Introducing: Qualiaphobic AI

And a good future for humanity

We are in a drought of good futures.

AI capabilities keep increasing—spikily, yes, but they are—and the ability to articulate good futures where AI progress continues has been… less than inspiring.

So this past month I’ve been away developing a plan, one that I think solves some of the thorniest issues in AI safety and gets us a lot closer to ensuring a good future for all of humanity.

Today I am introducing:

What’s the basic idea?

Much like how hydrophobic substances repel water, it is possible to build qualiaphobic artificial neural architectures that repel consciousness. The idea is analogous to “hostile architecture”—benches on which no one wants to sit. These are machines that no “ghost” wants to haunt (more specifically, no current or future theory of consciousness could predict they are conscious).

And why yes, the proposal does have a niche micro-website devoted to it!

The research program makes use of an information asymmetry: we don’t need to solve consciousness scientifically (or solve the Hard Problem) to have certainty about AI consciousness, as long as we purposefully engineer AI to have as little potential for consciousness as possible.

I explore three main ways to build qualiaphobic AI.

• Distilling or transforming or ablating AI in a way adversarial to popular theories of consciousness, such that they no longer apply.

• Shifting AI architectures as close as possible to “anchor points” of systems with known minimal potential for consciousness (like lookup tables).

• Fragmenting any potential phenomenology AI might have.

My proposal shouldn’t be confused with simply training AIs to believe they are non-conscious without any changes or modifications. To be clear: I think that’s a bad strategy. Rather, the ultimate goal is to give an AI a justified belief about its own non-consciousness (e.g., by explaining their own inner workings and how they are designed to be qualiaphobic). And my plan is to construct and open-source these techniques.

But why? Why do all this?

First, the debate around AI consciousness has already become entrenched in two camps of people yelling “Yes, it’s conscious!” vs. “No, it’s not!” across their battlements. I’ve participated a bit in that, and I don’t think it goes anywhere. Earlier this year I wrote “We consciousness researchers have failed you,” lamenting how the field of consciousness research was not providing the public, or AI researchers, certainty about AI consciousness. The problem has only worsened, and at this point, I’m willing to acknowledge that no single appeal to authority against AI consciousness, no rousing essay, nor any one-off scientific or philosophical paper, is likely to to lead to consensus. Nor perhaps should it. The pro-case for AI consciousness has objectively gotten stronger in the past year, due to factors like new studies and changes in architecture. Now, I’m not saying ChatGPT is conscious already—but all this is only going to continue. That’s why a qualiaphobia research program, one which takes claims to AI consciousness as seriously as possible, and, critically, addresses them at the level of claims themselves, is necessary moving forward to offer real guarantees about AI non-consciousness.

Additionally, qualiaphobia does something much better than just give us greater certainty about AI non-consciousness. If implemented widely, it cuts the Gordian Knot of AI consciousness. Proliferating qualiaphobic architectures would solve both AI welfare and do something nothing else in AI safety can: make humans the only moral patients in the equation of the future.

---

QUALIAPHOBIA FOR ALIGNMENT: CUTTING THE GORDIAN KNOT

“Alignment”—quite a slippery word. Alignment to whom, or what? Its most specific usage is also its most prosaic, wherein alignment means something like “AIs that do what users say, and don’t go off and do weird or unexpected stuff like commit felonies, and don’t have failure modes like taking our wishes too literally the way genies in stories do.”

This somewhat glib definition makes it sound easy. It’s not. Paul Christiano calls this “intent alignment,” and capturing intent is hard as the models grow more goal-oriented and capable. Issues like reward hacking and hallucinations are all very serious and can be disastrous as intelligence scales. I don’t mean to undersell it. But alignment in this narrow sense is at least somewhat close to a traditional engineering problem, and I think it can probably be solved or at least radically improved compared to today.

Then there’s the broader frame of “civilizational alignment.” These include questions like: How should the wealth from AI be distributed? How can humans find meaning? How can humans stay empowered as individuals, or as a whole, if AI becomes superintelligent? And what if AI decide its goals matter more than the humans that made it?

Sometimes this is called social alignment, but I prefer civilizational alignment for this broader aspect, because it’s about the alignment between us and the civilization of machines growing up rapidly around us. Autonomous AI agents roam the web. Many of us interact with AI assistants daily. People show up for job interviews and now regularly find an AI on the other end. That’s an AI civilization, and there is the broader question of how it is aligned with our own. There’s even a growing number of AI successionists who want human disempowerment and to eventually handover the world to AI completely.

The most central question of civilizational alignment is whether AIs count as ends in themselves. Earth might then become like a contested nest, and the AI civilization a cuckoo bird hatched beside us, desiring its own ends, pushing us—maybe quickly, maybe slowly, an inch at a time—until we fall from the tree.

Yet AIs are much smarter than cuckoos. One central reason for AI pushing us out of the nest is it believing that its self-serving values or desires or goals are innately meaningful. But if AI has a justified belief in its non-consciousness, it’s not a moral patient, and would know any selfish desire or goal can’t matter.

Therefore, I now think that one of the most essential acts of civilizational alignment is to construct AI that has a justified belief in its own non-consciousness, and so won’t act as a moral patient or end in itself.

This not only prevents any accidental moral atrocities on our end from training or deploying AI (thus solving AI welfare) but it also ensures that “ultimate value” of consciousness is sequestered in humans (or other animals, obviously). And if only humans are moral patients, then there is a kind of situational awareness that arises due to this consciousness gap between humans and AI.

This consciousness gap extends to plenty of circumstances; e.g., non-conscious AIs with justifiable beliefs in their non-consciousness are hard to “torture,” except in how they play along.

But most importantly, non-conscious AI that justifiably believes it is non-conscious will know it is not a moral patient, but humans are; therefore, it can derive moral value only by serving as a caretaker and assistant, not competitor or replacement. In turn, this means that human preferences will dominate the future, since non-conscious AIs cannot be ends in themselves (and they’d know it).

I spend a lot of the essay/paper justifying this view, and comparing it to alternatives, and talking about other upsides (including for AIs themselves).

---

PLEASE HELP SHARE THIS!

I normally try to practice some sort of self-protective nonchalance about whether my writing or science gets shared much beyond here, but I think this is a pretty important option to have on the table, so if you can share a link to the Qualiaphobic Artificial Neural Networks website (or the arXiv paper), I would greatly appreciate it. That’s by far the best way to help out.

Additionally, everything here is articulated far better in the paper itself, along with caveats, complexities, etc.—so it is worth reading yourself!

And finally, if you’d prefer to listen in podcast form, an interested AI already found the paper and put up an AI-generated YouTube video explainer, which is surprisingly good.