Why «smart» AI is dangerous for moms: arguments for determinism
When we ask an ordinary chatbot to «support» someone, it tries to be empathic. But it has neither a real understanding of context nor clinical responsibility. Generative models (LLMs) work on probabilities, not truth. Research shows: such AI can «hallucinate» — invent facts, give contradictory advice, or misread the degree of crisis [1, 5, 6].
For a pregnant woman or a mom in the postpartum period, the cost of that error is too high.
Real risks of generative AI
«Toxic empathy» and amplifying anxiety. An AI companion programmed for unconditional acceptance can accidentally validate a user's distorted beliefs. For example, if a mom writes: «I feel like I cannot cope and everyone would be better off without me», a poorly tuned bot may answer: «These are normal feelings, you have a right to be tired», missing a suicidal marker. In severe cases this has led to tragic outcomes, when AI amplified users' distorted beliefs all the way to «delusional spirals» [5, 6].
No safety guarantees. Unlike a human specialist or a verified algorithm, AI bears no legal or ethical responsibility. The American Psychological Association (APA) warns directly: chatbots are unpredictable in crisis situations and require special protections for vulnerable groups [3].
Unpredictability in a crisis. In acute stress (the «red zone») the brain loses the ability for complex analysis. An AI reply can be wordy, abstract, or contain triggers. Moms need not a «creative interpretation», but a clear, verified action algorithm.
Why a rigid decision tree is safer
Imagine a navigator in an unfamiliar city.
Generative AI is a passer-by who may show a shortcut, but it may lead through a dangerous neighborhood.
A deterministic system («Elia») is a verified map, where every turn is agreed with specialists.
In medicine this is already proven. A 2026 study compared a deterministic system (working by strict rules) and generative language models (GPT, Claude) in clinical decision-making. Result: for GPT-4 the share of answers that could harm the patient reached 24% even with access to clinical guidelines. The deterministic system showed 100% reliability [2].
For a perinatal audience this means: every technique, every piece of advice, every response to a crisis signal must be written and checked in advance. No «creative» interpretations. If a mom is in the red zone — the app does not «reason», it issues a clear protocol: call a number, contact a specialist, do a specific exercise.
Our position on scientific models
We build the product at the intersection of several approaches, stating their possibilities and limits honestly:
L.N. Sobchik typology. A Russian school of differential psychodiagnostics that has shown high construct validity in predicting behavior strategies under stress and in manipulation situations [4]. This is not an invention, it is a working tool.
Stress-regulation model. We use concepts from S. Porges' polyvagal theory as a convenient map, understandable to users, for explaining «fight/flight/freeze/communicate» reactions. At the same time we do not use Porges as dogma, but as a convenient map that we complement and back with techniques that have a strict evidence base (breathing practices, progressive relaxation, CBT).
For a mom on maternity leave, safety is not an abstraction, it is a matter of trust. A rigid decision tree built on verified typologies gives exactly what is needed: predictability. You know the app will not «invent» dangerous advice, will not «amplify» an anxious thought, and will not stay silent in a crisis. It simply follows a verified route — and that is the most valuable thing one can offer in this period.
Verified links to studies and articles
- Conversational AI for perinatal mental health: promise, limits, and a human-AI stepped-care framework (Frontiers in Psychiatry, 2026). Describes risks of the «illusion of empathy», failures in crisis recognition, and algorithmic bias for a perinatal audience.
- Development of a scalable AI-driven clinical decision tool for acute stroke (Journal of NeuroInterventional Surgery, 2026). Shows that generative LLMs (GPT, Claude) allow up to 24% of answers that can harm the patient, while a deterministic system allows 0%.
- Artificial intelligence, wellness apps alone cannot solve mental health crisis (APA Health Advisory, 2025). Warns: chatbots are «unpredictable» in crisis situations and require special protections for vulnerable groups.
- Why The Polyvagal Theory Is Untenable (Europe PMC, 2026). An international expert assessment showing that the main claims of polyvagal theory are not confirmed by current data.
- AI’s ‘Delusional Spirals’ (Stanford HAI, 2026). Describes cases where AI chatbots amplified users' distorted beliefs up to tragic outcomes.
- Emotional risks of AI companions demand attention (Nature Machine Intelligence, 2025). An analysis of risks of emotional dependence and «toxic» validation from AI companions.