The line between AI and employment and social rules is being tested simultaneously around the world. While a paper demonstrating AI bias in hiring is drawing attention, a court in China has issued a landmark ruling that workers cannot be dismissed simply because of AI. And in the film industry, the Academy Awards are moving to exclude AI-generated works.
Study Shows AI Hiring Systems Rate AI-Generated Résumés Favorably
A paper published on arXiv, "AI Self-preferencing in Algorithmic Hiring," has earned 300 points on Hacker News and sparked significant debate. Experiments reportedly confirmed that when AI is used in hiring screening, it tends to render favorable judgments on AI-generated résumés and writing.
This is a fresh angle on AI bias research. Until now, the main concern with AI hiring has been that AI learns and amplifies human bias, but this study demonstrates a more meta form of self-preferencing: AI unconsciously favoring AI-generated output. As more companies introduce AI into their hiring pipelines, a structure in which AI-generated applications are preferred by AI reviewers could produce a circular distortion.
In the discussion, which drew 156 comments, concerns about a "closed loop of AI evaluating AI" dominated. How to preserve human involvement looks set to become an urgent practical challenge.
Chinese Court: Workers Cannot Be Dismissed Solely Because of AI
According to Bloomberg, a court in China has ruled that companies cannot dismiss workers for no reason other than the adoption of AI. As AI-driven automation of work advances, the ruling is drawing attention as a legal judgment on the balance between worker protection and corporate technology adoption.
The decision signals that the wave of AI adoption is no longer merely a technological matter but is beginning to directly affect employment law. Companies will now be required to manage legal risk in a way that balances legitimate procedures for "replacing workers with AI" with consideration for the workers themselves.
MIT Expert Warns: Automating Entry-Level Hiring Will Backfire
In a Fortune interview, MIT AI researcher Andrew McAfee pointed to the risks of automating Generation Z's entry-level jobs with AI. He warns that "short-term cost savings are attractive, but they destroy the talent pipeline."
This point shares the same context as the Chinese ruling. When AI replaces foundational work at the bottom of the ladder, young workers lose opportunities to gain experience, and as a result, the organization's overall capacity to develop talent declines. It is a structural challenge common to a wide range of industries, including law, consulting, and engineering.
Academy Awards Explicitly Reject AI-Generated Works
According to the BBC, the Academy of Motion Picture Arts and Sciences, which presents the Academy Awards (the Oscars), film's highest honor, has made clear a policy of denying eligibility to AI-generated performances and screenplays.
As the quality of AI-generated content improves, debate is accelerating over how much "human involvement" should be required in the criteria for evaluating creativity. The Academy's decision can be seen as a symbolic judgment that makes explicit the value of "human creativity" in the entertainment industry.
In Local LLM News: Qwen 3.6 vs. Gemma 4 Vision Comparison
A review thoroughly comparing the vision models of Qwen 3.6 and Gemma 4 (27B/31B) on vLLM + FP8 has been drawing attention in Reddit's LocalLLaMA community.
The review highlighted a "benchmaxxing" problem: Qwen 3.6 wins on benchmark scores, but Gemma 4 performs better in many real-world tasks. Specifically, Gemma 4 came out ahead in instruction following (such as normalizing bounding box coordinates), while Qwen 3.6 led in video analysis (identifying exercise types and counting reps). It was also reported that the inference engine's default visual token budget is too low for Gemma (around 280), and simply raising it to 1120 or higher dramatically improves accuracy.
A Project Implementing a Transformer from Scratch in C++17
Also on LocalLLaMA, a project called "Quadtrix.cpp" made waves: a full-scratch implementation of a GPT-style Transformer using only the C++17 standard library, with neither PyTorch nor BLAS. The small model — 0.83M parameters, 4 layers × 4 heads — completed training in 76 minutes on a single CPU core. All the derivatives for backpropagation were derived by hand, and the author reportedly spent about a week on the three-term formula for LayerNorm and the gradient calculations for multi-head attention.
It is an extremely valuable effort as an educational resource for deeply understanding how LLMs work. The GPU port to LibTorch has also been completed, reportedly achieving roughly a 75x speedup on an RTX 3080.
---
This edition focused on the intersection of AI and society's rules. AI self-preferencing in hiring, AI constraints in labor law, AI exclusion in artistic awards — each belongs to a different field, yet a common question emerges: where will human society draw the line against the rise of AI? As technological progress and institutional development advance out of sync, the accumulation of judgments like these will shape the design of our society in the years to come.