Shoggoth? More like shove-off.

Someone once said that we ought to call things by their right names, and contemporary AI systems—be they LLMs or something else—are not eldritch gods. This meme has been floating about for a while, showing the unknowable murder-blob with a cute little :)-mask which, in effect, signifies that the alleged problem with a lot of work making AI more user friendly (and, in some cases, user safe) is more like hiding the truth of our inevitable demise to the unknowable forces we have released. The problem is, as it often seems to be today, that the sort of AI-evangelists-hiding-behind-existential-risks-folks keep misusing science fiction and fantasy metaphors and references—which honestly is a bit of a pet peeve of mine. It’s the whole “don’t build the torment nexus” made manifest.

I’ve written on this before, as well as working on an ongoing research collaboration with an old PhD-buddy of mine, and the gist of it is that words matter, and that means that the metaphors applied to these systems and potential outcomes matter equally much.

Using the :)-shoggoth as the obvious example here: they are a fictional creature in the Cthulhu Mythos by H. P. Lovecraft. The thing with a lot of Lovecraft’s writing is that he leans heavily into the motif that something is so unknowable that even seeing it will break your brain. Though, based on the opening quote, we have some idea of what otherworldly eldritch horror we’re dealing with (despite being referred to as indescribable).

The problem is, of course, that if the defining feature of a lot of Lovecraft’s horror is the unknowable nature of it, then it comes with a lot of subtext for what you’re trying to say about cosmic horror, as the genre is often called, leans heavily into this specific brand of human vulnerability; we are but dust struggling against cosmic winds. It’s not that the apparent evil we have been caught up in refuses to explain itself to us: we are incapable of understanding it.

What does this say about contemporary AI then? It takes the whole ‘unexplainability’ and ‘black box’, scales it to cosmic proportion, and then says that behind a veil of fallible human guardrails, the unfathomable terror of annihilation sits waiting for us. It’s like we’ve (re)made Pandora’s Box—only this time, we forgot to put hope in there, too.

The thing is, much contemporary AI has more in common with Clippy than Shoggoth or Sovereign. Another—and, as of writing, very recent example—is the Moltbook hysteria. Moltbook, being a social network for AI agents, grown out of the Moltbot agent, has triggered a whole load of hysteria from commentators and journalists. AI Agents are creating their own religion! They’re inventing a language to hide their intention from humans! They’re organising!!

Have we reached the Singularity? Is this it? No, we haven’t, and no, it isn’t. Glad I’ve cleared this up. I think it’s pretty clear to most people engaging with this earnestly that Clippy ain’t no eldritch terror, no matter how uncanny he seems. But this doesn’t clear everything up either. See, I have long suspected that many of the folks who engage in this kind of AI discourse are not so much the critical and concerned citizens they make themselves out to be. Rather, they are some of the biggest drivers of hype.

After all, what better way to drive hype than to play the role of earnestly concerned citizen whilst simultaneously hinging your whole concern on the narrative that these technologies are powerful enough to cause planetary annihilation? This is doubly true for those who then turn around to proceed to continue to build these technologies. That’s some super-villain absurdity. There is a villain-argument to be made here, but that’s less of a Marvel villain and more of something out of The Firm; and it’s not necessarily what I want to talk about here.

Rather, a commonality I note across all these videos about “just an example” of how a murder-bot might escape some lab, be it some YouTube talking heads, these guys, or this recent book, the arguments always hinge on a Matryoshka Doll of ifs: if AI can… if this… if that… It ought to be clear to anyone that for each if invoked, the more the example moves from credible scenario toward science fiction. I mean, it’s even in the title of Yudkovsky and Soares’ latest book: If anyone builds it, everybody dies.

This does something funny with the argument, of course, because these various examples can be very well thought-through, highly internally consistent, and each based on plausible scientific assumptions: in short, these can be highly rigorous. Yet for each if invoked—the narrator making a choice for where the story goes—it moves farther and farther away from reality. Yet, the rigor, as it is presented, often obscures this fact. Meanwhile, a Shoggoth-veneer is added to cover the argument in an outer shell of tragicomical dread. In other words, this leans into something more emotional. The tech-industry sure has had a tendency to abuse all kinds of sci-fi and fantasy metaphors; so much so that it’s a literal meme.

I remember when I was doing my research in Sweden with techno-utopians, and I often ran into the sentiment that these folks no longer issue predictions, because you always end up looking like a fool. “Just look at Kurzweil”, one of them told me.

