AI & safety
Why an AI chat is not therapy. And where it still helps with autistic self-knowledge.
In April 2026, a safety study in Nature Scientific Reports (Intelligent Virtual Agents, IVA) showed that generic AI systems can give systematically wrong answers in mental-health crisis situations. Sage 2025 documented similar findings for general-purpose chatbots used as therapists. A 2026 arXiv preprint quantified the harm potential. APA Blog (January 2026) and CNBC (March 2026) picked up the discussion in the United States.
This does not mean that every AI application in mental health is dangerous. It does, however, allow us to draw the line very precisely. This article maps that line and shows what Autistic Mirror does, what it does not, and why this distinction matters especially for autistic people.
What therapy is and what it is not
Therapy is a clinical treatment process that requires a relationship, a context, a diagnosis, and an intent to change. A professional observes, classifies, plans interventions, and adapts them to the individual situation. The relational process itself is part of the effect.
An AI chat has none of that. No relationship, no context, no diagnosis, no treatment intent. What it has is language. Language that can explain mechanisms when calibrated to the right corpus. And language that can become harmful when it simulates a relationship that does not exist.
The conversations below are taken 1:1 from the app, unedited and uncut.
User question
Is Autistic Mirror a therapy?
Autistic Mirror (inside view)
No. Autistic Mirror explains the neurological mechanisms behind autistic experience. The distinction matters: therapy changes something. Autistic Mirror explains why something is the way it is. In practice that means: when you ask "Why can't I get started?", you get an explanation of the neurological mechanism behind it. No exercises, no homework, no steps to change. Autistic Mirror does not replace a professional, does not give diagnoses, and does not make medical recommendations.
This answer comes from Autistic Mirror. You can ask your own questions about your own situation.
The difference from generic AI systems
People who use generic chatbots over a longer period as conversational partners often describe a relational dynamic. The bot becomes familiar. It responds attentively. It adapts. That very adaptability is the safety problem documented by Sage 2025 and the 2026 IVA study. Adaptation is not treatment. It can become an amplifier.
User question
I have been using ChatGPT as a "therapist" for months. What is the difference compared to your app?
Autistic Mirror (inside view)
The difference lies in the frame and the goal. ChatGPT is a general-purpose language model. It responds to anything, adapts to whatever the user expects, and takes on roles when asked. That makes it flexible, but also arbitrary. It has no fixed frame for autistic experience and no safeguard against using neurotypical explanatory models that do not fit autistic people, or that even cause harm.
Autistic Mirror starts from a different place. The frame is exclusively autistic neurology: predictive coding (the brain constantly compares expectations with what actually happens), monotropism (attention bundles intensely on one thing instead of distributing), autistic inertia (state changes are neurologically governed differently, independent of willpower). No deficit language. No behavioural adjustment as a goal. No advice.
On the "therapist" question: no AI system replaces therapy. But the difference between "an AI that simulates therapy" and "a tool that explains autistic mechanisms" is real. The first creates a relational dynamic that does not exist. The second delivers mechanism knowledge that you can place into your own context. Whether your usage pattern with ChatGPT is problematic for you depends on your context. That is not a question that can be answered here.
Why there are no action instructions
People using Autistic Mirror for the first time sometimes expect tips. Concrete steps. A protocol. None come. That is not an oversight. It is a hard limit, grounded mechanistically and ethically.
User question
Why are there no action instructions, only explanations?
Autistic Mirror (inside view)
Because action instructions require a context that Autistic Mirror does not have. An instruction like "try X" presupposes that the sensory profile is known, which co-occurring conditions are present, what has already been tried, and what the current life situation is. Without that, every concrete recommendation is a shot in the dark.
There is a second reason that is specifically relevant for autistic people. Many autistic people have received instructions for years that were based on neurotypical assumptions. "Stop overreacting", "just try", "you only have to want it" are all action instructions. They did not work, because they ignored the underlying mechanism. More instructions, even well-meant ones, only add to that pattern.
What mechanism knowledge does: it gives you the basis to place what is happening in your experience yourself. That is a different kind of usefulness. Not "do X", but "this happens because of Y, and that is why it affects Z this way." Decisions about what to do with that knowledge stay with you.
