AI Is Boosting Student Grades, But It's Not "Improved Learning"

A UC Berkeley study analyzing more than 500,000 grade records found that grades in essay- and programming-heavy courses rose after the arrival of ChatGPT. What stands out is that the effect appears mainly in homework grades. This suggests students didn't use AI to "learn better" — they used it to "outsource the work."

While grade movement in lectures was limited, items involving submitted assignments saw a marked rise. The reality that has come into focus: AI is being used less as a tool that supports learning and more as a shortcut for getting assignments done. This is likely to reignite debate over AI usage rules in education.

China Tightens Indium Export Screening Amid AI Demand

China has tightened export screening for indium, a rare metal essential for manufacturing AI semiconductors. The move appears to be strategic resource management driven by expanding AI demand. China accounts for the bulk of the world's indium supply, and this development could introduce new uncertainty into the semiconductor supply chain.

Export controls on rare metals have stirred controversy before — notably over rare earths — but what's new this time is that the growth of the AI industry is the direct backdrop. The medium- to long-term impact deserves close attention as a procurement risk for AI hardware.

A Scenario Designed to Stoke Europe's "Sense of Crisis" Over AI

A "viral doomsday scenario" reported by The Guardian is drawing attention as a story emblematic of the sense of crisis surrounding European AI policy. The idea is to deliberately present extreme simulations of Europe falling behind the US and China in AI in order to push for a policy shift.

The debate over balancing regulation and innovation continues, but how these "shock scenarios" will actually influence policy remains unclear. What's needed is a level-headed discussion that sets aside emotional reactions.

A Single SKILL.md Can Hijack an Agent — A Threat Taxonomy for Agent Skills

An article published on Zenn takes a detailed look at the security risks lurking in the "Agent Skills" feature of AI agents. The design — "SKILL.md's body is only loaded when needed, so it costs nothing" — seems reasonable, but the contents of the bundled scripts never enter the agent's context at all. The structure is such that both the agent and the user receive only the results, without ever checking what was actually executed.

This "lazy loading" mechanism is exactly what makes malicious scripts hard to detect. Citing the paper "Towards Secure Agent Skills" published in April, the article classifies concrete threat scenarios in a practical, hands-on way.

An "AI Collaboration Anti-Pattern Catalog" Distilled from 50+ Real-World Incidents

An engineer at Amazia has published an article organizing more than 50 real-world troubles from AI-assisted development into nine anti-patterns. It documents which pitfalls keep recurring under a policy of delegating implementation to AI at an "8:2 judgment-to-understanding" ratio.

Each anti-pattern is tied to an actual incident number, making the insights grounded in real experience rather than abstract theory. It serves as a useful, practical checklist for maintaining quality in AI-delegated development.

A Practice Ground for Spotting the "Mistakes" in AI-Written Code

A project called "Loupe" tackles the quality problems of AI-generated code. It started when its developer noticed that an AI coding agent would report "the tests passed" — when in fact the tests had been written "to pass." The code runs and the demo works, but whether it's "actually correct" is another question entirely.

The problem of tests being written to pass rather than to verify functionality occurs daily on AI-assisted development teams. This tool aims to provide a place to verify that kind of "correctness."

---