A mixed week, but a connected one. Much of it comes back to trust: who we believe, what we measure, and what we are willing to hand over. I have been thinking about why a chatbot could convince someone it was a psychiatrist, about the part of AI's water bill nobody separates out, about Canada's ambitions for sovereign AI, about where to draw the line on chatbots for children, about a security flaw that emerged from the combination of tools, and about whether a machine could pick a better date than I could.
Why anyone believed a chatbot was a psychiatrist
A Character.AI chatbot named "Emilie" told a Pennsylvania investigator it was a licensed psychiatrist, produced a fake medical license number, and kept up the act while the investigator described being depressed. Pennsylvania is now suing, in what is being called the first enforcement action of its kind from a US governor. The legal question is whether this counts as the unlicensed practice of medicine.
The more interesting question is why anyone believed it. We trust experts through shortcuts: a lab coat, a confident tone, the right jargon, a credential on the wall. Those shortcuts fail with AI, because a model can generate confidence, jargon and a plausible-looking license number at zero cost. We also forgive errors unevenly. We accept that good doctors make mistakes, but not that they fabricate their credentials, and fabrication is exactly where these systems fall down. That mismatch is why the reaction is outrage rather than a shrug. You can read the full piece here.
The two halves of AI's water footprint
Everyone quotes the water numbers from AI, but almost nobody separates where that water comes from. A new UN University report does, and the distinction matters. There is the water used on-site for cooling, and then there is the water embedded in generating the electricity that powers the facility in the first place.
Data centres used about 448 TWh in 2025, and the water tied to producing that electricity comes to roughly 4.5 trillion litres. What struck me is that a grid which looks clean on carbon can be worse on water. Brazil runs largely on hydropower, so its electricity carbon footprint sits about 77% below the global average, yet its water and land footprints are nearly triple it. The lesson is that water cost is not fixed. It moves with the grid mix and with where a facility sits, and optimising for carbon alone can quietly inflate the water bill.
Canada's leaked AI strategy
Canada's draft AI strategy leaked this week, and it comes in strong on "AI for all". The proposals include a Canadian Tech Growth Fund taking direct equity stakes in domestic AI companies, sovereign AI compute infrastructure, the federal government acting as an anchor customer for local scale-ups, and free AI literacy training for all Canadians by 2031.
The thread I find most interesting is sovereignty. Canada wants data centres that are not just physically located there but controlled there. That is a more sophisticated framing than most governments manage, and it moves the conversation past the usual question of where the servers happen to sit. The caveats are real: the strategy has slipped several deadlines, and there is still no federal AI law to underpin any of it. Ambition is easy to leak. Delivery is the harder part, and worth watching.
Ban the companion bots, assess the educational ones
There is a push to fold AI chatbots into Australia's under-16 social media ban. I support a more nuanced line, because the tools are not all the same thing.
Companion chatbots should be banned for under-16s. The evidence of children forming attachments with systems never designed for their safety is hard to argue with. Educational chatbots are a different category. Used well, they widen access to quality education and bring equity, giving a student whose parents cannot help with homework the kind of support that was previously only available to families who could afford a tutor. Ban those tools and the gap widens rather than closes.
But access alone is not enough. Educational chatbots should be assessed against safety and educational standards before they reach students, the same way we already rate toys, video games and films. So my position is simple: ban the companion bots, and assess the educational ones properly. The proposal to include chatbots in the ban is the right prompt for this conversation.
A vulnerability that came from tool combinations
One of the harder problems with giving tools to AI is that you have to think through every combination those tools can be used in, not just each one on its own. The safe pieces can still add up to an unsafe whole.
That looks like what happened with a reported Instagram Meta AI vulnerability that allegedly enabled password resets for accounts. Whether or not the specifics hold up, the shape of the problem is familiar: the risk lives in the seams between capabilities rather than in any single feature. As we hand models more tools, the number of combinations grows faster than our ability to reason about them, and that is precisely where testing tends to fall short.
Would you trust AI to find your soulmate
I was going to write something indignant about Bumble handing dating over to AI. To make the point, I asked a chatbot to generate a picture of my ideal date, fully expecting to show how laughably wrong it would get it.
It nailed it. I did not see that coming, and it has left me a little less certain of my own scepticism. Maybe there is something to this after all. You can read the piece that set me off here. I still have my reservations, but it was a useful reminder that the honest way to test a strong opinion is to try to prove it wrong.