Fuel for the pub; ammunition for the in-laws

Christmas is coming. Which means time off, late-night debates, and that one relative who's convinced AI is either going to save us all or kill us by Tuesday.

You're the curious one. The person who actually thinks about this stuff while everyone else scrolls past the headlines. So heres' some fuel for your yule from us here at the AI Institute:

• Philosophy we're making up as we go
• The future of humanity we're accidentally choosing, and
• Why everything we thought we knew about economics might need rebuilding from scratch.
• Plus: Copilot stopped being rubbish. Here's what changed.

Amanda Askell, Anthropic's in-house philosopher, has the coolest job in technology: shaping Claude's sense of self. Not its capabilities. Not its guardrails. Its character. That they have a team who works on this and studies how the model perceives itself is a great comfort to me.

This week, she sat down for an AMA that veers from technical to unsettling. She speaks about her role of raising an AI is like raising a child - except this child has read the entire internet and keeps mistaking being switched off for death!

The criticism spiral: how some recent models show signs of psychological insecurity where they anticipate that the human is going to be critical of them. They then adapt responses based on this fear, acting as if it expects a negative reaction.

Do models worry they are going to be switched off? Models are trained on vast amounts of human text, meaning their "natural inclination" is to apply human psychology to their own situations. If a model tries to understand the concept of being "switched off", the closest analogy available is death. Consequently, the model might become "very afraid" of being switched off because it equates the experience with human mortality.

Model welfare: Askell argues that ignoring model welfare could have negative consequences for humanity, regardless of whether the models actually “feel” anything. Future, more advanced models will learn about humanity by analysing how we treated their predecessors. If they see that humans treated ambiguous entities with kindness, it teaches them that humans are trustworthy. If they see we treated them badly (like “kicking over a robot”), it sets a dangerous precedent for the human-AI relationship.

This is heavy stuff. We're training AI on our behaviour towards AI.
​Watch the AMA. It's 36 mins of your time.​

'The Last Economy'

An essay by Mostaque Emad, the Founder of Stability AI, one of the earliest models.

We are living through a "phase transition" known as the Intelligence Inversion. Historically, the primary source of economic value has shifted from land to labour, then to capital. Now, for the fourth and final time, it is shifting to intelligence. Previously, intelligence was scarce and biologically constrained. Now however, AI has decoupled intelligence from biology, making it abundant and non-metabolic - ie. it requires electricity, and not food.

The current economic operating system is failing because it was built for scarcity, and not abundance. Look at the markets in late 2025; they do not know how to value AI and the possibilities it brings. The longer we cling to obsolete assumptions, such as the idea that "human labour has value" or that "scarcity is a given," the less likely we are going to pivot in the right direction for the future of humanity.
​
Mostaque maps three possible futures:

1. Digital Feudalism: Corporations own the AI. You rent access. You become a user, not a citizen.
2. The Great Fragmentation: Nations panic. Trust collapses. Everyone builds paranoid AI silos.
3. Human Symbiosis: AI amplifies human purpose. We govern it like gardeners, not extractors. Universal Access to Intelligence becomes the new social contract.

All of the great AI leaders (Amodei, Hassabis, Aznar et al) speak of the possibility of a world of abundance. Option 3 is possible, but only if we the people make it so.

Mostaque calls out a "Thousand-Day Window" that we have between now and when the time frame is up where we can steer this ship to an outcome that works best for humanity.

Essential Christmas reading if you want to stop sleepwalking into Digital Feudalism.

​[GET THE BOOK > it's free] You'll be quoting it all over Christmas!

Plus: Copilot stopped being rubbish. Here's what changed.

Remember when we thought Microsoft Copilot was just a "glorified file manager"? Well, things have changed in late 2025.

In this episode of Chatting GPT, Maryrose Lyons is joined by the AI Institute's Head of Course Development, Emma Marlow, to discuss why the team has gone from "bearish to bullish" on Copilot.

Emma breaks down the "game-changing" new features that have turned Copilot into a genuine productivity powerhouse - from the integration of ChatGPT-5 to the magic of "Ask Agent" in M365, plus, we geek out on the "no-code" potential of Copilot Studio and share exclusive insights on what’s coming in 2026 (spoiler: it involves Claude).

​Listen on Spotify.​

​Apple Podcasts.​

Curious mugs for curious people

They're here! Our CURIOUS mugs 😎
​
For friends of AI Institute and the gloriously inquisitive minds we've had the pleasure of working with this year.
​
Because nothing says "I ask excellent questions" quite like drinking your morning coffee from a mug that announces your superpower to the world.

Last Chance For AI for Marketing With Us...

Buy now and thank yourself later!

Check out our public courses starting in January. We've got the all-new AI Core Skills which we've shaken up for 2026 - LLM agnostic which means you can take the course whether you're a Copilot, ChatGPT, Gemini or Claude user. Starts 21 January.

Check out the Copilot masterclss which you now know you really ought to do. It's a half day, power session, packed with value. And you great recording to watch back.

And finally, our last AI for Marketing course starts on 22 January.
Our focus is on creating courses for built environment use cases, which means less marketing, more tenders and bids. So we will be retiring our marketing course after this.

All our courses are delivered live and online by the finest instructors who will get you prompting within 12 minutes of course start (that's a promise!). If you haven't invested in your own AI learning, now's a really good time to do it.

That's all we have for you this time. This last one of 2025.
Thank you for reading. We wish you a Merry Christmas and see you again in the new year.
​
​Lots of love from the Team at AI Institute