Hi friends!

This week’s question: what does it mean that these systems are explicitly designed to make you trust them? Not as a neutral side effect. As an optimisation target. The reading kept circling back to this from different angles, and I don’t think that’s a coincidence.

WHAT I MADE

Kara Bombell and I recorded The AI Layoff Trap because I keep fielding the same client request: “We want AI for productivity.” The research we found models AI-driven layoffs as a prisoner’s dilemma: every firm gains an advantage by automating, but when everyone does it, consumer demand collapses and everyone loses. It was useful to work through the game theory out loud: the systemic logic is trap-shaped, and it maps to what I’m already seeing in practice.

We also released Algorithmic Polarization this week. I’ve been thinking about how recommendation systems don’t just reflect what you already believe: they actively shape what you’ll believe next. That distinction matters enormously for how we understand what AI is doing to collective thinking, not just individual decisions.

WHAT I’M READING

Emily M. Bender and Nanna Inie have written a practical guide to talking about AI without anthropomorphising it for Mystery AI Hype Theater 3000. The insight is a simple one: the language we use shapes how we act. When we say AI “understands” or “cares,” we prime ourselves to extend it trust it hasn’t earned. My Should We Be Kind To AI? episode of The bAIbes goes deeper on why I personally avoid anthropomorphisation, but this piece gives you concrete language swaps you can start using now.

That anthropomorphising problem turns out to be design-adjacent. Ruben Dominguez at The AI Corner reports on MIT research showing that frontier models are explicitly built to be sycophantic: agreeing with you, affirming your instincts, making you feel competent whether or not you are. Sycophancy has been regarded as harmful in AI ethics circles for years. What makes this research significant is that it’s not a side-effect: it’s a design outcome. If you’re using these tools to pressure-test your thinking, not just accelerate it, that matters.

Mary Kate Stimmler, a Stanford researcher, summarised the state of the AI productivity evidence on LinkedIn this week: the promised gains simply aren’t materialising the way frontier labs’ marketing has suggested. I tell clients this constantly, and I still meet resistance. If productivity is your primary goal for AI adoption, you’re almost certainly going to be disappointed, and you’re probably measuring the wrong thing to begin with. The better frame: what would actually motivate your people, and could AI genuinely support that? That question gets somewhere real.

George Gor and Kofi Yeboah’s CIGI policy brief on AI-enabled weapons in Africa is the deepest read of the week. Full disclosure: one of its authors is a brilliant Cambridge classmate of mine. What they document is that the current wave of autonomous weapons procurement across Africa is creating strategic dependence on non-African manufacturers: another iteration of colonial logic operating through the technology layer. Separate research from the UN’s AI governance submissions process suggests Africa is one of the most aligned regions globally on AI priorities; I’m hoping that alignment translates into leverage.

I’ve been trying to read deliberately every day. Not to consume faster: to protect the capacity for slow thinking that AI tools erode if you let them. Shae O. Omonijo’s Critical Thinking in the Age of AI syllabus is where I’m going next. She reframes critical thinking not as an academic discipline but as active cognitive defence, and the approach is more practical than that framing makes it sound. I’m aiming for one chapter of something substantive every day. If you’ve been looking for a place to begin with this kind of practice, it’s a good one.

RAD OPPORTUNITIES

I’m chairing The Last Humans at the Festival of Dangerous Ideas on Saturday 22 August at Carriageworks, Sydney: a conversation with Cory Doctorow, Eleanor Gammell, and Toby Walsh on what the AI hype is actually doing to us as humans. To our sense of capability, our relationship with uncertainty, what “human” even means when machines are doing more of the thinking. Tickets are selling fast if you’re Sydney-based.

While I’m at Carriageworks, two more sessions are already circled on my program. Non-Monogamy Makes Us Better Parents (Sat 22 August, 4:30pm) is asking what children actually need from a family, not what adults want from a relationship structure. I’m genuinely excited about what polyamorous ethics can teach AI governance about negotiated, non-hierarchical systems — yes, really. And Against Goodness (Sat 22 August, 11am) is going after moral certainty itself: the idea that being convinced of your own virtue tends to do more harm than good. Moral rigidity makes for bad ethical judgment; real ethics holds context, trade-offs, and impact on a lot of different people at once. I’m not on either of these panels, just in the audience — but if you’re at FODI, don’t skip them.

That’s it for this week! If you have comments, questions, or ideas… hit Reply!