Three announcements from 23 September share one thing: they are about your voice and what an AI service does with it. OpenAI links ChatGPT Voice to your email and calendar, Meta gives glasses wearers an opt-out from AI training, and Alibaba cuts the prices of its speech models sharply. For anyone working with client or patient data, the question shifts from "what do I type in a chat window" to "what does the assistant get access to".
ChatGPT Voice gains access to email, calendar and Slack
According to 9to5Mac, ChatGPT's voice mode from 23 September works with plugins, including email, calendar and Slack. You can ask aloud whether there is a free half hour in your calendar and have an appointment placed there, or have it search for a particular email (Digital Trends). Voice now also runs on all three GPT-6 models (Astra, Sol and Luna) and works alongside ChatGPT Work.
Handy, but mind what it means. A spoken request can now pull data from your mailbox and calendar, and calendars hold names, appointments and sometimes diagnoses or case numbers. If you work in healthcare, at a law firm or in HR, first check which plugins are switched on in your working environment and whether they fall under your data processing agreement. The reports say nothing about new privacy settings or EU availability, so do not assume anything changes to the existing GDPR profile of ChatGPT.
Meta glasses: opt-out for images, not for audio
Meta announced that users of its AI glasses can refuse to have their "visual data" used to train AI. According to Engadget, after an opt-out those images are also no longer shown to external contractors who label them. Meta itself: "If you've opted out and your glasses are asked to translate a menu, we use your images to answer your question and then they're gone." (loosely translated)
Two caveats. The opt-out applies only to images: spoken interactions with Meta AI may still be used for training and review, although you can delete those recordings manually. And an opt-out changes nothing about the point the Dutch Data Protection Authority made this week: filming others with camera glasses is almost never allowed without good reason. You can read that in our post of 22 September. A better training setting at Meta does not settle the rights of the people you film.
Qwen Audio 3.1: transcription up to 95% cheaper
On 23 September, Alibaba released Qwen-Audio-3.1: five models for speech recognition, text-to-speech and realtime conversations. According to The Decoder, the price for text-to-speech falls by about 70%, for realtime by about 85% and for speech recognition by up to 95%. The ASR-Next variant recognises multiple speakers with timestamps and detects emotion and background noise.
For conversation reports in customer service or the consulting room, that sounds attractive. But the source does not state which languages are supported, nor where the data is processed. Earlier this month we already saw a Qwen model that was available only via an API outside the EEA. A low price per minute of audio does not outweigh a conversation containing special category personal data ending up on a server outside Europe.
What you should do now
- Check which ChatGPT plugins are enabled in your organisation now that they are also reachable by voice.
- If you use Meta glasses for work, switch on the opt-out and delete voice recordings. Set out in your house rules where the glasses may not be worn.
- Test cheap speech models only with non-sensitive audio until the processing location is clear.
You can find the current privacy profile per provider on the model comparison and the safety page.