Three different kinds of data
It helps to separate what is being collected, because the three types carry different risks and are often governed by different settings.
Audio is the most sensitive, since a voice recording identifies you in a way a vocabulary score does not. Transcripts sit in the middle: they contain whatever you said, which may include far more than you intended. Progress data is usually the least sensitive and the most useful to keep.
| Data type | Why it exists | Question to ask |
|---|---|---|
| Audio | Speech has to be heard | How long is it kept? |
| Transcript | Corrections and reports | Can I delete individual sessions? |
| Progress data | Adapting difficulty | Is it exportable? |
| Account details | Billing and login | Standard account questions |
| Usage analytics | Improving the product | Can I opt out? |
A service that answers all five clearly is being straightforward. One that answers none is telling you something too.
The questions worth asking
You do not need to read a full privacy policy to get the important answers. Four questions cover most of it, and the answers are usually findable in a few minutes.
The training question is the one people care about most once it is pointed out: whether your recordings are used to improve models, and whether that can be switched off.
- Are recordings used to train models, and can I opt out?
- How long is audio retained, and is deletion available?
- Is data shared with third parties, and which ones?
- Where is data stored, and under which jurisdiction?
- Can I export or delete everything if I stop using the service?
Habits that make it moot
The most effective protection is not a setting; it is not putting sensitive information into a practice session in the first place.
This costs nothing pedagogically. Rehearsing a hospital appointment works identically with an invented condition, and a job interview roleplay does not need your real employer's name to be useful.
Building the habit early is easier than remembering to be careful in the moment, especially when a scenario gets absorbing.
| Japanese | Romaji | Meaning |
|---|---|---|
| 個人情報 | kojin jōhō | Personal information |
| 機密情報 | kimitsu jōhō | Confidential information |
| 仮の名前 | kari no namae | A placeholder name |
| 架空の会社 | kakū no kaisha | A fictional company |
| 設定を確認します。 | Settei o kakunin shimasu. | I will check the settings. |
| 録音を削除したいです。 | Rokuon o sakujo shitai desu. | I want to delete the recording. |
Practising with placeholder details also removes a small hesitation that slows learners down: the worry about what they are saying out loud.
At work
Using a language tool on a company device raises a separate question, because the rules are set by your employer rather than by the app's privacy policy.
Many organisations restrict services that process audio on managed laptops, and some prohibit discussing internal matters with any external service regardless of its terms. Rehearsing a real client meeting with real details can breach that even when the tool is entirely trustworthy.
The safe version is straightforward: practise the situation, invent the specifics.
| Scenario | Risk | Safer version |
|---|---|---|
| Rehearsing a client call | Client name and details | Invented client |
| Practising a status update | Internal project data | Generic project |
| Interview practice | Current employer details | Describe the role generally |
| Medical vocabulary | Your own condition | A textbook example |
| Reading a real email aloud | Whatever it contains | Retype a neutral version |
What is reasonable to accept
Some data collection is inherent to the service working. Speech must be processed to be corrected, and progress must be stored for difficulty to adapt, so demanding that nothing be retained means demanding a tool that cannot do its job.
The reasonable position is proportionality: collection that serves the feature you are using, with retention you can see and control, and no secondary use you did not agree to.
Treating this as a spectrum rather than a binary makes it easier to choose sensibly rather than either ignoring it entirely or refusing every tool.
- Expected: audio processed to give you feedback.
- Expected: transcripts kept so you can review sessions.
- Reasonable to control: training use, retention period.
- Worth avoiding: unclear third-party sharing.
- Worth avoiding: no deletion path at all.
A short checklist
Ten minutes at the start saves revisiting the question later with a year of sessions already stored.
Do it once, per service, and then stop thinking about it.
- Read the retention and training sections of the policy.
- Turn off any training use you do not want.
- Check your employer's rules before using a work device.
- Decide your placeholder name and company now.
- Note where the deletion and export controls are.
None of this should stop you practising. It should take one session's worth of time and then never interrupt you again.