What should be the role of AI in the classroom? That’s an emerging issue that teachers and administrators are grappling with. For some, it may mean reverting to pre-Google days.

This week, Google said it will turn off its Gemini AI tool by default for accounts of school students, giving school administrators options to opt in.

But Gemini has effectively been bundled with Google’s search engine: a school that wants no AI at all for its students risks losing Google Search, too. Schools that want students to continue to access the search engine will have to accept features such as AI-generated summaries. In other ways, teachers are going back to the future.

My colleague reported on how, in Silicon Valley, take-home writing assignments are increasingly giving way to quick oral exams, where teachers ask students to verbally explain their reasoning. Homework is being replaced by more supervised in-class writing and live discussion, where a teacher can ensure the thinking is being done by the student, not their AI assistant. And the lessons themselves are being redesigned around something AI still struggles to fake convincingly: a student’s own reasoning and perspective, not just the correct factual answer.

Outside the classroom, Americans also had a chance to cast their vote on the AI boom this week in the midterm elections. Scroll down to learn more.

OpenAI employees warned about operational security lapses for years Google extends voluntary buyout offers to employees in several countries AI shopping agents could threaten Amazon and Walmart's ad business Elon Musk's business empire braces for Democratic wave as Texas flips blue Chinese chipmaker YMTC puts IPO on hold as weaker markets weigh, sources say AI sales startup Nooks secures $500 million in funding for expansion In this week’s midterm elections, Americans sent AI companies a message about data centers: slow down. Of the 41 data center-related ballot measures voted on this week, all but one broke in the anti-data center direction, according to Ballotpedia.

Meanwhile, in Texas, Georgia, and Ohio — all places with plans for significant data center capacity growth — Democrats flipped governorships, with voters electing candidates who have promised to pause new data center permits and tighten construction rules. Texas matters most of all: investment bank Jefferies estimates the state accounts for roughly 26% of the entire planned US data center pipeline.

Texas is as good a test case as any for what happens next.

Governor Greg Abbott had already imposed a permit pause tied to reviews of how proposed data centers would strain the state’s power grid and water supply; completing those reviews could eventually reopen the approval process. Governor-elect Gina Hinojosa wants to go further, keeping development paused until state lawmakers pass new protections for communities and utility customers, and revisiting the tax incentives developers currently enjoy.

That likely means a longer, less predictable wait, not just for a technical review to finish, but for an entire legislative process to play out. “This election could extend an existing pause and change the conditions for restarting,” said Mona Dajani, partner and global co-head of infrastructure, energy and AI at law firm Cooley.

“For developers, it’s a meaningful difference between waiting for an audit and waiting for lawmakers to agree on new rules. They still have equipment commitments, financing costs, and customer deadlines while they're waiting.”

Not every developer is equally exposed. A project still comparing potential sites has room to wait, or to simply build somewhere else instead. A project that has already secured land, ordered equipment, lined up utility upgrades, and promised a tenant a delivery date does not have that flexibility — relocating at that stage is expensive, and local permits and grid connections don’t just move. That dynamic tends to favor the biggest players, according to Dajani.

Hyperscalers that already operate data centers in multiple states can shift workloads around a delay, or redirect new investment elsewhere, as capacity and customer needs allow. A smaller developer betting everything on one unfinished campus and one tenant has none of that cushion — a single delay can put its financing, equipment orders, and revenue projections all at risk simultaneously. The big players’ advantage also comes on the financing side.

When a project’s opening date becomes uncertain, financing tends to get more conditional: developers may need to put in more equity upfront, hold larger reserves to cover delay costs, or seek extensions on commitments they’d already lined up. Some lenders may simply choose to wait for approvals to clear before releasing money at all — meaning a signed financing deal doesn’t guarantee that the cash is actually accessible when a developer needs it. None of this means demand for AI infrastructure is going away.

But this week’s results suggest support for the infrastructure boom needs more than just Silicon Valley’s lofty ambitions. To be sure, this isn’t a simple story of voters versus the industry.

The same newly elected leaders now have to balance a pause on permits against the practical reality that data centers also create jobs — including union construction jobs belonging to many of the same voters. My colleagues detailed the split opinions inside unions in Nevada. Voters may not be able to vote on how fast AI research accelerates.

But they can, and did, vote on how fast the infrastructure behind it gets built in their backyard — which, in turn, will influence how fast AI can impact on society and how costly that will be.

Security researchers already have a name for what just happened: Black October. Two critical vulnerabilities in PHP, the programming language running roughly 70% of the internet’s websites, surfaced last month after sitting dormant in the code for years — found only because AI models were finally capable of chaining together the complex steps needed to prove they were exploitable.

The chart quantifies it: new vulnerabilities tracked by NIST’s database jumped from 4,289 in October 2025 to 17,059 in October 2026, nearly a fourfold increase in a single year, with the line rising steeply since spring.

Finding a bug used to require real skill, and turning that bug into a working exploit could take an experienced professional weeks or months. AI has thrown that timeline out the window: models can now subject a codebase to round-the-clock scrutiny, finding one flaw after another, and turn a working exploit around in hours instead of months. Security professionals warn AI-assisted attacks are likely to get worse before, as some believe, AI can itself be used to write more secure code and find vulnerabilities faster.