 |
| The Permissions setting under Plugins in ChatGPT decides when ChatGPT asks before it uses connected apps, in voice mode as well. |
OpenAI rolled out a significant update to ChatGPT's voice mode on Wednesday, allowing the AI assistant to pull information directly from connected third-party services like Gmail, calendars, and Slack while users converse with it. The feature, which works across web, iOS, and Android platforms, displays written responses in the chat interface as the conversation unfolds.
The update means users can now ask the voice assistant questions about their email inbox, upcoming calendar events, or specific Slack threads without switching between apps. Free and Go tier users will also have access to the feature, though limited to the plugins their respective plans support. Additionally, Voice mode is now available within ChatGPT Work—the dedicated workspace for documents, presentations, and spreadsheets—though users will need both Voice and Work access enabled to utilize it there.
No New Permissions Granted
Importantly, the update does not create any new access rights. According to OpenAI, "Existing app connections, permissions, and usage limits apply." This means if a user has already connected their Gmail account with sending permissions—a capability ChatGPT has supported for two Google accounts simultaneously since August—voice mode will operate under those same existing rights.
Screen Approval Required for Actions
The most significant limitation of the new feature is spelled out clearly in OpenAI's Voice FAQ. The documentation states: "If an action needs your approval, Voice will prompt you to review it on screen. Use the on-screen controls to approve or decline; spoken approval is not supported."
This means telling ChatGPT "yes, send it" after it reads out a draft email will not actually send the message. Whether you're driving in your car or have your phone tucked away in a pocket, any task requiring approval will wait until you physically look at and interact with your screen.
For those looking to stay current with ChatGPT's evolving feature set, the official release notes provide comprehensive updates on new capabilities and changes.
Permission Settings Determine When ChatGPT Asks
Whether ChatGPT prompts for approval at all depends on a setting located under Settings > Plugins > Permissions. Three primary options exist:
- "Always ask" – ChatGPT requests permission before reading app data or making any changes
- "Allow read actions" – ChatGPT reads data without asking but still prompts before making changes
- "Allow low-risk actions" – Lower-risk items are approved automatically, while other actions still require confirmation
In a personal German account tested by our team—which had never connected a plugin—the third option was selected by default. OpenAI classifies sending emails, messages, comments, or invitations as higher-risk activities, along with creating, editing, or deleting files, modifying security settings, and conducting financial transactions.
The Voice Mode FAQ also references an "Allow all actions" setting that skips approval prompts entirely, though this option did not appear in the personal account we examined.
Understanding Read vs. Write Access
For voice mode users, read access proves more consequential than sending capabilities. Under the "Allow read actions" setting, asking "What's in my inbox today?" will yield an answer from Gmail without any prompt, and the low-risk option typically behaves identically. Those wanting a prompt every single time should switch to "Always ask."
OpenAI cautions users that "Changing an app permission does not disconnect the app or revoke access already granted." To fully cut access, users must navigate to Settings > Plugins, select the specific app, and disconnect the account there. The managing app permissions guide offers detailed instructions for this process.
Privacy Considerations and Data Retention
The conversations themselves leave records. OpenAI retains audio clips from Live conversations for 30 days, deleting them alongside the chat unless the company requires them for security or legal purposes. The new privacy center does not alter these default retention policies.
For additional perspectives on this update, 9to5Mac's coverage highlights three key ways OpenAI enhanced the voice experience, while TechCrunch's report examines how the mobile app is gaining voice-based agentic capabilities.
The Bottom Line
OpenAI's integration of connected apps into ChatGPT's voice mode represents a meaningful step toward a more capable AI assistant. However, the requirement for on-screen approval before any significant action keeps users firmly in control—at least for now. The system is designed to prevent accidental sends or unauthorized actions, though it also means true hands-free operation remains limited when approvals are needed. For users who value convenience and are comfortable with read-only access, the feature offers genuine utility. For those concerned about privacy or accidental actions, the granular permission settings provide necessary safeguards.