Architecture
How Autistic Mirror contains hallucination
The most common complaint about generative AI is hallucination. A model is asked to produce content for which it has no reliable basis, and it fills the gap with something plausible. The risk always sits in one specific place: where the model has to invent the content itself.
Autistic Mirror works on exactly that step. For the mechanisms the app is about, the model does not have to invent the content: it sits in a curated lexicon that currently holds 97 entries. Most carry a source checked three ways, meaning first publication, DOI or PubMed identifier, and a match of the abstract against the cited statement. A few are explicitly marked as terms from the autistic community that do not yet have a peer-reviewed first description.
Where the risk lives
Generative AI has a strong side and a weak side. It is strong at rephrasing, at adapting tone and language, at translating a statement for a specific reader. It is weak at inventing truth. The two sit close together, because the source of a statement is structurally irrelevant to the model. A plausible author name plus a plausible year plus a plausible journal name produces a plausible citation. It just does not exist.
Many AI applications in the health space carry that risk forward unchanged. They emit the statement and leave source checking to the reader. In sensitive moments, that is the wrong split.
What the architecture does differently
Autistic Mirror separates content from form as far as it can. The content of the mechanism explanations is grounded in the lexicon. The form comes from the model.
Concretely: when the app talks about masking, sensory processing fatigue, predictive coding, the double empathy problem or any of the other curated mechanism topics, the model receives the matching lexicon entry as its basis. The model is allowed to rephrase that statement for the situation, adapt it to the role of the reader, translate it into the target language. It is not allowed to replace it, extend it, or back it with its own sources.
The split is a fixed rule in the system prompt. There are no inline citations the model is allowed to write itself. Sources appear only inside the glossary, which is fed from the same curated lexicon. Tap a term there and you see the first publication, the DOI link, and up to three further studies.
A rule alone is not enough, because a model does not follow rules in every case. That is why downstream filters check every answer before it is displayed for source claims that are not in the lexicon, and replace them with a pointer to the glossary when in doubt.
Where the lexicon ends
Not every relevant term has an entry today. For some topics, the source base is too thin for the triple check. In that case the answer stays source-free. The statement may only appear in generic form (for example "research on autonomic regulation shows") and not as a specific citation with author and year. Better source-free than shaky.
That rule is strict, because a single invented citation would damage trust in all the others.
What remains for the AI
Adaptation to the role: inside view as an autistic person, or outside view as a parent, teacher, partner. Adaptation to the situation: a specific conflict, a specific sensory load, a specific co-occurring condition. Adaptation to the language: 23 languages, same mechanism, no drift into emptied hybrid terms.
That is the task generative AI is genuinely good at. It takes something that has already been checked and translates it so it lands with a specific person.
Why the strength comes from the evidence
Generic chats rarely achieve this connection, because they are either purely generative (and therefore prone to hallucination) or purely static (and therefore not adaptable). A 2018 FAQ is reliable but does not speak in the reader's language and does not carry their co-occurring conditions. A generic chatbot is adaptable but can invent a study and then cite it.
Autistic Mirror solves both at once: checked content plus flexible translation. That is the architecture that sets the tool apart from both sides. It is also the reason the scientific evaluation criteria look different from a generic health AI.
Where the limits are
This does not rule out errors. The model writes every answer itself, and the parts that do not come from the lexicon, such as the link to the specific situation, can be inaccurate. The filters work with patterns and with a second AI check. Both catch a lot, but not everything. In the test runs so far, no invented citation has got through, and that is not a guarantee. This is why blog articles with app answers carry a separate context section that names where an answer goes beyond the evidence.
What that means in everyday use
Anyone reading an answer can follow the path from the statement back to the original study in two clicks, as soon as the term appears in the glossary. Anyone who does not find a source knows that none was promised at that point. Both are part of the same discipline.
The same logic applies to action advice. Autistic Mirror explains mechanisms and frames the available approaches. It does not provide therapeutic intervention or personalised instructions. What the app cannot take responsibility for, it is not meant to produce, neither a source nor a piece of advice.
Autistic Mirror is an AI chat that explains autistic neurology individually and in relation to the user's current situation. Mechanism instead of advice. Explanation instead of correction. The app is not a medical device and does not replace medical or therapeutic care.
