ChatGPT Gets Into Apple Messages
Plus: Palantir’s NHS holdout, robot data without robots, and an ad-review gap for deepfake tools.
By Saeed Ezzati5 min read
The audio edition
Listen to this newsletter
0:003:16
Read transcript
ChatGPT can now work inside Apple Messages on a Not just draft a in a separate window but read and conversations them responses send texts and messages The integration launched August 20th and also works with Codex and ChatGPT Work the Assistant into an ongoing communications service That usefulness comes with unusually broad permissions Setup requires authorization in both ChatGPT and macOS including full disk access contacts tools and access to on device message history OpenAI says the integration uses AppleScript and AccessAbility to operate messages In other words retrieval and action arrive together ChatGPT can help find a past exchange but it can also communicate as the user or remove messages OpenAI says the plugin runs locally on the Mac and does not create an index of all messages Those are meaningful stated limits but the company has not fully explained what local processing means operationally The clearest control is approval before every send OpenAI discourages persistent approval because turning it on removes the user's final chance to review a before it goes out in their name For this kind of Assistant the permission screen and the send review step are not incidental setup They are the product's safety boundary That same question who controls the data surface and who gets to on it in the NHS Greater Manchester's Integrated Care Board remains the only one of England's 36 boards not actively using Palantir's Federated Data Platform or FDP It continues with a locally built analytics and data science platform which it says is more flexible primary care data unavailable on the FDP and in some functions is two to years ahead Those claims have not been independently established and Palantir disputes them The UK government now has six months to decide whether to a break clause in a contract worth more than 400 million its planned 2031 expiry The issue is technical but also about public trust and dependence on one foreign vendor Anthropic offers a product design version of the same shift from information to action Artifacts in Claw Desktop describe an unreleased system internally called PARCA that records meetings live speaker attributed transcripts and notes then turns follow ups into assignments for cloud co code or a Each action can include an description full prompt execution type and an auto run setting But public Mac and Windows builds show no usable interface so its release and execution remain unresolved Education provides a more control model Washington Leadership Academy assigns AI permission per task a zero to four rubric from no use to unlimited use Students are taught prompting output evaluation and tool selection while classroom bots are sometimes designed to give hints or even wrong answers that students must challenge A Stanford review of 14 studies found mixed results often improved with AI access but gains frequently disappeared when access was removed The practical lesson across all four stories is to watch the control layer Permissions review gates provenance and assignment specific boundaries will determine whether AI remains an assistive tool or quietly becomes an authorized actor
