What it means for us

G7 floats a “trusted partners” plan to restore allied access to restricted US models

Days after Washington blocked foreign access to Anthropic’s frontier models, the G7 in Évian spent its final session on how to let allies back in. Macron’s argument was blunt: nobody buys an AI model they fear can be switched off overnight. By June 19, the White House had softened, with the American president telling Axios he no longer views Anthropic as a national-security threat — same company, different week. Snide editorial comments aside, the lesson for communicators is pretty clear: this isn’t just about geopolitics; it’s that every executive with an AI strategy now needs contingency language, not just “we take AI seriously” language.

People ask the chatbot, then don’t click (or, why SEO is no longer enough)

The headline number is stark: roughly one in ten people now use AI chatbots to get news, but only about 4% click through to the original source. The answer arrives before the link does. For communicators, that means the old question — “can people find us?” — has become “how are we represented when the machine answers for us?”

Le Monde wonders what to do when a paying reader shows up as a bot

Le Monde has spent months blocking crawlers — and now faces a different problem: the subscriber who sends an AI agent to fetch and summarise the paper on their behalf. Its CTO wants a standard that lets an assistant say, in effect, “I’m acting for a real human who pays you.” This is the “my agent is my reader” era arriving early, and every organisation will soon have to decide how it treats humans, agents, and impersonators in the same channel.

USA Today builds AI “shell files” to beat Google’s summaries to the punch

Reporters pre-build article skeletons — background, links, context — so breaking stories publish in seconds, before Google’s AI Overviews can eat the click. They tested it at the Winter Olympics and are running it through the World Cup now. What does this matter to communicators? I remember these being called “modules”. Nothing new under the sun. However, a good reminder to have everything ready, steady and good to go. Meanwhile, Keith Richard’s obituary gathers dust.

China wants 80% of its AI data-centre power to be green by 2030 — up from 11%

AI data centres run power-hungry GPUs around the clock while renewables come and go, and Chinese demand is set to climb by 300–500 billion kilowatt-hours by 2030. AI may feel weightless on the screen, but it eats electricity like a raccoon in a bakery — and “responsible AI” can no longer stop at model safety. It now includes energy, water, and where the thing is physically built.

Norway draws a hard line: a near-ban on generative AI for ages 6–13

Younger children won’t use generative AI in school; older students will, with supervision and a focus on judgment. The PM framed it around the basics — reading, writing, maths. So, not everywhere, not never, but where it fits. Perhaps an interesting model for organisations too. Some use AI to encourage, some to supervise, some to teach judgment before output.

A forensics expert warns the “is this real?” problem is getting worse

Hany Farid, who does this for a living, says synthetic images, video, and audio are now convincing enough that ordinary people can’t reliably tell — and verification usually arrives long after a post has sunk in its hooks. The threat is not just fake content; it’s the slow erosion of shared reality (nope, didn’t steal that from The Matrix). For communicators, be prepared to show your work.

Retailers ask the EU to exempt AI-made ads from labelling

EuroCommerce — Amazon, H&M, Inditex, Ikea — argues that using AI to drop a sofa into a digital room isn’t the same as a deepfake, and that labelling everything makes disclosure meaningless. They have a point…until a customer asks, “wait, was any of that real?” For communicators perhaps the better test for our work isn’t “did AI touch this?” It’s “would the audience feel misled if they knew how it was made?” I always knew that Aflac duck couldn’t really speak.

The minister and the false positive

A reader flagged this one for me — thank you! — and please keep them coming. It slipped past my deadline last week but it’s too good to leave behind. Die Zeit ran 30 speeches by Germany’s digital minister, Karsten Wildberger, through an AI-detection tool called Pangram: one (delivered to the Atlantic Council in Washington) came back as fully AI-written, seven more had AI-written fragments, and 22 were flagged as human. The ministry shrugged — AI is a “support tool,” no disclosure needed, a human checks everything.

The reason this matters for us is that the question “did a machine write this?” is now going to trail every speech we put our names to. But before anyone panics, look at what these detectors actually do. I handed my monthly AI group an “AI tells” prompt built to catch exactly where a speechwriter had leaned on AI — and it returned many false positives. Enough of them to genuinely piss off (and rightfully so) those who’d written every word themselves. Let’s remember: a flag is not a verdict.

And remember what these models trained on: the open internet — which is to say, our work. Decades of speeches and beautiful comms output, scraped and absorbed. So, when a detector calls your writing “AI-generated,” there’s a real chance it’s reacting to your work, because the machine learned to sound like you. Which leaves us flagging the AI that’s flagging our work as AI — work that was ours to begin with. Nice.

Influencer recommending sofa fingered as fake

The Guardian found brands quietly using AI-generated influencers to fake customer experiences (surely not!) — some creators even signed NDAs to keep quiet about it. One Dubai label’s “satisfied customer” turned up wearing the clothes and — in a detail no script could improve — a glorious extra finger. Nothing says authentic testimonial like a spokesperson who failed "hands" in art class. Do I need to say it? If your audience feels tricked, no efficiency gain survives the cleanup. It’s also just stupid.