AI Daily Briefing — 2026-05-26

During Google I/O, DeepMind CEO Demis Hassabis declared humanity stands at the “foothills of the singularity,” a moment when AI surpasses human intelligence. However, the context of his statement—a segment on scientific AI featuring WeatherNext, a tool that predicted Hurricane Melissa’s landfall and potentially saved lives—revealed a tension between grand visions and practical achievements. While WeatherNext’s real-world impact is significant, it underscores a divide in AI for science: specialized tools versus agentic, LLM-based systems that could autonomously conduct research.

Agentic systems are advancing, with Google Cloud’s Pushmeet Kohli noting a shift toward AI that “begins to do science.” This raises questions about investing in specialized tools like AlphaFold or WeatherNext, as autonomous AI scientists could collaborate with humans or work independently. Despite this, Google continues developing tools like AlphaGenome and AlphaEarth, and AlphaFold remains widely used by over three million researchers. Yet, signs of realignment emerge: Nobel laureate John Jumper now works on AI coding, reflecting a prioritization of agentic systems, as coding abilities are critical for their success.

Across the industry, agentic researcher systems show promise, with OpenAI recently announcing breakthroughs. This evolution suggests a future where AI and humans collaborate as peers, or AI drives scientific progress alone. While specialized tools remain valuable, the focus is shifting toward autonomous systems that could redefine discovery.

Artificial intelligence companies are racing to build systems that truly understand the external world, moving beyond the limitations of large language models. Recent breakthroughs have pushed so-called world models into the spotlight, sparking a critical question: can AI learn to comprehend reality as humans do?

In a recent roundtable discussion, MIT Technology Review AI’s editor in chief Mat Honan, senior AI editor Will Douglas Heaven, and AI reporter Grace Huckins explored this frontier. The conversation, recorded on May 21, 2026, delved into how AI might bridge the gap between data processing and physical-world interaction.

The panel examined emerging research and bold visions from leaders like Yann LeCun, who has proposed a new architecture for AI. They also touched on related innovations, such as how Pokémon Go is giving delivery robots an inch-perfect view of the world, hinting at practical applications for these advanced models.

As AI continues to sprint forward, the question remains whether machines can develop a genuine understanding of the world around them. The discussion highlighted both the promise and the profound challenges ahead in this next chapter of artificial intelligence.

Pope Leo XIV released his first encyclical, Magnifica Humanitas, on Monday, framing artificial intelligence as a modern lens for ancient human struggles. While the document, titled “safeguarding the human person in the time of artificial intelligence,” uses AI as its entry point, the pontiff’s core concerns are timeless: widening inequality, democratic decay, and the consolidation of power among elites who disregard the common good. The 200-page text, presented alongside Anthropic co-founder Chris Olah, argues that technology developed by a small, unaccountable minority cannot serve humanity broadly.

The encyclical warns that AI amplifies existing disparities, allowing those with wealth and data to manipulate information, sway elections, and control economic systems. Pope Leo calls for “clear criteria and effective oversight” rooted in community participation, rejecting an AI arms race driven by geopolitical or commercial dominance. He insists that technical power does not grant the right to govern, echoing critiques of unregulated tech influence that predate modern algorithms.

These dynamics are not new. The pope’s predecessor, Leo XIII, addressed similar industrial-era power concentrations in 1891. Today, examples abound—from Elon Musk’s Twitter takeover to tech elite funding of anti-regulation campaigns. Notre Dame law professor Paolo Carozza, a Meta Oversight Board chair, told TechCrunch AI that AI-driven misinformation undermines truth and democracy, while data harvesting threatens cognitive freedom. The encyclical ultimately reframes AI as a magnifier of age-old challenges, urging collective governance over elite control.

Elon Musk appears to be abandoning his long-standing vision of a solar-powered economy on Earth, according to a recent SpaceX IPO filing. The document reveals that his AI company, xAI, is heavily investing in natural gas turbines to power its data centers, with plans to spend $2.8 billion more on the fossil fuel. This marks a stark departure from Tesla’s original Master Plan, which aimed to transition the world from hydrocarbons to a solar electric economy.

While xAI has purchased $697 million worth of Tesla Megapacks for battery storage, it has not bought a significant number of solar panels. Instead, the filing emphasizes space-based solar power as the future, with SpaceX claiming orbital arrays can generate over five times the energy of terrestrial panels due to constant sunlight. Musk and other tech executives are now exploring the possibility of launching AI data centers into orbit to bypass earthly opposition and power constraints.

However, the economics of space-based data centers remain daunting. Power costs for satellites are far higher than terrestrial facilities, and protecting hardware from space conditions is expensive and complex. It is also unclear if AI training can be effectively distributed across multiple satellites. Musk likely views current gas-powered data centers as temporary, expecting SpaceX to launch gigawatt-scale servers into orbit within a few years.

The filing warns of “terawatt-scale annual AI compute growth,” far exceeding today’s 40-gigawatt data center usage. This first-principles thinking assumes a massive power shortage on Earth, but the risk remains that Musk’s space-based solution may not materialize as quickly or cheaply as he hopes.

The National Transportation Safety Board temporarily closed its public docket system after discovering that AI tools were used to recreate the voices of pilots who died in a UPS cargo plane crash. The incident highlights growing concerns about how artificial intelligence can manipulate sensitive data from official investigations.

Federal law prohibits the NTSB from including cockpit voice recordings in its public dockets. However, the accident file for UPS Flight 2976 contained a spectrogram image—a visual representation of sound frequencies. YouTuber Scott Manley pointed out that this image could be reverse-engineered into audio, and individuals soon used AI platforms like Codex to reconstruct the cockpit recording.

The NTSB restored public access to its docket system on Friday but kept 42 investigations closed for review, including the UPS crash case. The agency stated that the recreated audio was circulating online, violating privacy protections for deceased crew members and their families.

This event marks a new frontier in AI ethics, where technology can resurrect voices from the dead using publicly available data. Regulators now face the challenge of balancing transparency with protections against digital exploitation in accident investigations.