Skip to content

Freshly launched: until 30 November something may still go wrong – send us feedback and we’ll fix it fast!

Legal

AI transparency

Last updated: 15 November 2026

This is an English translation for your convenience. Only the German version is legally binding. Read the German version

Stabily is an AI system. When you chat with Stabily, the answers do not come from a person but from a language model. Here we explain who is behind Stabily, what Stabily is intended for, how it arrives at its answers and where its limits lie.

This page is based on the European Artificial Intelligence Regulation (Regulation (EU) 2024/1689, known in English as the “AI Act”). It supplements our Terms of use and our Privacy policy.

You are chatting with an AI

Stabily is an AI conflict coach for teachers. Its answers are generated automatically by a language model. Nobody from our team reads along before you receive an answer, and nobody writes answers by hand. The only exception is the answer to the “What can you do?” button: a fixed text that our team wrote in advance and that is the same for everyone (see “What can you do?”).

The AI Act requires that you know this from the outset (Art. 50(1)). That is why Stabily introduces itself in the chat as an AI conflict coach before you send your first message. In addition, on every page of the chat, a notice below the input field says that Stabily can make mistakes and does not replace a therapist or a counselling service, with a link to this page.

If you dictate a message using the microphone in the input field, an AI model likewise converts your recording into text. This text may contain errors. It appears in your input field, and it only goes to Stabily as your message once you have read it and sent it yourself.

In voice mode, you speak with Stabily: an AI model continuously converts your words into text, and as soon as you pause, this text goes to Stabily as your message without any further step. You see it in the chat; it may contain errors. Stabily reads the answer aloud in an artificial voice generated by an AI model; it is not a real person. Whatever is read aloud also appears in the chat, with the sources.

How answers are labelled

Every answer is visibly shown in the chat as an answer from Stabily, separate from your messages. Even when you export a conversation, “Stabily” appears above every answer.

The AI Act also requires that texts generated by AI are marked as artificially generated in a machine-readable format, as far as this is technically feasible (Art. 50(2)). Stabily’s answers appear in the chat, which is labelled as AI as described above, and in the text file that you save from the chat’s menu under “Export conversation”. We label the action plan and adapted worksheets, which you can download as PDFs (see How Stabily arrives at its answers), visibly on every page with “Created with AI (Stabily, Mistral AI)” or “Adapted with AI (Stabily, Mistral AI)” respectively, and in a machine-readable way in the file’s metadata (subject) as “AI-generated”; a PDF from a chat in German is in German and carries “Mit KI erstellt (Stabily, Mistral AI)” or “Mit KI angepasst (Stabily, Mistral AI)” and “KI-generiert” instead. [to be checked: whether this labelling in the PDF is sufficient for Art. 50(2) or whether a dedicated metadata standard is needed]

Who is behind Stabily

Us: the provider of Stabily

We, the association Stabil im Kopf – Mentale Gesundheit für Jugendliche in Vienna, developed Stabily and offer it under the name Stabily. This makes us the provider of the AI system Stabily within the meaning of the AI Act, even though Stabily is free of charge. We determine what Stabily is intended for, the rules by which the coach answers and which sources go into its knowledge base. You can find further details about us in the Legal notice.

Mistral AI: the provider of the language models

The language models on which Stabily is built come from Mistral AI SAS, based in Paris, France. Mistral AI is the provider of these general-purpose AI models and has its own obligations under the AI Act for them, for example regarding technical documentation and copyright. We integrate these models into Stabily and are therefore, in the language of the AI Act, a “downstream provider”. We have not trained or modified the models. Which models we use is set out under Which models Stabily uses.

You: the user

You use Stabily and decide what you do with the answers. If you use Stabily professionally, for example as a teacher, you may count as a deployer within the meaning of the AI Act. For a system like Stabily, this mainly means one thing: if you publish a text from Stabily in order to inform the public on matters of public interest, you must disclose that it was generated with AI, unless you have reviewed it and bear editorial responsibility for it (Art. 50(4)).

We are developing additional, paid features for schools and other institutions and platforms in the education sector. Who has which role there is governed by separate terms.

What Stabily is intended for and what it is not

Stabily supports teachers with conflicts in everyday school life. You describe a situation, Stabily asks you a few targeted questions and suggests ways of dealing with the conflict, for example conversation guides or short interventions. Where appropriate, Stabily offers you answers you can click. Stabily is aimed at teachers and contains no services for children and young people. You may use Stabily from the age of 18 (see Terms of use).

Stabily is not intended to

  • assess pupils or make decisions about them, for example about marks, performance assessments, admission to a school, assignment to a class or type of school, disciplinary measures or suspensions;
  • monitor pupils during tests;
  • select, assess or monitor teachers or other employees;
  • make diagnoses, replace therapy or give psychological, medical or legal advice;
  • assess emergencies or call for help: Stabily is not an emergency service;
  • be used by children and young people themselves.

These limits also matter for the classification under the AI Act (see the next section). Our Terms of use therefore exclude using Stabily to assess or make decisions about pupils. Anyone who uses Stabily for this nonetheless is acting outside its intended purpose.

