AI-written papers hit peer review & NASA pivots to lunar living - Tech News (Mar 26, 2026)

We’ll start with that research bombshell. A team unveiled what they call “The AI Scientist,” an end-to-end system designed to automate much of the machine-learning research loop: proposing an idea, checking the literature for novelty, running experiments, analyzing results, and writing up a paper. They also built an automated reviewer modeled on major conference guidelines, and they claim it tracks human accept-or-reject decisions surprisingly well. The most attention-grabbing detail: with oversight and prior approval, the researchers submitted AI-generated manuscripts to an ICLR workshop under blind review; one scored above a typical acceptance cutoff, and was then withdrawn because it was AI-made, per their precommitted protocol. The takeaway isn’t that machines are “done” replacing scientists—limitations like shallow ideas and bogus citations are still a real problem. It’s that the workflow is becoming scalable, and peer review may be the first system to feel the pressure.

On a related theme—understanding what AI is actually doing—Anthropic published new interpretability work that acts like a kind of “microscope” for Claude. In plain English, they’re trying to separate what the model says it’s doing from what it’s actually doing internally. Their results suggest a key caution for anyone who relies on step-by-step explanations: models can produce convincing reasoning that doesn’t match the internal process that led to the answer. They also report hints of planning—like aiming for a rhyme before writing a line—and that some concepts appear shared across languages instead of living in separate “English” or “French” compartments. If you’re using AI in high-stakes settings, the message is clear: fluent explanations are not the same thing as reliable auditing.

Now to big-tech AI strategy, where the lines between competitors keep blurring. Reporting says Apple has been granted unusually deep access to Google’s Gemini models within Google’s own data centers. The practical angle here is distillation: using a powerful model to generate high-quality outputs, then training smaller models that can run cheaper and, crucially for Apple, run directly on devices. That could help Apple ship AI features that are faster, more private, and less dependent on a network connection—while still benefiting from a frontier model as a “teacher.” The interesting tension is that Apple is also building its own foundation models, so this looks like a two-track plan: partner for speed now, but keep a path to independence later.

Staying in the AI infrastructure lane, Arm made a statement by launching its first in-house processor aimed at AI data centers. Arm’s identity for decades has been “we design, you build,” licensing CPU blueprints to partners. Putting an Arm-branded chip into the market changes that dynamic. Arm says the focus is server racks running agent-like AI workloads, not phones or laptops, but it still raises an obvious question: when the platform provider becomes a competitor, how do customers respond? The fact that some major Arm ecosystem names publicly cheered the move—while a couple of the biggest did not—adds a little intrigue about where alliances may shift next.

Google, meanwhile, is pushing urgency on a different front: cryptography. The company says it has moved up its internal target for “Q Day” readiness to 2029—effectively a warning that the industry should accelerate the shift away from today’s mainstream public-key cryptography toward post-quantum alternatives. For Android, Google outlined a path to weave post-quantum signatures into core device trust checks. The why is straightforward: even if a cryptographically relevant quantum computer isn’t here yet, attackers can harvest encrypted traffic now and decrypt it later. And for signatures—used to prove software is authentic—waiting until the last minute is especially risky, because you can’t retroactively prove yesterday’s software updates were trustworthy.

Let’s shift to courts and accountability, where social platforms just had a rough run. Juries in California and New Mexico delivered rare wins for parents and child advocates, finding Meta liable in both cases—and YouTube liable in the Los Angeles trial. What’s notable is the legal strategy: these cases emphasized product design and alleged addictiveness, rather than focusing on user-posted content. That matters because it can bypass the usual legal shields that tech firms lean on. If appeals don’t wipe these verdicts away, the ripple effects could be substantial: more lawsuits, more pressure to redesign youth-facing features, and more leverage for regulators who argue the platforms have treated harm as a cost of doing business.

And while we’re on Meta, CNBC reports the company laid off several hundred employees across multiple groups, including parts of Reality Labs. Meta describes it as routine restructuring, but it fits a familiar pattern: trim spending in older bets while pouring resources into AI to keep pace with OpenAI, Google, and Anthropic. It’s another reminder that the AI race isn’t just about models—it’s about budgets, headcount, and which projects survive the next planning cycle.

Now for space, where NASA is trying to turn the Moon from a destination into an address. The agency unveiled an ambitious Moon base strategy estimated around twenty billion dollars, explicitly framing it as a shift from short visits to long-term habitation—“this time, the goal is to stay.” On top of that, NASA appointed longtime engineer Carlos Garcia-Galan to run the Moon base effort, and signaled a pivot toward a surface outpost as the central objective, with the Lunar Gateway potentially pushed later. The significance here is less about a single structure on the Moon and more about a management reset: a sustained presence forces decisions on logistics, supply chains, landing cadence, partner roles, and how to keep a program from drifting into endless architecture debates.

Over in fundamental physics, CERN pulled off a surprisingly practical milestone: transporting antiprotons by road for the first time. They test-drove a portable containment system around the campus, keeping the particles trapped and intact despite vibration and movement. Why should anyone outside particle physics care? Because some of the most sensitive antimatter measurements are limited by the environment near big accelerator equipment. If you can safely move antimatter to quieter labs, you can run cleaner experiments—potentially sharpening tests that probe why the universe has so much more matter than antimatter. The next step is longer trips to outside facilities, which turns this from a clever demo into a real logistics challenge.

Finally, a policy story that’s increasingly hard to ignore: AI-generated sexual abuse material. In Germany, a high-profile allegation involving AI-generated pornographic images has intensified calls to modernize laws that were built for older forms of image abuse. A cross-sector plan is urging lawmakers to criminalize creation and distribution of pornographic deepfakes, curb “nudify” apps, and force faster takedowns—while the government says a draft bill is ready to close loopholes. This is part of a broader shift: as generative AI makes fabrication cheap and fast, legal systems are being pushed to treat synthetic abuse as abuse, not merely “speech” or a civil dispute.

And one more to close the loop on AI’s societal footprint: OpenAI’s nonprofit foundation says it plans to distribute a billion dollars in grants over the next year, focused on life sciences and health, plus programs meant to reduce AI’s downsides, including impacts on jobs and mental health. The funding promise is huge—but it also reopens questions about governance and public-benefit commitments, given OpenAI’s unusual structure and its high-stakes commercial trajectory. In other words: the money is real, the needs are real, and scrutiny will be, too.