We may be sleepwalking into an AI-fueled culture shock; what's next for AI and math; how AI and copyright turned into a political nightmare for UK Labour; Perplexity's CEO hypes up agents

How Morgan Stanley built an AI tool to modernize its legacy code; Alibaba emerges as a winner in Chinese AI; Google's Veo 3 can make very convincing deepfakes; UK civil servants see the ROI of AI

In an article for The Guardian this week, readers’ editor Elisabeth Ribbans—whom I respect for how promptly and openly she responds to feedback, including mine—described a recent complaint she’d received about a word. This particular word wasn’t connected to a political scandal or a hot-button social issue. Instead, it had to do with the North American spelling for the past participle of get, gotten. For British readers, that single syllable carried the unmistakable scent of imported Americana. Other complaints followed: faucet, airplane, hot flash. (For the record, I would rather drop dead than use the word bonnet to describe the hood of a car.) But back to the topic at hand: these aren’t just linguistic quirks. They’re signs of a cultural drift that readers can feel, even when they can’t always articulate why it bothers them.

And it should bother them.

Because language isn’t just a neutral vessel for communication, it is culture in motion. It holds identity, geography, history, humor, and pain. It’s the difference between a cuppa and a soda. Between the King’s English and California casual. So when AI models are trained mostly on American English, you’re not just training them to say elevator instead of lift, you’re embedding (quite literally) an entire worldview.

The complaints from British readers of The Guardian are a small but powerful signal of what happens when culture starts getting steamrolled by scale. As Elisabeth explained, The Guardian has grown over the last two decades, reaching a global audience that has pushed it to use more US English, particularly in content from its American desk. But that trade-off comes at a cost: UK readers are now staring down articles written in a voice that sounds increasingly foreign. If that’s happening to a respected 204-year-old British publication, imagine what’s happening inside AI models trained mostly on American data.

This is exactly why we need to stop thinking about low-resource languages and start thinking about low-resource cultures.

Some of the most interesting examples of low resource cultures come from the Arabic world. Take Nile-Chat, an Egyptian Arabic AI model—I believe it is a masterclass in why cultural context matters more than token counts. Built by the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Nile-Chat is fluent in Masry (the Egyptian dialect of Arabic) and Arabizi, the Latinized script Egyptians use on TikTok, Instagram, and WhatsApp. It understands local slang, humor, and nuance. Ask a generic model what drink to serve guests in Cairo and it might suggest Coca-Cola. Ask Nile-Chat and you’ll get karkadeh or a proper glass of tea because it knows hospitality here is about more than carbonation.

Nile-Chat includes a retrieval process for augmenting local cultural knowledge by parsing web pages representing specific local cultural concepts through web search.

Most language models, especially those trained on English internet content, would flounder with phrases like zy elfoll. Nile-Chat knows it means perfect. A generic model might translate it as like a flower, completely missing the social context. And that’s the crux: without exposure to how people actually talk, especially in informal or hybrid registers like Arabizi, models flatten local realities into generic outputs.

Which brings me back to the copyright debate and the Make It Fair campaign that The Guardian and other media organizations support.

The argument over whether copyrighted British content should be allowed in AI training sets is important. But there’s a danger looming that we don’t talk about enough: if that content is excluded, UK users will be left with AI tools that echo only American idioms, assumptions, and references. That’s not just annoying, it’s cultural erasure. Bit by bit, word by word, British linguistic and cultural norms will get squeezed out of a machine mind designed by Americans to think like Americans.

So yes, protect creators. Defend copyright. But also consider other, deeper issues at stake. AI models are cultural mirrors and if the training data doesn’t include your culture, don’t be surprised when the reflection doesn’t look like you.

In the race to build global AI, we need more than just data diversity. We need cultural depth. Nile-Chat shows what’s possible when you care enough to teach a model not just how people speak, but why they speak that way. If we ignore that lesson, we risk building machines that talk to everyone but truly understand no one.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

• MIT Technology Review: What’s next for AI and math

• Politico: How AI and copyright turned into a political nightmare for Labour

• Bloomberg: Chasing Big Money With the Health-Care Hustlers Making Deepfake Ads in South Florida

• Wired: Perplexity’s CEO Sees AI Agents as the Next Web Battleground

• WSJ: How Morgan Stanley Tackled One of Coding’s Toughest Problems

• Wired: How the Loudest Voices in AI Went From ‘Regulate Us’ to ‘Unleash Us’

• Time: Google’s New AI Tool Generates Convincing Deepfakes of Riots, Conflict, and Election Fraud

• FT: UK civil servants who used AI saved two weeks a year, government study finds

• WSJ: AI Is Here for Plumbers and Electricians. Will It Transform Home Services?

• The Information: How Alibaba Helped China Take the Lead From the U.S. in Open-Source AI