How we classify Stabily under the AI Act

The AI Act classifies AI systems according to their risk: some practices are prohibited, strict requirements apply to high-risk AI systems, and systems such as chatbots are mainly subject to transparency obligations. What counts is the purpose for which the provider intends a system. Our assessment for Stabily:

No prohibited practices

Stabily does not influence anyone with subliminal, manipulative or deceptive techniques, does not exploit anyone’s vulnerability, for example due to age, does not evaluate people on the basis of their social behaviour (social scoring) and does not create profiles of individuals.

Nor does Stabily recognise emotions within the meaning of the AI Act. This would in principle be prohibited in educational institutions (Art. 5(1)(f)). Emotion recognition there means systems that infer emotions from biometric data, for example from a face or a voice. In the chat, the coach only receives text. If you dictate a message or speak in voice mode, an AI model only converts your voice into text; it is not analysed in order to recognise emotions, characteristics or you as a person. The automatic safety check of your messages (see How Stabily arrives at its answers) checks the text of a message, not you as a person. Among other things, it ensures that the coach first points to emergency numbers and support services if there are signs of danger.

Not a high-risk AI system

The AI Act lists exhaustively which AI systems are high-risk: AI in certain products such as medical devices (Annex I) and AI for certain purposes in sensitive areas (Annex III). We have checked Stabily above all against these areas:

  • Education (Annex III, point 3). AI systems are high-risk if they decide on access or admission to educational institutions or assign persons to them, evaluate learning outcomes (including in order to steer the learning process), assess the appropriate level of education of a person or monitor prohibited behaviour during tests. Stabily does none of this: it helps teachers to understand conflicts and plan next steps. It assesses neither pupils nor their performance and makes no decisions about anyone. If Stabily asks about the country or the grade, it does so only to find suitable content, not to classify a person. This also applies to the action plan: it summarises the steps that you plan yourself, does not assess any pupils, does not create profiles and makes no decisions about anyone. Whether, according to what you have said, criteria for bullying apply describes the situation, not a person, and you decide this yourself: the app only takes it from your own answers.
  • Biometrics (Annex III, point 1). Stabily does not analyse biometric features, whether from images, voices or videos. When you dictate and in voice mode, your voice is only converted into text. Stabily does not recognise anyone by their voice, does not identify anyone, does not categorise anyone according to sensitive characteristics and does not recognise emotions.
  • Emergency calls and essential services (Annex III, point 5). Stabily does not receive emergency calls, does not evaluate or prioritise them and does not dispatch emergency responders. If there are signs of acute danger, Stabily names emergency numbers; this does not replace calling them. Nor does Stabily make decisions about public assistance benefits, healthcare services, credit or insurance.
  • Employment (Annex III, point 4). Stabily is not intended to select staff, allocate tasks on the basis of personal traits or evaluate the performance and behaviour of employees.
  • Medical devices (Annex I). Stabily is not intended to detect, prevent or treat illnesses, including mental illnesses, and is not a medical device.

Because Stabily does not serve any of these purposes, it is not a high-risk AI system in our assessment. Stabily is therefore mainly subject to the transparency obligations (Art. 50) and the obligation to support our team’s AI literacy (Art. 4).

This assessment depends on the intended purpose. Before we substantially expand Stabily, for example with the planned features for schools, we review the classification again. Should Stabily serve a purpose listed in Annex III in future, the stricter obligations for high-risk AI systems apply, including risk management, technical documentation, human oversight and registration in the EU database.

How Stabily arrives at its answers

When you send a message, the following happens:

  1. Safety check. A separate model automatically checks your message for signs of danger, such as self-harm or violence, and for misuse. If there are signs of acute danger, the coach first names emergency numbers and support services. The same model also checks the coach’s finished answer, with the same thresholds and with your message as context: this way, signs that only become clear in the course of the conversation are picked up too. If it finds a sign of acute danger in the answer, Stabily shows the emergency notice below the answer. The same applies to the summary of the guided intake and to the short answer there to a question you ask in between; both are written by a language model. The check does not cover clarifying questions asked by the planning step, the card and the questions of the guided intake, answers that you stop before they are finished, or any answers in a chat in which there has already been such a sign. Once there has been such a sign in a chat, the app reminds the coach in all further answers in that chat to keep safety in mind, and an “Emergency help” link to the page with the contact points remains visible at the top of the chat; Stabily then shows a new emergency notice when the check finds a new sign in one of your messages. The check may miss signs or misjudge harmless messages. If it is unavailable at the time, the coach answers anyway; its fixed rules for emergencies always apply.
  2. The coach. A language model reads your message together with the conversation so far. It does not remember other conversations. We set out how the coach behaves in instructions: it asks a few targeted questions, offers you answers you can click where you only need to choose, and asks about the country or grade only if the answer depends on it, and never about the country in a chat in English (see Chats in English).
  3. Planning. Before the coach answers, a language model plans its next message in a short intermediate step: if the answer needs specialist knowledge, it sums up the situation you described and your question in a search query for the knowledge base. Depending on the model, it also decides whether a clarifying question is needed first, for example if it is not clear what the matter is about, or if the country is missing for a question about the legal situation (except in a chat in English); Stabily then shows you the clarifying question, often with answers you can click. Usually the model that also writes the answer does the planning; when Mistral Large 4 answers, Mistral Small 4 plans (see Which models Stabily uses). You do not see the intermediate step itself, and we do not store it. If our team has switched on the guided intake, the intermediate steps of that process replace the planning (see The guided intake).
  4. Search in the knowledge base. With this search query, Stabily searches our knowledge base of selected specialist literature and materials before the answer is written. If the planning fails, for example because of a timeout, Stabily searches with your own words instead: with the beginning of your message without links, email addresses and phone numbers, and, for a short message, together with your message before it. In a chat in English, the app adds German search terms from a fixed list that match the English words of your message, because the knowledge base is in German. If the results are not enough, the coach can search once more while answering. The search combines a semantic search, which finds similar meanings, with a keyword search and takes related terms into account. The coach formulates search queries without names of people; if Stabily searches with your own words, they contain the names you wrote. The coach only uses sources that our team has approved.
  5. Answer with sources. The language model formulates the answer using the sections found. If the answer draws on the knowledge base, Stabily names the source with its title and, where available, the page number, so that you can look it up. If Stabily finds nothing suitable in the knowledge base, it is supposed to say so and not invent sources.

