The throughline in this week's AI news is a widening gap between what AI systems are permitted to do and what users can see, keep, or take back. Google's Gemini is being allowed to act for users - making calls, logging into sites, buying things - while chat retention quietly stretches from 72 hours to two months even with history off. Meanwhile Nvidia has stopped shipping advanced AI chips to China despite White House approval, showing that capability expansion is now conditioned as much by corporate and geopolitical strategy as by product ambition.
Agents Get the Keys
Google's expansion of Gemini Tasks is the clearest signal of where consumer AI is heading. As TechCrunch reported, the Task mode can coordinate calls, log into websites, and make purchases on the user's behalf. That is a meaningful shift in the trust model: the assistant is not merely drafting text but acting with the user's credentials and, potentially, the user's money. For US consumers, the promise is convenience - fewer forms, fewer hold queues - but the exposure is concentrated. Delegated login and delegated payment mean the AI becomes a chokepoint for identity and fraud. The product logic is straightforward: whoever holds the task holds the user. Whether US households are ready to hand over those tasks is a separate question, and one the product rollout implicitly answers for them by normalizing the delegation before the norms exist.
Retention Without a History Switch
Running in the opposite direction is Google's change to Gemini chat retention. TechCrunch reported that Google is extending retention to two months, up from 72 hours, and that the change applies even when history is off, rolling out globally over the coming weeks. The detail that matters is the phrase "even with history off." A setting named for privacy no longer functions as a user might reasonably read it. For US consumers, this is a governance question disguised as a settings toggle: if the off switch does not mean off, the consent model is doing less work than it appears to. Regulators and litigants in the US have grown attentive to exactly this kind of gap between labeling and behavior. The practical result is that conversations with an AI assistant now carry a longer evidentiary and privacy tail than many users will assume, and the burden falls on the user to understand a policy most will never read.
The Export Freeze That Approval Didn't Prevent
Nvidia's halt on advanced AI chip shipments to China complicates any simple story about export controls working. As The Verge reported, Nvidia has stopped sending its advanced AI chips to China despite approval it received from the White House in December last year. The announcement came from interim CEO Colette Kress during an earnings call on Tuesday, and The Verge framed it as a notable shift for a company that, under former CEO Jensen Huang, had pursued the China market aggressively. The lesson for US technology companies is that a government green light is not the same as a business decision to proceed. Export permissions can be granted and still go unused when the strategic, financial, or political calculus shifts. For the US market, that means supply that might have gone abroad stays closer to home - a dynamic that affects allocation, pricing, and competitive positioning among domestic AI buyers, without any formal policy change required.