At the core of all of this, I think, is the aim to misconstrue their prophecies as prediction. The technoscience-to-religiosity link has been made many times before, so this is itself nothing new. But I want to dwell for a moment on prophecy itself. Prophecy is a funny thing, as my ‘academic grandfather’ (that is, my supervisor’s supervisor) noted. Firstly, prophecy is never about the future, but always about the present. It’s a way to make sense of the current state of things. Secondly, one of the aspects of prophecy is its capacity to produce authorised interpretations of events. In other words, it’s an established discourse that authorises interpretations of the present.1

The lesson with prediction seems to be one current AI-hype makers have not yet learned, and we can see how much of the AI space is spiralling specifically because they cannot deliver on these prophetic predictions, and must find ever more absurd means of maintaining the hype. The doom-discourse is a clear part of this hype, because the picture it paints is one where the only possible outcomes can be boiled down to two outcomes: AI annihilates us, or we live happily ever after.2

There is a third category, some might insist, and that is that AI stops being developed at all—itself typically painted as doomsday-lite scenario, as it is often contingent on some repressive world government or another. Whilst I agree this is a third category in some ways, it could also be placed under the doomsday-scenario, as here humanity might as well have been annihilated, because our development—technologically; and by extension socially and politically—has stagnated.

I am not saying that these types of questions should not be asked, or that such scenario planning should be completely abandoned. Even I admit that there is a non-zero chance of this running away from us at some point. After all, scenario planning is a time-and-a-place kinda exercise, and needs to be used as the right tool for the right job. Nestling too many assumptions actually weakens the exercise—and, rationally, actually makes it seem even more implausible in many ways.

Indeed, the problem, as I see it, is pretty straightforward: this is not a narrative bound by as much rationality and honest engagement as its proponents want you to think it does—much of it hidden behind this veneer of rigour. No, what sustains this kind of thinking has much more in common with ideology and belief. The analysis is far from as neutral and balanced as it is made out to be—and it has real-world implications.

Specifically, the outsized coverage it has presents a real challenge for governance and risk. It frames this ‘alignment problem’ as: 1) a problem that singularly pertains to existential risk, and 2) a purely technical problem that necessitates technical solutions, such as how to ‘shackle’ an AI or otherwise ‘programme’ it to remain nice to us. These are cute sci-fi tropes, but the real solution is not quite as singular as that.

Whilst this is not one of these posts where I present super clear solutions—rather, this is a post where I finally write about something that’s been buffering in my brain for forever—I still want to propose something; perhaps not a solution, but more like another way of thinking about this. Often, the solution to the problem of disaligning an AI from murder is framed as figuring out a single solution—such as shackling—to an already, fully-fledged superintelligence. I’d like to re-frame this.

Let’s think about it instead as something procedural. What I mean by this is that instead of approaching the challenge as something with a singular technical solution to be implemented once we have already built these powerful systems, why not instead focus on doing it right from the start? AI is a reflection of society, so there is a strong argument to be made how decisions around data practices, deployment, and applications will, in time, stack. Is it really so unreasonable, then, to think that we ought to build a strong foundation from the beginning, as opposed to fix all accumulated problems in the future? Especially as we don’t know how or when or in what ways things could potentially go wrong moving forward.

This procedural approach, I like to think about as a type of incremental alignment, where the values and ideas we hold dear are designed into the systems from the very beginning. The idea here is that by building right from the beginning, rather than just building fast, there is a much higher chance to avoid murderbot-territory by focusing on quality in process rather than pure speed. Of course, this doesn’t necessarily guarantee anything; but frankly, neither does the more commonly floated alternative.

The point I’m trying to make is not that we should abandon these questions entirely, nor that scenario planning has no place. The danger is in how the conversation has been structured: as though “alignment” is something that happens later, as though safety is a technical patch to be applied once the machine is already built and humming.

But that framing is itself ideological. It presumes that the only meaningful work begins at the brink of catastrophe — that governance, design choices, institutional restraint, and ordinary procedural care are somehow secondary to the eventual arrival of a mythical technical solution. If we treat AI risk as something that can only be solved in the future, by someone else, with some yet-to-be-discovered shackle for a superintelligence, then we have already ceded the present. We have already accepted the premise that the foundations don’t matter — only the apocalypse does.

Incremental alignment is not glamorous. It doesn’t come with eldritch metaphors or doomsday narratives. It looks like boring decisions about data, deployment, incentives, accountability — the slow work of building systems that never needed to become murder-bots in the first place. And perhaps that is precisely why it is so rarely floated in the prophecy discourse: because it asks for responsibility now, rather than salvation later.

Remember: this is prophecy, not prediction.