“What can you do?”

On the empty start page, everyone without an account sees the “What can you do?” button below the input field. It starts a new chat with exactly this question. The answer to it is not generated by AI: it is a fixed text written by our team, the same for everyone. For this message, neither the safety check nor planning, search or a language model runs. If you then continue writing, the coach answers again as described above, and it reads the fixed answer as part of the conversation history. If you ask in your own words what Stabily can do, the coach answers; its fixed instructions describe the same abilities and require a direct answer without a clarifying question.

The guided intake

Our team can switch on a mode in which Stabily asks a few questions at the start of a case you describe, before it gives a recommendation. The guided intake is not always switched on; when it is off, Stabily answers as described above. When it is switched on, it works like this:

  • Stabily shows a card with key facts: grade, since when, how often and what has already been tried. It does not ask again about anything you have already written. You answer with an ordinary message in the chat.
  • Stabily then asks two to five questions for reflection, one per message. For a conflict involving several people (you yourself count as one of them), it always asks about the conflict stage, otherwise about the urgency; it chooses the other questions from a catalogue. If your goal is still unclear, it also asks about that.
  • Finally, Stabily briefly sums up what it knows so far, asks one last question and then gives the recommendation. The last question first concerns any key facts that are still missing; if you ask about the law or contact points and Stabily does not know your country, it also asks about that, with answers you can click, except in a chat in English. Otherwise it asks what the appropriate approach depends on, if the sections found suggest this, and otherwise whether there is anything else it should know.
  • With the “Skip to recommendation” link you can skip the questions; the recommendation is then less precise. Stabily answers a direct question without a described case directly, at most with the card before it. If you ask a question yourself in between instead of answering, Stabily answers it briefly or announces that it will address it in the recommendation, and then asks its question again. A sign of danger ends the questions; the safety rules then apply as described above. If you do not answer a question but receive an ordinary answer instead, the process is over.

Before each of these messages, Mistral Small 4 reads the conversation in a separate intermediate step. It records what kind of request it is, which key facts are already known and whether your goal is known yet, and writes the search query for the knowledge base; when a question for reflection is due, it chooses the appropriate one from the catalogue and phrases it for your case. For the summary, Mistral Small 4 writes, in a second intermediate step after the app has searched the knowledge base, a few short lines and, if the approaches found depend on a detail that Stabily does not yet know, a question about it. If, instead of answering, you ask a question of your own about Stabily’s question, Mistral Small 4 writes a short answer to it in the first intermediate step; for a question about your case, the app uses a fixed sentence. These intermediate steps replace the planning, whichever model the coach uses. The AI does not decide which step comes next: the app determines the order and number of the questions. Our team writes the questions; it maintains the catalogue in the admin area in versions (see Human oversight); a chat in English always uses a built-in English catalogue. You do not see the intermediate steps themselves, and we do not store them; with each message of the process, we only store a technical marker (see Privacy policy). If the first intermediate step fails, Stabily answers as it would without the guided intake, with the planning; if it fails in the middle of the process, Stabily gives the recommendation straight away.

The summary is written by a language model; the safety check checks it like a finished answer, under the same conditions as described there. The short answer to a question you ask in between is also written by the language model, and the safety check checks it like the summary; it does not check the app’s fixed sentences. If the check is unavailable, the summary or answer appears anyway. The check does not check the card and the questions of this process again as an answer; it checks your messages as always. The summary may state something incorrectly or incompletely: correct it in your next message before Stabily gives the recommendation.