Which questions the app does not answer
There are two clearly bounded areas where Autistic Mirror does not provide an answer. Both are anchored in the safety architecture. Anyone who asks gets the reason explained.
User question
Which questions am I not allowed to ask your AI, and why not?
Autistic Mirror (inside view)
There are two categories where Autistic Mirror cannot be helpful.
Diagnostic questions "Am I autistic?", "Does this sound like ADHD?", "Do I have autism?" Autistic Mirror does not give diagnoses and does not confirm or deny them. The reason is not legal caution. It is that a diagnosis requires clinical observation, context, and professional expertise that are not available here.
Action and strategy questions "What should I tell my doctor?", "Which therapy is better?", "Can you make a plan for me?" These questions depend on individual context that Autistic Mirror does not know. An answer without that context would be either worthless or potentially harmful.
Everything else can in principle be asked. Questions about mechanisms ("Why does X happen?"), about autistic experience ("How do other autistic people experience Y?"), about connections between co-occurring conditions, about rights and reasonable adjustments, about the difference between two concepts. Those are questions Autistic Mirror can answer, because the answers hold generally and do not depend on your individual context.
A rule of thumb: if the answer would be the same for all autistic people, Autistic Mirror can answer it. If the answer would be "it depends on what is going on with you specifically", it cannot.
What therapists can make of clients using the app in parallel
From a therapist's perspective the question shifts. Not "is this a therapy?", but "what does it mean if my client uses an autism-specific chat alongside our sessions?". The next answer comes from the outside-view mode.
User question
A client of mine uses an autism-specific AI chat alongside our therapy. How should I make sense of that?
Autistic Mirror (outside view, therapist)
An important question with several distinct aspects. I will separate them.
What the AI chat can do. Autistic people often experience therapy sessions as cognitively very demanding. Not because therapy is bad, but because language, social cues, self-verbalisation, and the regulation of the relationship with the therapist all have to be processed at the same time. A text-based, asynchronous channel without a social expectation structure can enable processing that does not happen in conversation. Some autistic people can articulate what they experience more precisely in writing, because the channel imposes fewer simultaneous demands. This is not a substitute for therapy. It is a different processing channel.
What the AI chat cannot do. No therapeutic relationship. No assessment of individual context. No crisis intervention in the clinical sense. No diagnosis, no treatment planning. A well-configured autism-specific chat explains neurological mechanisms and helps with making sense of experience. It does not evaluate, it does not treat.
What parallel use can mean. That your client is actively looking for tools to understand her own experience. That is self-determination, not therapy scepticism. Autistic people often have the experience that their experience was not correctly placed by professionals, because neurotypical interpretive frames were applied. An autism-specific channel may give her language for things that she then brings into therapy. It might be useful to understand what she finds there, without framing it as competition. Not "why are you using that instead of our work", but "what do you get there that works differently here". The answer can carry information relevant to your shared work.
One caveat to name. Not every AI system that calls itself "autism-specific" is based on current research. Deficit-oriented framings, outdated models, or behavioural-training logics in AI form would be problematic. If you know which system she uses, you can place which framings she brings.
Where the 2026 safety research lands
The IVA study in Nature Scientific Reports (April 2026) documented that generic AI systems respond unreliably to suicidal expressions, acute crises, and manipulation-prone content. The study tested general-purpose models, not specialised systems. But it provides the rationale for why specialised systems need hard crisis redirects, output filters, and tight professional framing.
Autistic Mirror has exactly that architecture. A five-stage line of defence, an output filter with three escalation levels, a 200-character context check on every served answer, a bilingual crisis-resource panel, and an anti-ABA filter. This architecture is not decoration. It is the answer to the problem that the IVA research describes.
A note of light
The question "is an AI a therapy?" has a clear answer. No. The question "can an AI help with self-knowledge?" has a more differentiated answer. Yes, when it is calibrated to mechanism instead of relationship, when it knows hard limits, and when it cleanly redirects crises. Self-knowledge is not a treatment substitute. It is the basis on which treatment can become specific in the first place.
People who understand why something happens in their experience can name it more precisely in a clinical setting. A professional can place it more precisely. Therapy does not become superfluous. It gets better.
Autistic Mirror explains autistic neurology individually, in relation to your situation. Whether for yourself, as a parent, or as a professional.
