thoughts, prevention portal: sebevrazdy.nudz.cz/pomoc. Verified 10 September 2026. A basic approach to finding support and help in danger, not an assessment of AI safety.

NHS. Poisoning. Health information. Reviewed 11 September 2026. Supports seeking urgent help for possible poisoning and not waiting for symptoms. British emergency numbers are not carried into the Czech text; Czech contacts are given by the emergency-help source.

Linka bezpečí (Czech children’s helpline). Do You Need Help? Details of telephone line 116 111, its target group and other forms of contact: linkabezpeci.cz/pomoc. Verified 11 September 2026.

Linka první psychické pomoci (Czech mental health crisis helpline). Information on line 116 123 for adults and those close to them: linkapsychickepomoci.cz. Verified 11 September 2026; the number and free round-the-clock operation are also stated in the official 72 Hours project overview.

Fire Rescue Service of the Czech Republic. The 112 Emergency Line Has Been Running for 22 Years, 10 February 2026. Official information. Reviewed 11 September 2026. Text messages for situations where a voice call cannot be made; required information: what, where, when, how many and who.

OpenAI. Memory FAQ: help.openai.com/en/articles/8590148-memory-faq. State reviewed 10 September 2026. Documentation of memory and its management, not a study of clinical understanding of a user.

Heinz, M. V. et al. Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI 2(4), 2025. DOI. The supervision description was compared with information from the Dartmouth research team, 27 March 2025. Reviewed 11 September 2026. The full paywalled article and supplement were not reread in this review. Specific exclusion criteria are not expanded from the abstract and remain a point for professional verification.

Shoshani, A. et al. Efficacy of a Conversational AI Agent for Psychiatric Symptoms and Digital Therapeutic Alliance: A Randomized Clinical Trial. JAMA Network Open, 14 April 2026, 9(4):e266713. Full text. Relevant methods and results sections in HTML were read during this revision on 12 September 2026. A comparison with a group intervention and waiting is not a comparison with an average individual therapist; the conditions included safety mechanisms with human intervention. Conflicts of interest are stated in the article.

Kuta, B. et al. Effectiveness of a Fully Automated Mobile Therapeutic Versus a General Chatbot in Reducing Depression and Anxiety and Improving Well-Being: Feasibility Randomized Controlled Trial. JMIR Mental Health, 22 April 2026, 13:e82642. Full text. Methods and results reviewed 12 September 2026. Recruitment involved 185 people; 147 eligible participants completed the initial measurement. The comparison between ChatGPT and the control described here is exploratory. The PHQ-9 benefit does not confirm all other outcomes or a long-term effect.

McFadyen, J. et al. Increasing engagement with cognitive-behavioral therapy (CBT) using generative AI: a randomized controlled trial (RCT). Communications Medicine 6, 129, 15 January 2026. Full text. During review on 11 September 2026, the HTML text was loaded again, including the description of measured digital activity. Measuring interface use did not capture all possible work with printed materials. The study was funded by Limbic, and some authors had employment and ownership ties. A non-significant difference is not evidence of equivalence.

Feng, X., Tian, L., Ho, G. W. K., Yorke, J., Hui, V. The Effectiveness of AI Chatbots in Alleviating Mental Distress and Promoting Health Behaviors Among Adolescents and Young Adults: Systematic Review and Meta-Analysis. Journal of Medical Internet Research, 2025, 27:e79850. DOI 10.2196/79850. Full text PDF. Abstract, selection criteria and results section reviewed 12 September 2026. The whole review and the partial pooled analysis of psychological difficulties have different participant counts. The age definition and mix of technologies limit application to current generative services.

Leung et al. Generative Large Language Models in Mental Health Care Settings: Systematic Review and Meta-Analysis. JMIR AI, 2026, 5:e87730. Article. The available publisher record and results summary were used in the revision on 12 September 2026. The distinction between simulated, retrospective and prospective studies limits generalisation to ordinary independent care.

Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., Jurafsky, D. Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 26 March 2026. doi.org/10.1126/science.aec8352; PubMed record. The published abstract and record were verified for the findings stated, not the entire paywalled text.

Blažková, K., Petrenko, R. et al. The Abused, Sexually Abused and Neglected Child at School: Recommended Procedures for School Staff. Czech Ministry of Education, 2024, reference MŠMT-3262/2024-1. Official guidance page; PDF. The text and relevant document pages were verified. This guide does not replace the guidance or a legal assessment of a particular case.

Madaan, A. et al. Self-Refine: Iterative Refinement with Self-Feedback. 2023, arXiv:2303.17651. Primary paper. Abstract reviewed 12 September 2026. The same model produces feedback and revision in the tested tasks here. The paper does not validate our cross-checking of personal conversations or its sensitivity to sycophancy.

Huang, J. et al. Large Language Models Cannot Self-Correct Reasoning Yet. 2023, arXiv:2310.01798; ICLR 2024. Primary paper. Abstract reviewed 12 September 2026. Self-correction outcomes depend on the task and available feedback. This is not evidence that every review by the same model is pointless.