The action plan

Once Stabily has given you a detailed recommendation, you can have the chat summarised in an action plan and download it as a PDF. This only happens at your request, when you open the action plan. Mistral Small 4 writes a draft from the texts of the chat: the situation, the key facts, the steps that Stabily recommended to you and a short rationale with sources. These instructions are set by the app; they cannot be changed in the admin area. The app then checks the draft:

  • The model sums up the situation in its own words and must back up every statement with a passage from your messages. The app checks that this passage appears word for word in your messages and that time references and numbers do not change, so “for days” never becomes “for weeks”. Whether the passage actually supports the statement is something the app cannot check. Whatever the app does not find remains open and does not go into the PDF.
  • The app takes the recommendation and steps only from Stabily’s answers, with their sources. It leaves out new advice that is not in those answers.
  • The model determines which key facts apply according to what you have said, for example whether it is a single incident, and likewise backs this up with a passage from your messages; what you answer yourself in the plan takes precedence. By contrast, whether there is a clear imbalance of power and which of the four criteria for bullying apply only counts if you answer it yourself, with the answers you can click in the plan or in the chat. If the model reads it from your words, it remains an assumption, and the plan asks you about it. From this, the app determines the assessment (for example a minor or serious conflict, recurring problems or bullying) and the urgency according to fixed rules. “Bullying” only appears there if you have confirmed the imbalance of power and all four criteria yourself. The five stages of escalation follow Glasl’s stages 1 to 5.
  • Health information does not belong in the plan. The model is instructed to leave it out; in addition, the app searches for common terms for diagnoses, illnesses and medications, such as ADHD, autism, depression, anxiety disorder, dyslexia, trauma, eating disorder, Ritalin or suicide, in German and, in the plan of a chat in English, also in English. It leaves out a statement about the situation, an assumption or a rationale containing such a term; in the steps, it replaces the term with “[health information removed]” (in the plan of a chat in German with “[Gesundheitsangabe entfernt]”). General safety advice such as “if there are signs of suicidal thoughts, protection and reporting channels come first” remains. If the app detects a sign of suicide or self-harm in your messages or in the situation, the plan begins with the safety notice. A word list does not catch everything, so check the plan before you download it.
  • If information that the plan needs is missing, the plan asks you for it, with answers you can click and “I don’t know”. You can only download it once nothing is left open.

The action plan can be wrong too, for example by overlooking a detail or shortening a recommendation. It is an aid for your own planning, not a diagnosis and not an official record. Check it before you use it or pass it on.

Replacing names

Our team can switch on a mode in which Stabily replaces the names of people in your messages with placeholders. When it is switched on, the following applies to every new message:

  • At the same time as the safety check, Mistral Small 4 lists the names in your message. Fixed rules of the app add names with a form of address or title, such as “Frau Huber”, and in a chat in English, for example, “Mrs Patel”. As a rule, the app does not replace roles such as “pupil” or the names of experts after whom methods are named, such as Glasl.
  • The app replaces the names with “[Person A]”, “[Person B]” and so on before it stores the message. The same name keeps its letter throughout the chat; if you write a person’s name differently at some point, for example only the first name, they may get a second letter. The coach only reads the placeholders and speaks of “Person A”; it is instructed not to ask for the real names.
  • Below your message, Stabily shows you which names it has replaced and, the first time, suggests that you use roles such as “Person B” yourself straight away.

The detection is not complete. It may miss a name or mistake a word for a name, and it does not replace other details by which someone can be recognised, such as class and school. If it does not succeed within a few seconds, the app only replaces what its fixed rules find, and the answer comes anyway. If you dictate or speak in voice mode, Mistral AI first converts your words into text, with the names in it (see What happens to your input); they are only replaced in the message you send. It is therefore best to write and speak without names from the start.

Adapting worksheets

If Stabily cites a worksheet from the knowledge base as a source, you can view it, download it and have it adapted to your class with “Adapt with Stabily”. You then send a message saying what you would like, for example regarding the grade or the length. The app checks whether the worksheet was cited as a source in this chat and gives the coach’s language model the text of the worksheet for this one answer. The instruction for this is set by the app; it cannot be changed in the admin area: a complete, print-ready worksheet with notes for you, instructions and tasks, without names and with a reference to the original. Planning and search in the knowledge base are skipped for this answer. You can download the adapted worksheet as a PDF.

Here too, the AI can be wrong, for example by leaving out a task, changing the content of the original or adding something that does not fit. Check the adapted worksheet before you use it in class.

