AI transparency

Last updated: 30 September 2026

This notice explains where AbilityMic uses artificial intelligence, what that AI can and cannot do, what leaves your device, and how to turn it off. It is our transparency information about AbilityMic's AI features, prepared with the EU Artificial Intelligence Act (Regulation (EU) 2024/1689) in mind, and our plain-language explanation for everyone else.

Looking for a plain-language overview? See AI Act and your communication — nine principles for AI in AAC.

1. In short

AbilityMic is an AAC app first. Symbols, text and speech output work offline, with no AI involved at all.
Online AI and features that use your data are off until you turn them on. That covers cloud AI, symbol generation, the assistant, My memory, partner speech as context, suggestion explanations and cloud voices. Each one is switched on by you or your caregiver and can be switched off again at any time.
On-device suggestions work from the first launch. Word suggestions and sentence proposals (grammar rules, plus Apple Intelligence or Gemini Nano on compatible phones and tablets) run on the device and send nothing. A sentence proposal never replaces your message unless you choose it. The Simple profile hides sentence proposals; the app does not currently have a separate switch for them.
You are talking to a machine when you use it. Sentence suggestions, generated symbols and generated boards are produced by an AI model. They can be wrong, odd or incomplete.
You decide what is said. A suggestion is never spoken on its own — a person accepts it, edits it or discards it first.
We do not analyse you. AbilityMic does not recognise emotions, does not identify people by voice or face, does not score users, and does not diagnose or assess anyone.
Nothing here is medical. AbilityMic is not a medical device and gives no clinical advice.

2. Where AI is used

Every row below is a separate feature. Online features and features that use your data have their own switch; on-device word suggestions and sentence proposals are on by default, and the Simple profile hides sentence proposals. “On device” means the content stays on your phone, tablet or computer. “Online” means the content is sent over an encrypted connection to the AbilityMic backend, which passes it to the AI provider and returns the result.

Feature	Where it runs	What leaves the device	Default
Sentence shaping — turns a few selected words into a full sentence	On device: grammar rules and, where the hardware supports it, Apple Intelligence on compatible Apple devices or Gemini Nano on supported Android devices; an optional local model on desktop. Online only if you enable cloud AI	Nothing on device; the current sentence draft when cloud AI is on	On device: on (hidden by the Simple profile; the desktop local model is off). Online: off
Word and symbol suggestions while building a message	On device	Nothing	On (on-device only)
“Explain this suggestion” — why a suggestion appeared	On device	Nothing	Off
My Memory (second brain) — a personal knowledge base of people, places, routines and preferences that feeds the suggestions	On device, encrypted; a caregiver sees only what you share with them	Nothing, unless you share an entry with a caregiver	Off
AI symbol generation — a new symbol from a description	Online	Your text description, not your photos	Off
AI board drafts — a starting board from a description	Online	Your text description	Off
Assistant tasks in the app	Online	The request you typed or selected	Off
Speech recognition (dictation)	Your operating system’s speech service; the AbilityMic backend only if you enable cloud speech recognition	Audio is processed as it is spoken and is not stored by us	Off
Speech output (text to speech)	Device voices by default; premium cloud voices are optional	The text to be spoken, when a cloud voice is selected	Device voices
Image to AAC — a board from a photo	On device	A text prompt derived on your device, never the raw photo, and only when online symbol generation is on	Off
Partner speech as context (what the person opposite said)	On device, kept in memory only and never written to disk	Nothing	Off

Which AI providers receive online content, in which country they operate, and on what legal basis, is listed in the Privacy Policy (sections 3, 5 and 12).

3. This output is generated by a machine

When an AI feature is on:

Sentences, symbols and boards are AI-generated content. A proposal is only ever a proposal: it replaces your message only when you choose it. On iPhone, iPad and Android it is a highlighted chip with an icon above the message bar, without an “AI” label — you tap it to use it, or keep building your message if you do not want it (on Android you can also dismiss it). On desktop the app shows where the proposal came from (grammar rules, on-device AI or online AI): on Linux next to “Use suggestion” and “Keep my sentence” buttons, on Windows behind the ⓘ icon.
Speech output is synthetic speech. AbilityMic speaks with a synthesised voice — a device voice or, if selected, a cloud voice. It is not a recording of a human speaker and is not represented as one.
The AI does not speak for you. It proposes wording; a person decides what is spoken, sent or saved.
We do not currently embed machine-readable provenance markers in generated symbols. Technical marking of AI-generated content is still being assessed (C2PA metadata is one example of the technologies considered), and we state this here rather than implying otherwise.

