What the AI does, and where it stops.
EaseTalk uses AI to hold up a conversation you can practise against. This page names the providers behind it, sets out what it cannot do, and is honest about the fact that speech recognition mishears the people this app was built for more often than it mishears anyone else.
In short
The characters you talk to in EaseTalk are generated by AI. It exists to give you something realistic to practise against, to prompt reflection, and to pull your practice into a summary a therapist can review.
It does not diagnose, assess, or decide anything about your care. It is not a medical device and it is not a crisis service.
AI output can be wrong, and speech recognition mishears atypical speech more often than typical speech, which means the people this app is built for are the people most likely to be misheard. Every practice shows you What we heard so you can see when that happens.
Your therapist can accept, edit or ignore anything the AI produces. Clinical judgement is theirs.
1What the AI is for
EaseTalk uses AI for three jobs, and no others.
Structured practice
AI plays the other side of a conversation, so you can rehearse a real situation without needing another person to be free at that moment.
Reflection prompts
After a practice, AI helps generate prompts that ask you how it went. What you write in reply is yours, in your own words, and it is not rewritten.
Summaries for clinician review
AI helps turn what you did into a summary your therapist can read quickly, if you have agreed to share with them. The summary is a convenience for them. It is not an assessment of you.
2What the AI is not
Clear limits, stated plainly
The AI in EaseTalk does not diagnose. It does not assess your speech or your condition. It does not recommend or change treatment. It takes no decision about your care, on its own or otherwise. EaseTalk is not a medical device and it is not a crisis or emergency service.
If you are in distress or need urgent help, contact your treating clinician or your GP, or call the emergency services on 112 or 999. Nobody at EaseTalk monitors the app for emergencies, and it would be wrong to let you believe otherwise.
Nothing the AI produces should be used as the basis for a clinical decision on its own, by you or by a clinician.
3Which AI providers we use
The previous version of this page named none of them. Here they are, with where they run.
| Provider | What it does in EaseTalk |
|---|---|
| Microsoft Azure AI Foundry | Runs the language models behind the conversations, summaries and reflection prompts. Hosted in the EU, in North Europe (Dublin) or Sweden Central |
| Microsoft Azure Speech | Speech recognition during a live practice, in the same EU regions |
| ElevenLabs | Turns a character's words into speech. It receives that text, not your audio |
| Your own device | Some parts of the app use the speech recognition and speech synthesis built into your phone, so that part does not leave it |
What is sent to these providers during a session, and what our commitments are, is set out in section 7 of the privacy policy. It includes the words the recogniser detected, the conversation so far, the scenario, the goals for that practice, and relevant account details including your name.
If this list changes, we update this page and the privacy policy before the new provider starts processing your data.
4Limitations you should know about
These are the real ones, not a disclaimer written to be skipped.
The output is probabilistic, so it can be wrong
Language models produce plausible text, not verified text. A character can say something inaccurate, out of character, or unhelpful. A summary can misdescribe what happened. Read it as a draft, never as a record.
Speech recognition mishears, and it mishears some people more
This affects the people EaseTalk is built for
Speech recognition systems are trained mostly on typical speech. They are measurably worse at atypical speech, which includes dysarthria, stuttering and aphasia. That means the users most likely to be misheard by EaseTalk are precisely the users it exists to support. We are not going to bury that at the bottom of a page.
This is why every practice shows you What we heard: the words the recogniser detected are put on screen, so a mishearing is visible rather than silent. If what you see is not what you said, the recogniser was wrong, not you. Use Sound Setup to give it a better chance, and tell your therapist when it happens, because that is useful information for both of you.
It does not know you
The AI has the scenario, the conversation so far, the goals for that practice and some account details. It does not have your clinical history, and it is not reasoning about your condition. It is holding up a conversation.
It can be interrupted
Live practice depends on a network connection. If it drops, the practice may end. You can end a practice yourself at any point, and nothing is scored against you for doing so.
5Human oversight
A person is responsible for care, and the software is not that person.
- Your speech and language therapist may accept, edit or disregard any AI output, including summaries and trend descriptions.
- AI supports the workflow. It never determines care, and no output changes anything about your treatment on its own.
- Nothing in EaseTalk produces an automated decision with legal or similarly significant effects on you within the meaning of Article 22 of the GDPR.
- If a summary is wrong, the therapist's own record of the session is the one that counts.
Our AI ethics page sets this out in more detail, including what we will not do with a recogniser's mistakes.
6Training on your data
We do not use your identifiable data to train AI models, and the AI service is configured so that your content is not used to train the underlying models either. The models run inside our own Azure AI Foundry deployment in the North Europe region rather than on a shared consumer service.
If that position ever changes, we will say so here and in the privacy policy before it takes effect, and where consent is required we will ask for it rather than assume it.
Where your data lives, and which company processes what, is set out in section 9 of the privacy policy.
7Where this sits under the EU AI Act
We consider EaseTalk to be a low risk assistive system: it supports a person practising communication and it supports a clinician's own workflow, it does not make or recommend decisions about care, and it is not a medical device.
We want to be careful about how strongly we put that. This is our own informal reading of where the product sits. It is not a formal conformity assessment, and it has not been reviewed by a notified body or a regulator. We state it so you know our position and can challenge it, not so it can be quoted as a classification.
Where the AI Act's transparency duties apply, our approach is to exceed rather than meet them: you are told you are talking to an AI character, you are shown what the recogniser heard, and the providers behind the features are named on this page.
8Acceptable use of the AI features
When you use the AI features, please do not:
- try to extract, copy or reverse engineer the models, the prompts, or the scenario definitions behind them;
- attempt to make the system behave outside its purpose, including trying to get it to give medical, legal or financial advice;
- use any output as the basis for a clinical decision, whether you are a clinician or not;
- put information into a practice that you would not want stored, including other people's personal or health information;
- use the service to produce content that is unlawful, abusive, or intended to harm someone.
We may suspend access to the AI features if they are being used in a way that risks harm to you, to someone else, or to the service. Our terms of service cover this more generally.
9How this fits with our other terms
These AI terms sit alongside our terms of service and our privacy policy.
- The terms of service govern your use of EaseTalk generally.
- These AI terms add specific terms for the AI features, and where they say something more specific about the AI, they apply to that subject.
- The privacy policy governs personal data. Where anything here appears to conflict with it on a data protection question, the privacy policy applies.
10Contact
Questions about the AI features, or something the AI produced that you think is wrong, can go to easetalkai@gmail.com, or mfaggiani@ucc.ie. If an AI output about your own practice looks wrong to you, please tell your therapist as well, because they are the person who can act on it.