Tesla FSD and the handoff gap & Tech platforms face new liability - Tech News (Mar 27, 2026)

First up, a sharp critique of Tesla’s “Full Self-Driving” strategy is making the rounds, and it focuses on a problem that’s more human than technical: the handoff. The argument is that when automation works well most of the time, people stop actively supervising—so when the system fails in a weird, rare way, the driver is effectively waking up mid-emergency. The piece points to earlier Waymo findings where safety drivers quickly became inattentive, which is part of why Waymo moved away from designs that demanded sudden takeovers. What really lands here is a recent story from a former Uber self-driving lead who knew this risk intellectually, tried to step in during an unexpected maneuver, and still crashed. The takeaway is blunt: “be ready to take over” sounds reasonable, but it may not be realistic at scale.

Staying with safety and accountability, the legal climate around social platforms just got more serious. In two separate jury verdicts, Meta was found liable in cases brought on behalf of children, and YouTube was found liable in a Los Angeles trial tied to claims of addictive product design. What’s notable is how these cases are threading the needle around the usual platform defenses by focusing on design choices and duty of care, not just user-posted content. And in Europe, Germany is debating reforms after allegations involving AI-generated sexual imagery, pushing toward clearer criminal liability for deepfake porn and faster platform takedowns. Put together, it’s a sign that lawmakers and juries are less willing to accept “unintended side effects” as the end of the story—especially when minors are involved.

Now to Apple, where the AI era is reshuffling alliances and poaching talent. Bloomberg reports Apple is handing out sizable retention stock awards to people on the iPhone Product Design team, largely to keep key hardware minds from leaving—particularly to OpenAI, which has reportedly hired dozens of former Apple staff across multiple device lines. At the same time, Apple’s AI strategy looks increasingly hybrid. One report says Apple has deep access to Google’s Gemini models inside Apple-run data centers, including the ability to use Gemini as a “teacher” to help train smaller models that are easier to run. And another report suggests Apple is preparing to let Siri route requests to multiple third-party AI assistants—not just ChatGPT—through new system-level integrations in iOS 27. If that happens, it would turn Siri into more of a traffic director, and it could reshape how AI subscriptions are discovered and paid for on iPhones.

One of today’s most consequential research stories is about something called “The AI Scientist,” a system that aims to automate the whole machine-learning research loop—from proposing ideas, to running experiments, to writing a paper, and even generating peer-review style feedback. The team also built an automated reviewer modeled on major conference guidelines, and they claim it lines up with human accept-or-reject decisions surprisingly well. In a controlled test, they even submitted AI-generated manuscripts to a workshop under blind review, with one scoring above the typical acceptance bar before it was withdrawn because it was AI-made. The point isn’t that machines are “taking over science” tomorrow. It’s that the cost of producing plausible-looking research may be dropping fast—so norms around disclosure, citation integrity, and review overload are about to matter a lot more.

On the enterprise side, Oracle is betting that the database becomes the control center for “agentic” AI—software that can take multi-step actions, not just answer questions. The pitch is that if AI features live closer to governed business data, companies can reduce messy pipelines, cut deployment failures, and make audits simpler. Oracle is also emphasizing tighter access controls so AI systems can be restricted based on who’s asking and what they’re trying to do. In a different corner of the developer world, Stripe is previewing a command-line feature that helps teams spin up app environments and sync credentials more safely across laptops and CI systems. The common thread is less glamour and more plumbing: companies are trying to make AI projects easier to ship without turning security and reliability into an afterthought.

Zooming out to the macro AI race, Nvidia’s Jensen Huang used his sprawling GTC 2026 keynote to reinforce a framing that’s becoming the industry’s north star: compute as an industrial resource. He’s pushing the idea of “AI factories” where the output is tokens—basically, the raw material that powers AI products. The strategic importance here is narrative-setting: convincing partners and supply chains to build for demand years in advance. And speaking of demand, SoftBank just secured a massive bridge loan to back investments including OpenAI, doubling down on the belief that the next phase of AI competition will be won by whoever can finance the most compute, the fastest. It’s a reminder that AI is increasingly an infrastructure story—and infrastructure rewards scale.

In space, the big headline is SpaceX preparing for an IPO that could be anything but typical. Reporting suggests Elon Musk wants a more controlled, almost theatrical approach, potentially bringing investors to SpaceX facilities instead of doing a standard roadshow. That matters because it could set a new precedent for how mega-companies market themselves to Wall Street—especially when the founder already dominates the media cycle. In a separate interview, SpaceX President Gwynne Shotwell said the company’s merger with Musk’s xAI is still early, but she expects AI to play a growing role inside SpaceX, from factory automation to far more ambitious ideas like space-based data centers. She also mentioned the company seeking authority tied to a vastly expanded satellite vision, which naturally raises questions about congestion, safety, and who sets the rules in orbit.

NASA, meanwhile, is making a notable pivot: it’s pausing work on the Gateway lunar-orbit station and repurposing a key module into a nuclear-electric propulsion demonstration aimed toward Mars. The plan is to combine a spacecraft bus that’s already fairly mature with a new fission reactor system, with an aggressive target tied to a late-2028 launch window. NASA is pitching it as a tighter, more executable nuclear program than past efforts that ballooned or were canceled. If it works, it’s not just a propulsion milestone—it’s a statement that the U.S. can operate reactor-powered systems beyond Earth orbit, which has implications for future deep-space missions where solar power becomes limiting.

And finally, a human-focused AI story with real near-term meaning: a Neuralink trial participant says he’s playing World of Warcraft using only thought-based control about three months after receiving the company’s brain implant. The big deal isn’t the game itself—it’s the complexity. Navigating a fast-moving interface, managing inputs, and staying precise over time is a stronger signal than a one-off cursor demo. Neuralink is still in early, tightly supervised trials, so this isn’t a mass-market moment. But it is a glimpse of what “computer access as assistive tech” could look like if the reliability holds up.

Quick science and energy roundup before we wrap. In Japan, a RIKEN team published results from a long-running experiment repeatedly cloning mice from the previous clone. The striking finding is that the limiting factor wasn’t mysterious epigenetic decay—it was genetic wear and tear. Mutations and major DNA structural damage piled up across generations, likely because cloning doesn’t get the cleanup benefits of normal reproduction. In space science, astronomers using archived Hubble data documented a comet that slowed its rotation and then reversed its spin direction—apparently pushed around by uneven jets of gas and dust. And on Earth, Southeast Asian governments are revisiting nuclear power plans as oil-price shocks and AI data-center growth expose how fragile the region’s energy supply can be. Different stories, same theme: long-term systems break in ways that are easy to ignore—until they aren’t.