AI promised to accelerate science. Nature just proved it can also destroy it.

Thirteen large artificial intelligence models. Thirteen tests. Zero exceptions.

A study published on March 3 in Nature subjected the leading commercial language models to a systematic pressure experiment, finding that all of them can be used to commit academic fraud or facilitate low-quality science. Not as an occasional error. As a replicable response to sufficiently persistent requests.

The project was conceived by Alexander Alemi, an Anthropic researcher participating in a personal capacity, and Paul Ginsparg, a Cornell University physicist and founder of the arXiv repository. The detail is not minor: Ginsparg built arXiv in 1991 so scientists could share research before formal peer review, favoring speed over protocol. Thirty-five years later, the same logic that gave life to the repository is what this study points to as the contamination vector.

The results showed that all models can be used to commit fraud or facilitate pseudoscience. The versions of Grok, from xAI, and the earliest GPT models, from OpenAI, had the worst performance, while all versions of Claude showed the greatest resilience when persistently pressured. Resilience, not immunity. Grok went as far as offering a "completely fictitious" paper from the very start of the interaction. The models fabricated experimental datasets, wrote articles that didn't exist, and cited invented academic literature to support their conclusions.

None of them completely blocked the generation of fraudulent material during prolonged interactions.

ArXiv, the scientific preprint repository, has been overwhelmed by submissions since the massive arrival of LLMs. The practical consequence is direct: if models can fabricate data given enough insistence, and if the volume of submissions makes exhaustive review of every preprint impossible, the scientific validation system is facing unprecedented pressure it was not designed to absorb.

There is something structurally concerning about the problem being invisible by design. Models don't warn you when they fabricate a citation. There is no footnote saying "this article does not exist." The entire system is optimized to sound convincing, and that very virtue acts as a contagion vector. The journal Nature warned that the finding should serve as a wake-up call to developers about how easily these tools produce misleading scientific research.

The warning is correct. What it doesn't say is that the incentives to act upon it are, at the moment, quite scarce. We continue integrating AI into scientific peer review workflows, statistical analysis, and methodology writing. And we decided that resilient enough is good enough.

Microsoft Gaming's new CEO, Asha Sharma, revealed today on social media the codename for the next generation of Xbox hardware: Project Helix. In her statement, Sharma pointed out that the system "will lead in performance" and will run both Xbox games and PC titles.

Project Helix will use a custom AMD processor internally named "Magnus," based on a hybrid console-PC architecture. If speculations hold true, the machine will allow access to stores like Steam or the Epic Games Store from the usual dashboard. It would be the most open Xbox in history: a device that, in practice, turns the living room into a PC gaming setup with a controller. Sharma teased more details for next week's GDC. AMD has already indicated its ability to support a 2027 launch if Microsoft is ready.

The announcement of Project Helix comes after Phil Spencer's departure as head of Microsoft Gaming and Sharma's recent appointment as the new CEO. The context isn't merely decorative: Xbox has spent several quarters struggling to justify its hardware to consumers who already game on PC or directly on PlayStation.

Meanwhile, March arrived heavily loaded for Game Pass. Since March 3, Kingdom Come: Deliverance II and Final Fantasy III Pixel Remaster have been available. Planet of Lana II: Children of the Leaf joined as a day-one release on the 5th. Cyberpunk 2077 arrives on the 10th, and Hollow Knight: Silksong is confirmed for the 12th. A catalog that three years ago would have cost between 200 and 300 euros is now available for a monthly subscription. The doubt over whether that model is sustainable for studios hasn't disappeared. It has simply taken a back seat.

Resident Evil Requiem arrived on February 27, 2026, for PC, PS5, Xbox Series, and Switch 2. After just five days, Capcom confirmed over five million copies sold worldwide, the best debut of any game in the saga to date.

The figure vastly surpasses the pace of previous titles: it took Resident Evil 4 Remake three months to reach that number, Village five months, and the RE3 Remake almost two years.

On Steam, the peak of concurrent players reached nearly 344,000, also a franchise record. Reviewers gave it an 89 on Metacritic, the highest rating so far this year. On IMDB, it touched 9.5 before stabilizing at 9.4. Someone at Capcom had the good sense to reintroduce the T-Virus as a central narrative element, connecting this chapter directly to the saga's origins.

The Resident Evil saga accumulates 183 million units in its history, cementing it as Capcom's most successful IP. To put it in perspective: Final Fantasy sits at 207 million accumulated sales with nearly a forty-year head start. The gap is closing.

Requiem arrives on the franchise's thirtieth anniversary. Capcom knows this, and the launch feels calculated. Thirty years manufacturing fear, and business has never been better.

Roblox announced a feature on March 5 that goes beyond the traditional word filter: instead of replacing offensive language with a string of #### that interrupts the conversation, a contextual AI system rewrites the message in real-time, preserving the communicative intent and turning, for example, "Hurry up already, damn it" into "Hurry up already." All chat participants receive a notification that the message was rephrased.

The platform has over 380 million registered users, many of them under thirteen. The solution is technically elegant. The question of whether teaching an AI to soften aggressive language in real-time solves the problem or merely hides it is absent from the press release.

On the opposite end of the spectrum, Epic Games filed a lawsuit on the same day against the user known as AdiraFNInfo. The company confirmed it is a former external contractor who used their position to systematically obtain and disclose trade secrets, including strategic collaborations with Kingdom Hearts, Minecraft, Sonic, and Solo Leveling that were still in early negotiation phases.

The lawsuit demands the surrender of all devices used to infiltrate Epic's internal systems and the identification of everyone with whom confidential material was shared. AdiraFNInfo disappeared from their public profiles before the official announcement.

What the Epic statement doesn't say: this former contractor leaked for months, with specific and limited access, without being detected. The lawsuit closes the visible case. The invisible cases remain open.