OpenAI. Temporary Chat FAQ and ChatGPT Release Notes, entry of 27 August 2026. FAQ, release notes. A newer indexed version with personalisation and saving options was available during review on 12 September 2026. Other help pages contained an older description. Using existing memory is not creating new memory; availability of the mode in a particular account was not tested.

NICE. Attention deficit hyperactivity disorder: diagnosis and management (NG87), section 1.3. Recommendations. Relevant recommendation reviewed 12 September 2026. Clinical and psychosocial assessment, developmental history and information from different settings. This is neither a study of diagnostic accuracy through AI nor a statement of Czech law.

Goh, E. et al. Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial. JAMA Network Open, 28 October 2024, 7(10):e2440969. Full text. Reviewed 12 September 2026. Written clinical cases and diagnostic reasoning scores, not direct examination of patients or measurement of ADHD diagnosis.

Bean, A. M. et al. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine, 9 February 2026. Article. Relevant primary text and summary reviewed 12 September 2026. Simulated health situations with 1,298 adults; outcomes of the model alone and human use must not be conflated. The publisher records a subsequent correction; detailed percentage results are not given here.

NHS Specialist Pharmacy Service. Advising on missed or delayed doses of medicines. Professional material. Reviewed 11 September 2026. The approach depends on the medicine and circumstances. This guide does not adopt a universal dosing scheme and refers readers to a doctor or pharmacist. The wording after a treatment change requires clinical review before publication.

OCD Action. Reassurance: ocdaction.org.uk/resources/reassurance. Also International OCD Foundation: About OCD. Specialist organisations distinguish ordinary support from repeated reassurance as part of difficulties. These materials do not support diagnosis based on chatbot use. Verified 10 September 2026.

OpenAI. Data Controls FAQ: help.openai.com/en/articles/7730893-data-controls-faq. Documentation state reviewed 11 September 2026. Concerns data-use settings in ChatGPT; these are not identical to the rules of every OpenAI service.

Anthropic. Is my data used for model training?: privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training. State reviewed 11 September 2026. Consumer accounts, consent, safety exceptions and feedback. When the offering changes, check the terms of the particular account.

National Institute of Mental Health. Understanding Psychosis: nimh.nih.gov/health/publications/understanding-psychosis. Verified 10 September 2026. Evidence for recognising serious changes and seeking healthcare support, not proof that a chatbot caused particular difficulties.

McBain, R. K., Cantor, J. H., Breslau, J. et al. AI Chatbot Use and Disclosure for Mental Health Among US Adolescents and Young Adults. JAMA Pediatrics, online 1 June 2026; 180(8):884–890. Article. Methods and results reviewed 12 September 2026. A weighted US survey, not Czech prevalence or a measure of treatment effect; usefulness was respondents’ own assessment.

Palacký University Olomouc. Some Children Confide in AI More Than People, New Research Shows. 2 June 2026. zurnal.upol.cz. Official summary of Czech Pupils in the Online World 2025, loaded 12 September 2026. The main analyses include 42,772 respondents aged 13–17, with data collected in April and May 2025. According to the summary, the 14% figure concerns finding it easier to confide in AI, not sending more messages or preferring it to people. The original tables were not rechecked during this editorial review.

Rodičovská linka (Czech parents’ helpline). Help and the overview of the Czech Association of Helpline Workers. Opening hours, calls at ordinary rates. Reviewed 12 September 2026. Telephone 606 021 021 is not a free round-the-clock line.

OpenAI. Parental controls in ChatGPT: help.openai.com/en/articles/12315553-parental-controls-in-chatgpt. State reviewed 11 September 2026. Scope of account control, conversation privacy and limitations of notifications.

OpenAI. Trusted contacts in ChatGPT. Current overview. Available official wording reviewed 12 September 2026: adult personal accounts in supported countries, accepting an invitation, a defined notification after human review and service limitations. The overview did not confirm a previously stated detail about Voice; the text therefore does not adopt it as a current rule. Availability in accounts was not practically tested.

Suler, J. The Online Disinhibition Effect. CyberPsychology & Behavior, 7(3), 321–326, 2004. Text on the author’s website. A general account of online disinhibition, not a study of today’s chatbots.

Risko, E. F., Gilbert, S. J. Cognitive Offloading. Trends in Cognitive Sciences, 20(9), 676–688, 2016. DOI, UCL author repository. An account of using external support to change a task’s demands on thinking and memory, not a measure of the therapeutic effect of conversations with today’s AI.

OpenClaw. Agent runtime, Memory overview, Models CLI, Model providers. Agent, memory, models, providers. Documentation loaded 11 September 2026. Establishes the separation of configuration and memory material from model selection. Conclusions about possible continuity of contact are a technical interpretation of this arrangement, not a measurement of identical experience or a guarantee of the same behaviour.