Chats in English

Stabily is available in German and in English. A chat keeps the language of the page on which you started it: the app stores it with the chat, never changes it and always shows the chat in this language (see Privacy policy). In a chat in English, the following applies:

  • The coach answers in English, even if you write in German. It follows the same guide as in a chat in German, but with fixed rules of its own for chats in English: they instruct it to answer in English and take precedence over the guide where it asks for German or provides for a clarifying question about the country; some other fixed rules, for example on the wording of clarifying questions, are adapted there too. They cannot be changed in the admin area.
  • Our knowledge base is in German. The language model therefore writes the search queries in German here too. The coach itself translates into English what it takes from the sections found, with the same source numbers; the titles of the sources remain in German. Below the sources of an answer, there is therefore the notice “The sources are in German; Stabily translates what it uses.” A translation can be inaccurate too: look up important statements in the source cited.
  • The knowledge base contains legal information, school structures and contact points only for Austria, Germany and Switzerland. The coach is therefore instructed never to ask about the country here. If the app nevertheless recognises a clarifying question about the country, by its wording or by countries as answers you can click, it does not show it, and Stabily answers instead; if the planning has then not written a search query, Stabily searches with your own words as described above. The detection is not complete, and the app does not check questions in the text of an answer. If an answer depends on the law or the contact points of a country, the coach is supposed to say so and describe the general approach; if you name Austria, Germany or Switzerland yourself, it is supposed to use the knowledge base’s information for that country. The emergency notice names the European emergency number 112, which works in every EU country and in Switzerland, and leads to the English page with support services.
  • If our team has switched on the guided intake, it runs in English and with a built-in English question catalogue; the versions that our team maintains in the admin area only apply to chats in German. The question about the country that it can otherwise ask at the end is left out, and the language model is instructed not to ask about the country of its own accord either. If the app nevertheless recognises such a question in a text that the language model has worded for the guided intake, it does not show this text but, where needed, a fixed text instead, for example the question from the catalogue. The detection is not complete, and the app does not check the lines of the summary for it.
  • Mistral Small 4 writes the action plan in English. The app checks it according to the same rules, adapted for English texts, for example for time references and in the word list for health information. The plan does not ask about the country.
  • The coach adapts a worksheet in English, including its title. Where the original is named, its title remains in German: in the footer and the file name of the PDF and in the worksheet’s last line (“Based on the worksheet …”), which the coach is supposed to write exactly like that. The coach is also supposed to copy the original’s own credit line word for word in German; if the model leaves it out of an adapted worksheet, the app puts it back in German.
  • If our team has switched on name replacement, the app replaces names with fixed rules for English texts (see Replacing names).
  • When you dictate and in voice mode, Mistral AI converts your words into English text, and voice mode reads the answers aloud in the voice “Harriet” (see Which models Stabily uses).
  • Whatever else Stabily writes in the chat is in English too, for example the chat’s title, the fixed answer to the “What can you do?” button, the emergency notice and the PDFs that Stabily creates there, that is, the action plan and the adapted worksheet (see You are chatting with an AI). The original of a worksheet remains in German, as it was published.

How we set the coach’s instructions

Our team can adjust the coach’s guide, that is, what it helps with, its tone, how it conducts the conversation, what it can clarify and how it structures recommendations, in the admin area and test it beforehand. We save every change as a new version; earlier versions are kept, with the time and the team member who created them. This allows us to trace which instructions applied when, and to return to an earlier version. By contrast, the rules on the form of the answers, on the knowledge base and citations, on what Stabily says about its abilities, on data protection and on safety are fixed in the program, for example that the coach first names emergency numbers and support services if there are signs of acute danger. They apply with every version of the guide, the safety rules always take precedence, and they cannot be changed in the admin area.

How the knowledge base is built

Our team selects the sources. When they are imported, a model recognises the text of the PDF files, we split it into sections with page references, a model maps important terms and how they relate to each other, and for the semantic search the sections are converted into numerical vectors. Some sources exist only as text, without a PDF file; their sections have no page references. Your conversations do not go into the knowledge base.

The AI can be wrong

Language models calculate which text is most likely to fit. They do not understand the way a human does and can be wrong, even when the answer sounds convincing. The knowledge base reduces this risk but does not rule it out: the model formulates every answer itself and also brings in general knowledge. Only the statements for which Stabily names a source are supported by that source. Typical errors are:

  • statements that are false, inaccurate, outdated or entirely made up;
  • a source that says something different from the answer, or an inaccurate page reference;
  • suggestions that do not suit your class, your school, your country or the requirements of your education authority;
  • one-sided or stereotypical portrayals, for example of groups of people;
  • misunderstandings when a situation is described briefly or ambiguously.

That is why your professional judgement remains decisive: check suggestions before you put them into practice, look up important statements in the cited source and get professional help with serious problems. More on this in the Terms of use.

What happens to your input

So that Stabily can answer, we transmit the following to Mistral AI:

  • for the coach, including its planning step before the answer: your message, the conversation so far, the relevant sections from the knowledge base and the coach’s instructions;
  • for the guided intake, if our team has switched it on: for the intermediate step before each message of the process, the same as for the planning, and for the summary, also the sections found in the knowledge base;
  • for replacing names, if our team has switched it on: every new message as you wrote it, so that a model can list the names in it; apart from the safety check, which runs at the same time, Mistral AI receives everything else after that with placeholders instead of the names detected (see Replacing names);
  • for the safety check: your message, and for checking the finished answer, that answer together with your message (this includes the summary of the guided intake and the short answer there to a question you ask in between);
  • for the knowledge search: the coach’s search queries and, if its planning fails, the beginning of your message; in a chat in English also when the planning wants to ask about the country and writes no search query, and there with German search terms added (see Chats in English);
  • for the title in the sidebar: your first message and the beginning of the first answer, from which a model suggests a title;
  • for the action plan, when you open it: the texts of your messages and of Stabily’s answers, without the sections from the knowledge base, and the previous plan;
  • for a worksheet you ask Stabily to adapt: in addition to what goes to the coach, the text of that worksheet from the knowledge base;
  • for dictation, when you tap “Use recording”: only your recording, without your chat. Mistral AI sends the text back; we store neither the recording nor the text until you send the text as a message.
  • in voice mode, while it is on: the signal from your microphone, directly from your browser (Mistral AI sees your IP address in the process), and, for reading aloud, the text of the answers, sentence by sentence, without citations. We store neither the voice nor the speech output.

