Gemini can now run your Mac. Google's own warning: it may delete files
Google is handing paying Gemini subscribers an assistant that clicks, types and moves files on their desktop, and its own warning says it can delete or share the wrong thing.

Google is rolling out computer use in its Gemini app for Mac, letting paying AI Plus, Pro and Ultra subscribers delegate multi-step tasks across local files, desktop apps and websites from a separate background window. The catch Google itself flags, that the assistant can make mistakes including deleting or sharing files, turns a productivity feature into a governance decision for anyone with sensitive data on a laptop.
Google is putting an assistant on your desktop that does not just answer questions, it acts. The company is rolling out computer use inside its Gemini app for Mac, a feature that lets the assistant carry out multi-step tasks across local files, desktop apps and websites, all from a separate background window. It is not for everyone: the rollout is limited to paying subscribers on Google's AI Plus, Pro and Ultra tiers. In other words, the ability to hand your machine a to-do list and walk away is now a paid product, not a demo.
Here is the part Google wants you to read twice. The company warns the feature can make mistakes, including deleting or sharing files. That is the entire tension of the agent era in one sentence: the same firm shipping the capability is telling you, up front, that it might take out the wrong thing. There is no hedge in that warning, and no fix offered alongside it. The user is the backstop.
For context, the assistant category spent years inside a text box. You typed, it typed back, and you did the clicking. Computer use breaks that contract. The model gets hands: it opens apps, navigates sites, moves files. The background window detail is easy to skim past, and it is the most consequential part of the design. A background agent is one you are meant to ignore while it works, which is exactly when a mistake is hardest to catch. An assistant that errs in front of you is annoying. One that errs behind you is expensive.
The paywall tells you where Google thinks the value is. AI Plus, Pro and Ultra are the tiers that get computer use, which makes agentic capability a subscription lever rather than a freebie. That is a deliberate bet. Consumer AI has historically been a costly business to run, and labs have been hunting for features that justify a monthly charge rather than a one-time novelty. An assistant that finishes tasks you would otherwise do yourself is a far easier sell than a smarter chat window. If it works, it is the strongest retention argument Google has.
The risk side is where this gets uncomfortable for anyone with something to lose. A tool that can delete or share files sits on top of the same folders that typically hold contracts, payroll data, client records and unreleased product plans. Google's warning is effectively a disclaimer, and disclaimers do not stop a file from leaving the building. Expect security and IT teams to treat this the way they treat any new software with write access: permissioning first, sandboxing second, and a long conversation about audit trails third. The feature's usefulness and its blast radius scale together.
On competition, Google is not alone in chasing this. The industry's biggest labs have been pushing assistants from answering toward acting, because action is where the pricing power lives. Whoever makes an agent reliable enough to trust with real work captures the subscription, the enterprise contract and the developer ecosystem that follows. Reliability, not raw model quality, is the bottleneck. And reliability is precisely what a warning about deleted files undercuts.
For executives, the practical question is not whether to use this. It is who gets to. A background agent with write access to local files is a governance item, not a productivity hack. Companies that let it run on a personal Mac with access to shared drives are accepting a risk they have not priced. The safer pattern is boring and familiar: scope the access, watch the logs, keep the sensitive directories out of reach.
The bigger signal is that the agent era has arrived with its training wheels still on. Google is shipping capability and caveat in the same breath, which is honest and also a warning shot. The coming months will decide whether "it can make mistakes" reads as a temporary disclaimer or a permanent feature of how we work. For now, the assistant is on your Mac, the window is in the background, and the delete key is in its hands.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Technology
Anthropic mathematician claims Claude helped crack the Riemann hypothesis
If the proof holds, AI has toppled math's most famous open problem; if it does not, it is a case study in announcing breakthroughs before anyone verifies them.
AI's $4 Trillion Wipeout Erases Nasdaq's Gains for the Year
A brutal week took roughly $4 trillion off the AI sector and left the Nasdaq flat for the year, and the damage ran the full length of the supply chain.
Anthropic mathematician says AI proved the Riemann hypothesis. Verification will take time.
A single unverified claim from inside Anthropic now sits against more than 160 years of open mathematics, and the gap between asserting a proof and surviving peer review is the entire story.


