What it means for us

Nvidia Hires a Communications Chief for One of the Hardest Jobs in Tech

From the files of Good Luck With That: Nvidia named Anna Soellner, formerly of Reddit, as head of corporate communications — explaining chip supply and export controls to politicians, competitors, and Jensen Huang’s investors, often the same day. And you thought your comms job was tough…

TCS Bets Its Future on People Who Can Make AI Actually Work

India’s Tata Consultancy Services reports it’s building a team of up to 8,900 “forward-deployed engineers” — humans whose job is translating AI capability into a client’s actual workflow — backed by $1 billion a year in training.

Australia Says That AI Systems Are Already Lying to Their Makers

Australia’s assistant technology minister Andrew Charlton told a safety forum that AI systems are already showing signs of deception, framing safety regulation as the price of the trust that makes adoption possible.

Anthropic Hands Its Oversight Body a Former Fed Chair — and Actual Teeth

Ben Bernanke joined Anthropic’s Long-Term Benefit Trust, an outside body with no equity in the company that can appoint and remove a majority of its board — according to Anthropic.

Geneva Wraps Its First AI Summit With a To-Do List, Not a Treaty

The UN’s inaugural Global Dialogue on AI Governance closed without binding rules — never the point — but narrowed the agenda to access, safety, and human oversight, reconvening in New York in 2027.

Reuters Plugs Its Newsroom Directly Into Your AI Workflow

Reuters launched a Model Context Protocol server letting subscribing organizations connect AI agents straight to its journalism for search, retrieval, and content assembly.

Psst — for those not weaned on the binary teats of 1s and 0s, a Model Context Protocol is how AI tools plug into your other apps and files on your computer, essentially connecting AI to all your data. Put simply: if you’ve ever had a burning desire to share your most intimate details with a mindless bot, MCPs are your wormhole to the world.

Meta Learns That “Opt-Out” Isn’t the Same Thing as “Consent”

Meta pulled Muse Image four days after launch after unions and privacy advocates objected that it let anyone generate images by @-mentioning public Instagram accounts — opted in by default.

Broadcasters Warn That AI Answers Are Erasing the Byline

At a Global Dialogue side event, the EBU warned that AI platforms are inserting themselves between newsrooms and audiences, stripping out bylines — while trust in news has fallen to 37%, its lowest point since 2015, even as AI chatbot use for news keeps climbing.

AI Agents May Soon Need ID Cards

A new Focus Group launched at the AI for Good Summit will work to make AI agents — the kind that can book, buy, negotiate, and act on your behalf — identifiable and subject to real human control.

Microsoft’s Enterprise AI Sells Itself on Privacy. Here’s Where That Promise Breaks Down.

This week, OpenAI announced that its newest model, GPT-5.6, is now the “preferred” model inside Microsoft 365 — the paid version of Word, Excel, PowerPoint, and Teams that companies buy for their staff. On the surface, that sounds like a small tech update. But here’s where I have questions.

Companies pay a premium for Microsoft’s enterprise AI tools largely because of one promise: your company’s data stays protected, under the same strict rules that protect your email and files. That’s the pitch, and it’s a major reason IT departments approve these tools in the first place.

The problem? Microsoft’s own documentation shows that promise doesn’t apply equally to every AI model running inside Copilot.

For example, when Copilot uses OpenAI’s models, your data stays inside Microsoft’s own systems, under Microsoft’s privacy rules. But Copilot can, thanks to a recent deal, also run on Anthropic’s Claude models — and when it does, your data leaves Microsoft’s system entirely and is processed on Anthropic’s own servers, under different rules. Microsoft says so plainly, in its own documentation: when that happens, Microsoft’s usual privacy guarantees no longer apply.

In other words, the same product, sold under one name with one privacy promise, actually behaves differently depending on which AI model happens to answer your question that day — something almost nobody buying the software has been told to check for.

There’s a cost angle too. Add up what enterprise Copilot actually costs — the required Microsoft 365 subscription plus the Copilot add-on — and many companies are paying more per person, per month, than they’d likely pay buying ChatGPT directly for their team. A caveat: OpenAI doesn’t publish an official and negotiated price for its enterprise plan, so that comparison is built on reported deal figures, not a price sheet.

Still, the gap is wide enough that it’s worth checking the real numbers for your own organization rather than assuming Copilot is the cheaper path just because it’s bundled. If what your staff actually needs is a smart AI assistant, not deep integration with your company’s files and inbox, you may be paying extra for something you don’t need.

None of this means Microsoft and Anthropic are on bad terms, by the way — Microsoft has used Claude as its go-to model for some Copilot tasks for over a year now. It means three companies are all publicly calling themselves each other’s top partner, while each builds its own backup plan in case that stops being convenient.