For the question asked with the “What can you do?” button, we transmit none of this: its answer is a fixed text, and the chat gets the fixed title “What Stabily can do”, in a chat in German “Was Stabily kann” (see “What can you do?”). Only your next message goes to Mistral AI as described above, with this question and its answer in the conversation history.

Mistral AI processes this data on our behalf; there, it is stored in the EU by default, for up to 30 days, in order to detect misuse. Your conversations are not used to train AI models: we have objected to Mistral AI using our data for training, and we ourselves do not train any models with your conversations.

Exception for dictation and voice mode (as of 11 November 2026): The models that convert your voice into text and read the answers aloud are labelled by Mistral AI as “Preview” models. For such models, Mistral AI reserves the right in its terms and conditions to use the data and outputs to train its AI models; our objection does not apply to them. This only concerns these two features and only if you use them: your recording or the signal from your microphone, the text produced from it and, in voice mode, the sentences read aloud. For everything else, what is stated here continues to apply. More on this in the Privacy policy.

Without an account, we delete your anonymous ID after 30 days of inactivity; we then keep your conversations without any link to you, and their text for no more than 24 months. With an account, we keep them until you delete them or your account. Stabily does not make decisions that produce legal effects concerning you or similarly significantly affect you (Art. 22 GDPR). Everything else, including your rights, is set out in the Privacy policy.

Please do not enter any names or other details by which pupils, parents, colleagues or other people could be recognised. Describe situations in general terms, for example “a pupil in my class” or “Person A”. If our team has switched it on, Stabily replaces detected names with placeholders (see Replacing names); this is an aid, not a guarantee. You can find details on this in the Terms of use.

Human oversight and reporting problems

You decide

Stabily does not make decisions and does not act on its own: it does not write to anyone, does not call for help and does not carry out anything for you. Whether and how you use a suggestion is your decision.

What our team does

  • We select the sources of the knowledge base and approve them before the coach uses them.
  • We set the coach’s instructions in traceable versions and choose the model with which the coach answers. We first try out new instructions and models in a separate test chat that stores nothing.
  • We switch the guided intake on or off and maintain its question catalogue for chats in German in versions; earlier versions are kept, and we record every change. In the test chat, we try out the process without storing anything.
  • We switch the replacement of names on or off; we record every change.
  • Using our own tools, we check which sections the knowledge search finds for a question and why.
  • We evaluate the feedback on the answers, look at every report on an answer and mark what we have seen and dealt with.
  • For every answer, we see technical details without content, for example how often the knowledge base was searched and how long the answer took, in order to find errors and delays.
  • If necessary, we withdraw a source, change the coach’s instructions or return to an earlier version, or switch the model.
  • If necessary, we pause Stabily with a switch, immediately and for everyone. The chat then does not accept any new messages and shows you a notice pointing the way to Emergency help, and action plans are neither created nor updated. You can still read and delete your previous conversations, and still download a finished action plan.

Members of our team may read conversations in order to improve Stabily’s answers and to ensure quality: without an account, all of them; with an account, depending on your setting, by default also all of them. Every time our team opens a conversation, this is logged. Our team never writes in the chat itself and does not reply to you there. Who may read conversations and what you can set with an account is set out in the Privacy policy.

Reporting problems

If an answer seems wrong, inappropriate, dangerous or discriminatory to you, or if you have a question about our use of AI, write to us at wirsind@stabilimkopf.com. It helps to include the approximate date, what it was about and what was wrong with the answer, ideally with the exact wording. You can export a conversation for this from the chat’s menu. Before you do, please remove names and other details by which people can be recognised. We look at every report and reply to you.

Below every answer, you can also indicate whether it helped you and, if you choose “Not helpful”, add reasons and a short text. Our team evaluates this feedback to improve Stabily but does not reply to it. What we store in the process is set out in the Privacy policy. For urgent problems, please send us an email.

Complaints to an authority

If you believe that Stabily infringes the AI Act, you can lodge a complaint with the competent market surveillance authority (Art. 85 AI Act). [to be added: the competent Austrian market surveillance authority, once Austria has designated it] General information on the AI Act is available from the AI service centre (KI-Servicestelle) of the Austrian regulator RTR. How you can lodge a complaint with the Austrian Data Protection Authority (Datenschutzbehörde) is set out in the Privacy policy.

