AI-generated papers test peer review & Wikipedia bans AI-written content - News (Mar 27, 2026)

We’ll start in AI research, where a team is pitching what they call “The AI Scientist”: a system meant to automate the full machine-learning research loop—coming up with ideas, checking whether they’re genuinely new, running experiments, analyzing results, and writing a paper. They even built an “Automated Reviewer” modeled on major conference guidelines, and say it matches human accept-or-reject decisions at roughly human-level agreement, including on a harder post-cutoff dataset designed to reduce contamination concerns. The eye-catcher: with organizer and ethics approval, three AI-written manuscripts were submitted to an ICLR 2025 workshop under blind review; one scored above the usual acceptance threshold, but was withdrawn under a pre-committed rule because it was AI-generated. The authors also highlight today’s weak spots—buggy implementations, shallow ideas, and made-up citations—while warning that as models improve, peer review could buckle without clear disclosure norms and safeguards.

That concern about trustworthy text is showing up in a very different corner of the internet: Wikipedia. The encyclopedia has updated policy to ban the use of AI tools to generate or rewrite article content. Editors have been battling over this for a while, and the argument is pretty simple: Wikipedia’s value is grounded in sourcing, neutrality, and verifiability, and AI text can sound confident while quietly bending meanings or inserting unsupported claims. There are narrow exceptions—translation work and small copyedits to your own writing—so long as humans review it and no new content is introduced. The bigger signal is cultural: Wikipedia is choosing to lean harder into “human-curated and source-backed” at a moment when AI-written summaries are flooding search and social feeds.

On the business side of AI, new reporting says Apple’s partnership with Google gives Apple extensive access to Google’s Gemini models inside Apple’s own data centers. Beyond powering features directly, the detail that stands out is model distillation: using a large model’s outputs to train smaller models that are cheaper to run and easier to tailor. The practical implication is speed and efficiency—Apple could ship more capable features while keeping more of the workload within its own facilities, and potentially push more intelligence closer to devices without needing the heaviest compute all the time. Strategically, it’s also a reminder that the AI race isn’t just about who trains the biggest model—it’s also about who can productize capabilities reliably, at scale, and on tight resource budgets.

And to fund this next phase of AI competition, the money keeps getting bigger. SoftBank says it has secured a forty-billion-dollar unsecured bridge loan, set to mature in 2027, to support investments in OpenAI and general corporate needs. It builds on SoftBank’s earlier commitment to invest heavily via Vision Fund 2, further tying Masayoshi Son’s strategy to the AI boom. The point to watch isn’t only SoftBank’s risk appetite—it’s the broader trend: access to capital and compute is becoming the key chokepoint. As AI infrastructure projects scale up, financing starts to look less like traditional venture bets and more like industrial-era buildouts.

Switching to health and biology, a Nature study is drawing a clearer line between chronic inflammation and cancer risk—without relying solely on DNA mutations. Researchers studying repeated colitis in mice found that even after the gut looks healed, some colon stem cells keep a lasting “epigenetic memory” of inflammation. In plain terms: the cells’ control system stays rewired in a way that makes growth programs easier to switch back on. The team linked that memory to persistent changes in chromatin accessibility—especially around AP-1-related regulatory sites—and showed the effect can persist for months. Organoids grown from previously inflamed tissue kept a more regenerative, hyperproliferative behavior even outside an inflammatory environment. When tumors were later triggered, recovered mice developed bigger early lesions, suggesting these primed cell “fields” can fuel tumor outgrowth. The encouraging note: short-term AP-1 inhibition at tumor initiation reduced tumor size specifically in the post-colitis setting, hinting at prevention strategies and biomarkers for people with inflammatory bowel disease.

Another cancer-and-immunity story has an unusual twist: some tumors appear to express a protein best known in the brain—the NMDA receptor—and that can provoke an immune response with two faces. Researchers found evidence of NMDA receptor expression in a small fraction of cells in triple-negative breast cancer, then built a mouse model where tumor cells could be made to express it. In some animals, the immune system produced strong anti-NMDA receptor antibodies and the tumors regressed, suggesting the antibody response can help control cancer. But the same class of antibodies is linked to anti-NMDA receptor encephalitis in humans, a serious neurological condition. In mice, certain antibodies could also trigger neurological effects. In a patient cohort, elevated anti-NMDA receptor antibodies correlated with tumor expression and better disease-free outcomes. The takeaway is a delicate balance: immune responses that are beneficial against cancer may carry neurotoxicity risks, and future therapies may need to separate the anti-tumor upside from the neurological downside.

Staying with cancer, researchers are also getting sharper answers on a difficult pediatric brain tumor: supratentorial ependymoma driven by the ZFTA–RELA fusion. These tumors often have relatively few DNA mutations, which has pushed scientists to look at development and epigenetics. The new work suggests the fusion doesn’t need to drastically rewire chromatin; instead, it mostly activates genes that are already “primed” in certain progenitor-like brain cells. Using single-nucleus maps of development and tumor samples, the study points to a specific developmental regulatory program—linked to PLAG and PLAGL motif activity—that’s active in cycling progenitors and normally fades with differentiation. Tumor cells looked like stalled or incomplete development, and the more differentiated-like malignant cells were largely non-cycling, implying differentiation can act as a brake even when the oncogenic program persists. It’s not a treatment announcement, but it does sharpen where to look for vulnerabilities: developmental windows, epigenetic states, and strategies that push cells toward maturation rather than endless division.

In infectious disease surveillance, a large genomic study is updating an old but important bacterial identity problem: capsule typing in E. coli. Traditional K-antigen serology has fallen out of routine use, and the new work proposes a practical genome-based scheme for group 2 and group 3 capsule loci. After screening more than eighteen thousand genomes, researchers cataloged far more capsule diversity than the classic phenotypic types capture, including capsules that don’t react to existing antisera. In European infection datasets, a relatively small set of capsule types dominated bloodstream infections and many drug-resistant cases—but the study also estimates that some less-famous capsule types may have higher “invasive potential.” Add in signs of rapid capsule switching in major resistant lineages, and the capsule starts looking like a moving target for vaccines, phage therapy, and outbreak tracking—one that may vary significantly by region.

Now to the courts, where there’s been a notable shift in momentum against social media platforms. 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 different this time is the legal framing: rather than focusing only on harmful user content, these cases leaned on product design—claims that features were intentionally built to drive compulsive use, particularly in young users. That approach can help plaintiffs sidestep Section 230 defenses that often block content-based claims. Meta and Google say they disagree with the verdicts and are weighing appeals, but the bigger impact may be downstream: more lawsuits, more pressure for regulation, and a renewed debate over what companies owe minors in an attention-driven marketplace.

Finally, geopolitics. Ukrainian President Volodymyr Zelenskiy says the United States has made prospective security guarantees for a peace deal conditional on Ukraine withdrawing from—and effectively ceding—the entire eastern Donbas region to Russia. Zelenskiy argues that without strong, enforceable guarantees, any settlement risks becoming a pause rather than an end, giving Russia space to restart hostilities later. He also warns that giving up Donbas would hand Russia important defensive positions and weaken broader European security. Zelenskiy says a security-guarantees document had been essentially ready earlier this year but now needs more work after recent talks, and he claims Russia is betting Washington will lose interest as the war drags on. He thanked the Trump administration for continuing Patriot air-defense deliveries, while also saying supplies remain insufficient and Ukraine is expanding its own long-range drone and missile production.