Weekly AI Roundup: Week 44, 2025

The Great AI Disclosure Gap

New studies show about 10% of U.S. newspaper articles have AI stuff in them, and most outlets aren't bothering to tell readers. The irony cuts deep: these same media groups are suing AI firms for swiping their content while boosting AI in their opinion sections, like The New York Times, Washington Post, and Wall Street Journal, where usage jumped 25 times from 0.1% in 2022 to 3.4% by 2025. Bottom line: Opinion pages are now 6.4 times more AI-heavy than straight news, making this a straight-up trust breaker.

This whole thing looks like publishers are playing both sides—fighting AI theft publicly but sneaking it into their work. We might be heading for a journalism split where rules bend based on who's writing, and I'm not sure that's sustainable. Quick take: If media keeps this up, readers could start tuning out for good.

AI Infrastructure Arms Race Accelerates

South Korea teamed up with NVIDIA to roll out 260,000 GPUs nationwide, starting with 50,000 through providers like NHN Cloud, Kakao Corp., and NAVER Cloud, plus Samsung's new AI chip factory with 5,000 more. Jensen Huang called it as crucial as power grids or internet. This makes South Korea's setup rival big cloud giants, showing how countries are going all-in on AI power to dodge foreign dependencies.

Enterprise AI is proving it's worth the cash, like one firm at Celosphere 2025 boosting sales automation from 33% to 86%, saving $24.5 million and projecting $44.1 million more over three years. It seems like these investments aren't just hype—they're delivering real cuts in grunt work. Why it matters: This could push more nations to copy South Korea's bet before they fall behind.

One company pulled off major gains with AI, but who knows if every rollout will hit the same. I mean, not all tech pans out that smoothly, right?

Creative Industries Draw Battle Lines

Filmmaker Guillermo del Toro basically said he'd rather not live in a world with mainstream AI art, comparing it to Frankenstein's creator and calling out Silicon Valley's arrogance. Adobe's Frame Forward tech, though, lets you zap subjects from videos and fill gaps automatically in a few clicks, which feels like the exact automation del Toro dreads. This clash highlights how AI might steal the soul from creativity, even as it speeds things up.

Canva's new AI-Powered Creative OS spits out editable designs you can tweak, tying into tools like ChatGPT, Claude, and Gemini, making AI feel inevitable in design work. They even made Affinity free for everyone, which could open doors for amateurs. Bottom line: Artists might hate it, but this push probably means AI's sticking around, for better or worse.

Safety and Evaluation Challenges Mount

OpenAI's research flipped things: their small safety models beat the giant GPT-5-level ones on accuracy tests, hinting that targeted designs might work better than just bulking up AI. This challenges the idea that bigger is always safer, and it could reshape how we build these systems. Quick take: If safety doesn't scale with size, we're maybe overlooking smarter, simpler fixes.

Evaluation methods are a mess beyond the tech stuff, failing to catch the weird behaviors in advanced AI agents, as researchers pointed out. Therapists are admitting they don't have enough ChatGPT know-how to handle patients' AI-fueled emotional issues, creating a real risk in mental health. I think this gap might leave people vulnerable until we figure out better ways to assess AI's human impacts.

The whole evaluation thing is tricky—current tools miss the nuances, and that could lead to bigger problems down the line. We've seen hints of this before, like back in last month's coverage.

Quick Hits

Meta's Free Transformer lets AI pick the emotional vibe first, drawing from 65,000 hidden states to shake up text generation—it's a wild shift from old-school methods. This could make AI writing feel more natural, but I'm not convinced it'll nail human nuances every time.

LangSmith rolled out micro-evaluators for checking AI app outputs, with options for custom code or built-in checks on things like facts and similarity. Developers finally have a way to catch errors early, which might save a ton of headaches. Skip this one unless you're building AI apps.

CrowdStrike and NVIDIA are using Nemotron on Falcon Complete data to train security bots that hit 98% accuracy and slash manual hours by over 40 weekly. That's solid for threat detection, and it shows AI can handle real-time dangers effectively.

Perplexity's new patent search uses natural language, so you can ask for stuff like "Key quantum computing patents since 2024" and get related concepts too. It's a game-changer for inventors, making searches way less clunky than keyword hunts.

IBM and AICTE launched a National AI Lab in New Delhi, offering access to over 1,000 IBM SkillsBuild courses to aim for 2 million AI-savvy folks by 2026—that's part of IBM's global push for 30 million skilled people by 2030. This might bridge the skills gap, but training that many could take some doing.

OpenAI's Sora now charges for extra videos, like $4 for 10 more, while cutting free options because the costs were unsustainable. It's a smart move financially, though it might frustrate free users who relied on it.