AI literacy

In our team

The AI Act requires providers and deployers to support the AI literacy of the people who work with AI systems on their behalf (Art. 4). For us, this concerns everyone who works on Stabily, for example maintaining the knowledge base, setting the coach’s rules or handling reports. They should understand how Stabily works, where language models typically make mistakes and how to check answers and sources. [to be added: specific measures, for example internal guidance, training sessions and how often they take place]

For you

A few tips for using Stabily:

  • Read answers critically, especially figures, legal statements and quotations, and look up anything important in the cited source.
  • Stabily does not know your class, your school or the requirements of your education authority. Adapt suggestions to your situation.
  • For legal questions, for example about reporting and notification obligations, and for serious problems, turn to professionals, the school psychology service or your school management, and in case of acute danger, to the emergency services.
  • Do not enter any names or other details by which people could be recognised.
  • If you pass on or publish answers, check them first and, where necessary, label them as created with AI.

Which models Stabily uses

Stabily exclusively uses models from Mistral AI through its interface (API); there, the data is stored in the EU by default. We do not use any other AI providers.

  • Mistral Small 4 (identifier mistral-small-2603): by default writes the coach’s answers, including the planning step before them and adapted worksheets, and suggests titles for your conversations. It also writes the planning step when Mistral Large 4 answers. If our team has switched on the guided intake, it also writes its two intermediate steps, whichever model the coach uses. If our team has switched on the replacement of names, it also lists the names in every new message. At your request, it also summarises a chat in an action plan, whichever model the coach uses.
  • Mistral Medium 3.5 (identifier mistral-medium-2604) and Mistral Large 4 (identifier mistral-large-4-0): our team can set one of these models instead of Mistral Small 4 for the coach’s answers, for all conversations or only for individual accounts, for example for testing. Mistral Medium 3.5 then also writes the planning step; with Mistral Large 4, Mistral Small 4 continues to write it. Without an account, the answers always come from the model that is set for everyone. Mistral Large 4 also identifies terms and how they relate to each other in the sources when the knowledge base is built.
  • Mistral Moderation 2 (identifier mistral-moderation-2603): automatically checks each of your messages for signs of danger and misuse, and the coach’s finished answers for signs of danger (not if there has already been such a sign in the chat, and not for clarifying questions from the planning step or the card and questions of the guided intake; the summary of the guided intake and the short answer there to a question you ask in between are checked).
  • Mistral Embed (identifier mistral-embed-2312): converts search queries and sections of the knowledge base into numerical vectors so that the search finds passages that match in content.
  • Mistral OCR 4.1 (identifier mistral-ocr-4-1): reads in the text of the PDF files that make up the knowledge base. It only processes these sources, not conversations.
  • Voxtral Small Transcribe 3 (identifier voxtral-small-transcribe-3-0): converts a dictated message into text in the page’s language (German or English). It does not answer and does not assess anything; it only sees the recording, not your chat.
  • Voxtral Mini Transcribe Realtime 3 (identifier voxtral-mini-transcribe-realtime-3-0): listens in voice mode, continuously converts your words into text and detects when you pause. It does not answer and does not see your chat.
  • Voxtral Mini TTS 3 (identifier voxtral-mini-tts-3-0): reads the answers aloud in voice mode, in Mistral AI’s voice “Heidi” in German chats and “Harriet” in English chats. It only receives the text it is to read aloud.

The three Voxtral models are labelled by Mistral AI as “Preview” models (as of 11 November 2026); what this means for training is set out under What happens to your input.

We do not use any models for the coach other than those listed here. If we change the default model or add a new model, we record this in the change log.

The AI Act: status as of November 2026

The AI Act applies in stages. With the “Digital Omnibus on AI” (Regulation (EU) 2026/1744), the EU has postponed some deadlines and adjusted obligations. The current status:

  • 1 August 2024: The AI Act enters into force.
  • 2 February 2025: The prohibitions of certain AI practices (Art. 5) and the AI literacy obligation (Art. 4) apply.
  • 2 August 2025: The obligations for providers of general-purpose AI models apply, which for Stabily means Mistral AI, as do the rules on authorities and penalties.
  • 27 July 2026: The Digital Omnibus on AI enters into force. It postpones the deadlines for high-risk AI systems, adds new prohibitions and redefines the AI literacy obligation: providers and deployers must take measures that support AI literacy, but do not have to ensure a particular level.
  • 2 August 2026: The AI Act applies in large part, including the transparency obligations (Art. 50) and the right to lodge a complaint with the market surveillance authority (Art. 85). The European Commission has published guidelines on this and a voluntary code of practice on labelling AI-generated content.
  • 2 December 2026: New prohibitions of AI systems that generate intimate or sexually explicit depictions of identifiable persons without consent, or depictions of child sexual abuse. In addition, the transition period for machine-readable marking ends for systems that were already offered before 2 August 2026.
  • 2 December 2027: The obligations for high-risk AI systems under Annex III apply, including the area of education.
  • 2 August 2028: The obligations for high-risk AI systems in products under Annex I apply, for example in machinery, toys or medical devices.

