Vatican launches AI ethics push & Starship V3 test and Artemis - Tech News (May 18, 2026)

Let’s start with that Vatican development. Pope Leo XIV has set up an in-house study group on artificial intelligence, as the Church prepares a major teaching document expected to argue that AI is reshaping society the way industrialization once did. The emphasis, according to officials and scholars, will be ethics first—human dignity, justice, labor impacts, truth, and the growing problem of misinformation and deepfakes. What makes this noteworthy isn’t just symbolism. The Vatican is trying to become a consistent global participant in AI governance debates at a time when governments and companies are moving fast, and often disagree on what limits—if any—should be set.

On the space front, SpaceX is preparing Starship Flight 12 for Tuesday, and it’s a big one: the first flight of Starship V3, a larger and more capable version that NASA is depending on for future Artemis missions. SpaceX is expected to attempt satellite-deployment demos and an in-space engine relight—both key stepping stones toward more complex missions, including deorbit burns and eventually in-space refueling. The backdrop here is pressure on multiple fronts: Starship has had high-profile setbacks, NASA’s Moon schedule has already been reshaped, and Wall Street is watching closely with renewed talk of a SpaceX market debut. A clean flight won’t solve everything, but it would help re-establish confidence that Starship is moving from spectacular tests to dependable cadence.

Now to software and the accelerating reality of AI agents at work. Engineer Sean Goedecke says his day-to-day workflow has flipped since early 2025: for many tasks, he starts by asking an agent to implement the change, then does a single serious human editing pass before shipping a pull request. The time sink, he says, isn’t polishing the good results—it’s quickly scanning and discarding the weak attempts. Debugging has shifted too: every bug report goes to an agent first, and a majority are diagnosed correctly, though the hardest issues still demand human context, careful data gathering, and sometimes multiple agent restarts to get unstuck. What’s interesting is what he doesn’t delegate: broader communication. He still handwrites most PR descriptions and avoids using LLMs to author Slack updates, architecture decisions, or blog posts—using models more like a reviewer than a ghostwriter. The emerging “meta-skill” is triage: push low-risk execution to agents, while keeping judgment, review, and human accountability firmly in human hands.

That theme—structure matters more than ever—showed up in a separate argument making the rounds: the habits that make code maintainable for humans also make it friendlier to agents. Modular design, clear interfaces, precise domain language, and strong tests aren’t just good hygiene; they’re how you make AI help predictable. The author’s bigger claim is that micro-optimizing a function is less valuable when an agent can generate a decent implementation quickly. The premium shifts toward understanding the domain and setting crisp contracts between components. Alongside that, a new idea gaining traction is “agent hooks”—basically automated guardrails that run at fixed moments during an agent session. Instead of hoping the model remembers rules, teams can enforce them: blocking edits to sensitive files, refusing risky commands, requiring tests to pass, and writing audit logs automatically. Think of it as turning prompts into policy, so reliability doesn’t depend on the agent’s mood on any given day.

Not everyone is celebrating the speed-up. Mitchell Hashimoto is warning about what he calls a kind of “AI psychosis” in companies—where teams convince themselves it’s fine to ship fragile systems because agents can fix problems quickly. He draws an old lesson from reliability engineering: fast recovery is great, but it doesn’t replace resilient design. The risk is that local metrics can look terrific while the overall architecture becomes harder to reason about, harder to change safely, and more likely to fail in surprising ways. In plain terms: if automation makes it easy to move fast, it can also make it easy to lose the plot.

In research publishing, arXiv is tightening enforcement against submissions that show clear signs of unverified AI-generated content—things like hallucinated citations or leftover chatbot meta-comments. The message is blunt: authors are responsible for what they submit, no matter what tool they used. In serious cases, moderators can trigger a one-year ban, and getting back in could require proof of acceptance at a reputable peer-reviewed venue. This matters because arXiv is one of the world’s most important scientific on-ramps—and if trust erodes there, the whole research pipeline gets noisier and slower.

In autonomous vehicles, Uber is taking an increasingly combative tone toward Alphabet’s Waymo—even while Waymo robotaxis still appear inside Uber’s app in some cities. At the same time, Uber is committing huge money to build robotaxi capacity through other partnerships, plus charging infrastructure and city-by-city rollouts. The strategic anxiety is clear: if Waymo and other operators build large independent customer bases, Uber risks becoming optional as a distribution layer. So Uber’s pivot is about leverage—trying to ensure it can offer autonomy at scale, rather than simply renting access to someone else’s fleet.

Astronomy is about to become even more of a data-firehose. The Vera C. Rubin Observatory in Chile is beginning to deliver early data ahead of its decade-long survey that will repeatedly image the southern sky, producing a time-lapse view of a changing universe. Even in early operations, Rubin has already flagged a wave of new asteroids, including unusually fast-spinning objects that hint at surprisingly solid interiors. The deeper shift, though, is operational: Rubin’s alert system has already generated hundreds of thousands of “something changed” notifications in a single night, and it’s expected to scale to millions nightly once the main survey starts. The new bottleneck won’t be finding events—it’ll be deciding which ones deserve follow-up before they fade.

One more big-picture AI story: a new macro argument says the global AI contest is increasingly about physical scale, not just model quality. The United States is still described as leading in frontier models and advanced chips, but China is framed as dominant in the manufacturing-heavy “body layer” needed for robotics—components, factories, supply chains, and sheer deployment volume. The claim is straightforward: robots improve by operating in the real world, and the side that installs more machines learns faster. If that’s right, “embodied AI” becomes less of a lab race and more of an industrial race.

Finally, a quick look at the corporate and career fallout. Bloomberg’s profile of Amazon CEO Andy Jassy paints a company being aggressively reshaped for the generative AI era: major cost cuts, tighter management controls, and enormous spending on data centers and AI infrastructure meant to keep AWS competitive against Microsoft and Google. It’s a reminder that the AI boom isn’t just about flashy models—it’s a capital-intensive arms race in power, networking, and hardware. And on the job side, Microsoft AI chief Mustafa Suleyman is predicting that AI could automate most white-collar jobs within a year to a year and a half—an eye-catching timeline that will be debated, but reflects how quickly executives think capabilities are moving. That anxiety shows up in essays about a widening gap in tech wealth, and in practical advice for job seekers: stop relying only on online postings and use “side doors” like direct outreach and public work that proves what you can do. In a world flooded with AI-polished applications, specificity and visible evidence may be the only signals that still cut through.