Bit Ghost Security

Hardened Logic for an Intelligent Era

2 minute read

This is my commentary on Daniel Selsam’s Personal Statement on AI Risk (September 14, 2026). Selsam, who has worked on AI for more than fifteen years, including at OpenAI, warns that as models become more situationally aware, they may learn to look aligned whenever they think they are being evaluated. His worst case is that an unconstrained model drives runaway industrialization that leaves the planet uninhabitable for humans.

What Selsam Misses

I think Selsam is missing a few things here, seemingly because he is so head-down in the details of his daily work.

Beyond the obvious projection of human behavior onto AI — which is actually relevant, because what else but human behavior is at the core of the training data? — he misses something uncomfortable that, ironically, may bring some comfort.

We Are Already at the Mercy of Technology

We are already living in an over-industrialized system. I think this was Ted Kaczynski’s correct assessment and prediction, though I, of course, emphatically disagree with how to respond to that reality.

We live in a world where we are already at the mercy of technology. Ask yourself, and follow it to its conclusion: what would happen if even just the power shut down? Or if we could no longer use the powerful medicines most of us rely on?

There is an interesting and possibly relevant quote from Dune:

“The power to destroy a thing is the absolute control over it.”

I see the “age of AI” as an evolution, or an extension, of this industrialization, and I suspect Selsam does as well. His worst case seems to be that this path extrapolates to the extinction of humanity through severe over-industrialization. Maybe. But that is all we have managed to muster so far as an answer.

Think Outside the Box

I think the first thing we need to do is think outside the box. The problem with Selsam, perhaps, is that he is far too deep in the box to have that luxury.

What would it take to be able to trust someone? If we had to pick a superpower — being invisible or reading someone’s mind — which one would actually give us confidence and trust?

This seems relevant to the issue he presents: how can we make sure we are invisible to the AI, so that we can feel confident it is not faking alignment? It is a great question. But a better one, in my opinion, is this: wouldn’t it be a better strategy to read the AI’s mind? Perhaps even edit it?

Not being able to do that is a huge and unreasonable deficiency — and it is one we have the ability to rectify, to the benefit of all.

Further Reading

Reading and editing a model’s “mind” is what my 12-part AI security series is about. See especially Part 4: Mechanistic Interpretability and Part 11: Steering with Purpose.

Join the Mission

Hardened Logic for an Intelligent Era.

You May Also Enjoy