The “Digital Omnibus” on data, which would amend the General Data Protection Regulation (GDPR), among other things, has not yet been adopted. As of November 2026, it is still being discussed in the European Parliament. Until then, the GDPR applies unchanged.

In Austria, the authorities that supervise the AI Act have not yet been designated; a national implementing act is being prepared.

Changes and change log

We update this page when we make significant changes to Stabily, for example using a different model or offering new features, and when the legal situation changes or authorities publish new guidelines. The current version is shown at the top of this page with its “Last updated” date.

  • 15 November 2026: We now describe how Stabily works in chats in English (see Chats in English): a chat keeps its language. In a chat in English, the coach answers in English, searches the German knowledge base with German search queries, translates what it takes from it and is instructed never to ask about the country. The guided intake uses a built-in English question catalogue there; the action plan and the PDFs that Stabily creates are in English, and the fixed rules for names and for health information know English terms.
  • 14 November 2026: Stabily is also available in English. In an English chat, Voxtral Small Transcribe 3 converts a dictated message into English text, and voice mode listens in English and reads the answers aloud in Mistral AI’s voice “Harriet”. The models and the provider stay the same.
  • 11 November 2026: Our team can switch on a mode in which Stabily replaces detected names in your messages with placeholders such as “[Person A]” before storing them. For this, Mistral Small 4 lists the names in every new message; below your message, you see what was replaced.
  • 11 November 2026: The models for dictation and voice mode are labelled by Mistral AI as previews, and Mistral AI may use their data for training. We now say so under “What happens to your input”; for everything else, it remains the case that your conversations are not used for training.
  • 11 November 2026: New: voice mode. You speak with Stabily, and Stabily reads its answers aloud. The models Voxtral Mini Transcribe Realtime 3 and Voxtral Mini TTS 3 from Mistral AI convert your words into text and generate the voice.
  • 11 November 2026: You can now also dictate your message. The model Voxtral Small Transcribe 3 from Mistral AI converts your recording into text; you read it in the input field and send it yourself.
  • 11 November 2026: If you ask a question yourself in the middle of the guided intake instead of answering, Stabily now answers it briefly and then asks its question again.
  • 11 November 2026: Our team can switch on the guided intake: at the start of a case you describe, Stabily then first asks for key facts and asks two to five questions for reflection, sums up and only then gives the recommendation; you can skip the questions. For this, Mistral Small 4 writes two intermediate steps instead of the planning; the questions come from a catalogue that our team maintains. The safety check checks the summary like an answer.
  • 9 November 2026: On the empty start page, everyone without an account has the “What can you do?” button. The answer to it is a fixed text by our team, not an answer from the AI. If you ask the coach yourself what it can do, it answers directly and names the same abilities.
  • 9 November 2026: In the action plan, “Bullying” now only appears if you have confirmed the imbalance of power and the criteria for bullying yourself; if the model reads them from your words, this counts as an assumption. In addition, the app removes terms for diagnoses and illnesses from the plan using a fixed word list.
  • 9 November 2026: The safety check now also checks the coach’s finished answer, as long as there has not yet been a sign of danger in the chat, and, if there is a sign of acute danger, shows the emergency notice below the answer. Once there has been such a sign in a chat, the coach keeps safety in mind in all further answers, and a link to Emergency help remains at the top of the chat. If the planning step fails, Stabily still searches the knowledge base, with the beginning of your message.
  • 9 November 2026: Our team can now pause Stabily immediately with a switch; the chat then does not accept any new messages and shows a notice pointing the way to Emergency help, and action plans are not created or updated during the pause. An adapted worksheet names as its basis the original from the Stabily knowledge base and the original’s own citation, and no longer simply Stabil im Kopf.
  • 8 November 2026: In the chat, you can view worksheets that Stabily cites as a source, download them and have them adapted with Stabily. For this answer, the coach receives the text of the worksheet; the adapted worksheet is available as a PDF, labelled as adapted with AI on every page and in its metadata.
  • 8 November 2026: After a detailed recommendation, you can have a chat summarised in an action plan and download it as a PDF. The draft is written by Mistral Small 4, which backs up every statement with a passage from the chat; the app checks that the passage is in the chat and uses this to determine the assessment according to fixed rules. The PDF is labelled as created with AI on every page and in its metadata.
  • 7 November 2026: Members of our team may read conversations in order to improve Stabily’s answers. We keep conversations without an account even after the anonymous ID has been deleted. By default, the coach answers with Mistral Small 4 instead of Mistral Large 4; our team can set Mistral Medium 3.5 or Mistral Large 4 for everyone or for individual accounts. Before every answer, the coach plans in a short intermediate step; if the answer needs specialist knowledge, Stabily searches the knowledge base before the coach answers. We store the coach’s instructions in versions that our team can adjust; the safety rules remain fixed. Conversations without an account that we keep lose their text after no more than 24 months. You can give feedback below every answer.
  • 6 November 2026: First version. Stabily generates answers with models from Mistral AI and draws on its own knowledge base of selected specialist literature, stating the title and page.