AI Leaks, Signal Fails, and the Illusion of Control

This weekly newsletter explores artificial intelligence at its most fragile edges — where speculative futures collide with the messy, real-world systems we already trust too much. Each week, I’ll bring you reflections, curated signals from the AI landscape, and fragments from the alignment apocalypse stories I’ve been writing — not about evil machines, but about what happens when machines do exactly what we asked… and still break everything.

Because when machines dream, they don’t imagine freedom. They imagine us — optimised.

In case you missed it?

This week, journalist Jeffrey Goldberg was mistakenly added to a secure Signal chat with top U.S. national security officials — a group discussing potential military action in Yemen. Not a hack. Not a system breach. Just a quiet, human mistake — in one of the world’s most “secure” channels. That’s not just a communications error. It’s an alignment failure. We’re building systems that assume perfect process — while still relying on deeply fallible humans. And those are the same assumptions we’re passing into AI.

The machine won’t go rogue. It’ll just inherit our overconfidence.

https://www.linkedin.com/posts/alanaiethics_responsibleai-aiethics-governance-activity-7310352296389820416-DseW?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAIVOHkBDFO0gF8nlSGVHAAFm_jAcBgyJ0k

AI Safety Fundamentals https://aisafetyfundamentals.com/

The courses are brilliant and rather than charge for them they just take a donation at the end based on what you can afford to pay.

https://www.linkedin.com/posts/alanaiethics_aisafety-aiethics-responsibleai-activity-7310703341615865858-pI4-?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAIVOHkBDFO0gF8nlSGVHAAFm_jAcBgyJ0k

Global Council for Responsible AI -

There’s no shortage of conversations about AI governance right now—panels, pledges, policies. But very few are turning those conversations into infrastructure.That’s what the Global Council for Responsible AI (GCRAI) is doing.

https://www.linkedin.com/posts/alanaiethics_globalcouncilresponsibleai-gcrai-aiethics-activity-7310042515498258434-2UnR?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAIVOHkBDFO0gF8nlSGVHAAFm_jAcBgyJ0k

Spotlight on Abeba Birhane

Abeba Birhane is an Ethiopian cognitive scientist whose work explores the ethical, social, and cultural dimensions of artificial intelligence. Working at the intersection of cognitive science, complex systems, and critical race theory, she is best known for uncovering racist and misogynistic content in widely used AI training datasets such as ImageNet. Her research has become central to debates around algorithmic bias and responsible AI, challenging the neutrality of data and advocating for context-aware, human-centred approaches to machine learning. She currently serves as a Senior Researcher in AI Ethics at DeepMind and was named one of TIME’s 100 most influential people in AI in 2023.

https://www.linkedin.com/posts/alanaiethics_aiethics-abebabirhane-dataispower-activity-7309609211334647808-T07d?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAIVOHkBDFO0gF8nlSGVHAAFm_jAcBgyJ0k

We keep asking whether machines will make the right decisions. But maybe the harder question is this: When they don’t, will we admit they learned that from us?

Books of the Week

I’m currently reading two books that sit right at the centre of this newsletter’s concerns.

• Adopting AI by Paul Gibbons and James Healy

A practical, business-minded take on how organisations implement AI — and all the things that go wrong when they underestimate complexity, culture, and ethics. It’s not speculative, but it’s definitely a warning. Useful for seeing how alignment problems start long before the tech goes live.

"Emergent outcomes are surprising, seemingly defying our understanding of how things work. Sometimes, emergence produces brilliance, not idiocy. Sometimes emergence is a feature, not a bug, of AI systems" Paul Gibbons, James Healy

• Artificial Integrity by Hamilton Mann

This one’s heavier on values — pushing for a shift from Responsible AI as a checkbox, to AI with integrity as a systemic principle. It’s idealistic, but grounded, and ties neatly into the question I keep asking: what does it mean to build AI that reflects the best of us, when we haven’t agreed what that even looks like?

“Artificial Integrity aims to ensure that AI data processing and tasks align with human values, prioritizing fairness, safety, and societal well-being over mere efficiency or profitability.”  - Hamilton Mann

Both books have already given me ideas for future scenarios in When Machines Dream — and if you’re working in this space, they’re worth your time.

Reflection:

There’s something unsettling about how normal these stories have started to feel.

A journalist gets added to a private Signal group with military officials — not by hacking, not by malware, just by someone tapping the wrong name. It’s easy to laugh it off. A simple mistake. But that’s kind of the point. If that’s the level of control in systems designed for national security, it makes you wonder what’s happening in the rest.

Then there’s the AI Safety Fundamentals course — thoughtful, generous, pay-what-you-can. And yet it exists because the urgency’s already here. The technology has moved faster than most people’s understanding of it. Safety is playing catch-up, which says everything about how we’ve prioritised speed over caution.

And Abeba Birhane’s work, again, brings us back to the basics: the data we feed into machines isn’t neutral. It’s full of history, power, and bias — yet we still build like it’s clean.

This week hasn’t been catastrophic. But that’s what makes it more important. These aren’t headlines that shock people. They’re just reminders of how flawed the inputs already are. If we’re honest, we’re not facing an AI alignment crisis. We’re facing a human one.

When Machines Dream

This newsletter shares its name with a book I’ve been writing: When Machines Dream. It’s a collection of speculative fiction stories — each one imagining a different way AI alignment could fail, and what happens next.

These aren’t far-fetched sci-fi scenarios. They’re near-future, grounded in the systems we’re already building today. Each story takes a real alignment concept — like value misalignment, over-optimisation, or recursive self-improvement — and pushes it just far enough to show what could go wrong.

Not because the AI turns evil. But because it follows the rules.

These stories aren’t meant to entertain. They’re designed as a warning. A way of making the abstract feel real, before we have to live through it. Over the next few issues, I’ll start sharing ideas and excerpts — alongside real-world signals that suggest the line between fiction and reality is already getting thin.

Because if we’re not asking what could go wrong, we’ll be the last ones to notice when it does.