4. What AbilityMic never does

These are deliberate product limits, not features waiting to be built:

No emotion recognition. We do not infer feelings, mood, stress or pain from voice, face or typing.
No biometric identification or categorisation. No voiceprints, no face recognition, no inference of characteristics from biometric data.
No diagnosis or clinical assessment. The app does not evaluate cognitive or language ability, does not screen for conditions and does not recommend therapy.
No decisions about a person. Nothing in AbilityMic decides on access to services, benefits, education or care, and no output should be used that way.
No social scoring, no manipulation. We do not rank users and do not use techniques designed to steer behaviour against the user’s interest.
No advertising profiling. No ad SDKs, no tracking pixels, no sale of data.
No silent switch-on of online features. A feature that sends content off the device or uses your data never becomes active because of an update; it needs a person to enable it. On-device suggestions are described in section 1.

5. Limitations you should know about

AI suggestions can be factually wrong, grammatically odd, or miss the intended meaning — always read a suggestion before speaking it.
Quality varies by language. Sentence support is strongest in English, Polish and Spanish and weaker in less common languages.
On-device sentence support depends on the hardware, the operating system version and the language; where it is unavailable, the feature simply stays off unless cloud AI is enabled.
Online features need connectivity. If the connection fails, AbilityMic falls back to offline communication rather than blocking it.
Emergency SOS does not use AI. It sends a fixed message to the contacts you configured, so that it cannot be affected by a model’s behaviour.
For a child or a user who cannot judge a suggestion independently, the parent, guardian or AAC professional decides which AI features are appropriate and remains responsible for supervising them.

6. How to control it

Privacy centre in the app — one screen listing every consent: cloud AI, online AI services, AI symbol generation, telemetry, medical data, parental consent. Each can be withdrawn separately.
Turning AI off leaves a complete AAC app — boards, symbols, speech output, history and smart-home tiles keep working.
My Memory has controls of its own — pause learning, forget everything, export, an entry-by-entry “why do I know this?”, the list of people with access, and a change log that records refused access as well as allowed.
Export your data — Privacy centre → “Export my data”.
Delete your data — “Delete all data” in the app, or delete the account and the server-side data at /account-delete.php.
Before online AI is switched on, the app shows what will be sent, what will not be sent, and that the setting can be reversed at any time.

Step-by-step instructions are in the user manual, chapter “AI in AbilityMic — transparency and control”.

7. How we classify AbilityMic under the AI Act

We assess each feature separately rather than the app as a whole:

AI transparency and user control — sentence shaping, suggestions, symbol and board generation, assistant tasks, speech recognition and synthetic speech. This page, together with the in-app notices, explains these features and the controls you have over them.
Not high-risk. AbilityMic is not used to decide on access to education, employment, benefits, credit or essential services, and is not a safety component of a regulated product.
Not a prohibited practice. See section 4 — the practices banned by Article 5 are explicitly out of scope for this product.
Not a medical device. AbilityMic supports communication; it does not diagnose, treat or monitor a condition. Its intended purpose is to help a person express what they already want to say.

AbilityMic is designed to support the transparency requirements applicable to its AI features under the EU AI Act. The exact requirements depend on the specific function, the type of generated content and AbilityMic's role in providing that function. We continue to assess technical marking requirements for AI-generated content, including generated symbols.

Our role: Kierepka.pl is the provider of the AbilityMic AI features, including where they are built on models supplied by third parties, and is also the deployer of the backend that proxies them. A full internal analysis, including the areas we are still working on, is maintained in our compliance documentation and is available to institutional customers on request.

8. AI literacy — what these words mean

AI model — software trained on large amounts of text or images that predicts a likely continuation. It does not understand your situation and has no intention.
On-device (edge) AI — the model runs on your own hardware; content does not leave it.
Cloud AI — the content is sent to a server, processed there, and the answer comes back.
Suggestion — a proposal shown to a person, which only takes effect when that person accepts it.
Hallucination — a confident but wrong output. It is a normal property of these models, which is why review matters.
Synthetic speech — a voice produced by software rather than recorded from a person.

This page, the manual chapter written for users and caregivers, and the in-app explanation of why a suggestion appeared are there to help you understand and use AbilityMic's AI features. They are educational material for users.

9. Questions, problems and complaints

AI questions and reports of a bad or harmful output: support@abilitymic.com. Please describe what was shown and what you expected; we aim to respond within 5 business days.
Privacy and data protection: privacy@abilitymic.com.
Accessibility barriers: see the accessibility statement.
Escalation: you may also contact the authority responsible for AI market surveillance in your country, and, for personal-data matters, your national data protection authority (in Poland: UODO).

We treat reports of harmful, discriminatory or unsafe AI output as defects: they are logged, investigated, and — where a feature cannot be made safe — the feature is disabled rather than shipped.

10. Changes to this notice

We review this page whenever an AI feature is added, changed or removed, and at least once a year. The date at the top reflects the most recent revision. Material changes are also announced in the app.