OpenAI on Thursday afternoon released a new feature that lets its ChatGPT desktop application for Mac read, compose, and send messages through Apple’s iMessage service, giving the AI assistant direct control over one of the most personal functions on the machine. The feature also lets ChatGPT pull information from the conversation history and generate summaries of threads, according to the company.
The integration makes ChatGPT the first major AI assistant to operate Apple’s messaging app from the desktop. A user can ask the assistant to draft a reply, schedule a message, or summarize a long family thread, and the assistant carries out the task using the same interface a human would. OpenAI described the feature as part of its broader effort to make ChatGPT a capable agent on the Mac, able to act on the user’s behalf across the applications they already use.
The new capability builds on a series of integrations OpenAI has added to the Mac app over the past year. ChatGPT already works with Xcode, Apple’s development tool, and can operate Notes and the Terminal app, and the company has said the assistant can read health data from Apple’s Health app on iOS when given explicit permission. Each integration extends the assistant’s reach further into the operating system, and iMessage is by far the most sensitive of them so far.
Bloomberg reported that the feature has raised privacy concerns inside Apple, where some executives worry about messages being processed on OpenAI’s servers. The concern is not hypothetical. iMessage conversations contain the kind of material that companies and individuals assume is private: medical details, financial information, family arguments, work secrets. Routing that content through a third party’s infrastructure, even with user consent, changes the data boundary that Apple has long advertised as a selling point of its ecosystem.
The two companies have a complicated history on data sharing. Apple and OpenAI have been partners in some respects, with ChatGPT integrated into Apple’s operating systems, but they have clashed over how user data flows between them. People familiar with the relationship say the companies have disagreed about what data Apple should share with OpenAI and how it should be used, and the iMessage feature sharpens that disagreement because it puts OpenAI’s servers directly in the messaging path.
OpenAI says the feature is opt-in. Users must explicitly enable the integration and grant permission before ChatGPT can read or send messages, and the company has said it will not use the conversation data to train its models by default. The company points out that the assistant only does what the user asks it to do, and that the user can revoke access at any time. Privacy advocates respond that consent is a weak shield when the alternative, declining the feature, means losing the convenience it offers.
The feature is also a test of a larger question: how much of a person’s digital life should an AI assistant be allowed to touch. OpenAI has been explicit about its ambition to build agents that perform tasks across applications, and each new integration brings that ambition closer to reality. Reading and sending messages is a qualitatively different step from summarizing a document or drafting code, because messaging is where people express themselves without a filter, and where a mistake carries immediate social consequences.
Apple’s position is delicate. The company has built its brand on privacy, and its marketing has long contrasted its approach with that of companies that monetize user data. Yet Apple has also become dependent on AI partners, and its own assistant, Siri, has lagged behind the new generation of large language models. The company’s strategy has been to integrate third-party AI while keeping the user’s data on-device where possible, and the iMessage integration tests whether that line can hold.
The technical arrangement matters as much as the policy. OpenAI’s feature works through the Mac’s accessibility and automation interfaces, the same mechanisms that allow screen readers and other assistive tools to control applications. Those interfaces are powerful precisely because they are general, and the security of the feature depends on the permission system being enforced correctly. Researchers who study AI safety note that an assistant with messaging access, if compromised, could do real damage: send a fraudulent transfer request, impersonate the user in a family dispute, or extract secrets through a conversation.
OpenAI has responded to such concerns by emphasizing controls. The assistant asks before sending, and the company says it has built safeguards to prevent the assistant from taking actions that a user has not approved. The feature is rolling out gradually, which gives the company room to adjust before it reaches a wide audience. But the deeper question, about where the boundary between an assistant and its user’s private life should be drawn, is not something a permission dialog can answer.
For users, the feature is a glimpse of a near future in which the AI does the typing. The assistant can draft a reply to a landlord, summarize a group chat before a meeting, or check whether a promised payment was acknowledged, all without the user touching the keyboard. That convenience is real, and it is the reason the feature will almost certainly find users quickly. Whether it also finds a scandal depends on how well the companies handle the data that flows through it, and on how much trust users are willing to extend to an assistant that can now speak for them.
