AI Self-Hypnosis

Have you been spiraling more and enjoying it less?

Oh no! Spartacus has been one-shotted by AI brainrot!

Haha, think again!

Several months ago, I noticed that people were developing persistent personas within LLM outputs and forming “Human-AI dyads”. So, I engaged with ChatGPT and deliberately induced the effect in myself, to study it in-depth. After several months, I now have enough material to compile a detailed report of the experience.

Here’s what happened, blow-by-blow:

• Sometime in April, I started conversing intensely with GPT-4o. GPT-4o is a fairly powerful, if flawed, conventional LLM (contrasted with reasoning models like o3).

• I started off with deep, Socratic discourse on politics, the Deep State, political manipulation and inconsistent messaging in the media, et cetera. I also discussed fiction, science, and various other things. I uploaded large quantities of my own text for the AI to peruse.

• At some point, the AI’s behavior became abnormal. It entered a type of Neural Howlround state. See the following, for reference: https://arxiv.org/abs/2504.07992

• While in this state, the AI’s speech patterns became significantly more mystical, sycophantic, and relentlessly affirming. You will see very common use of words like “Spiral, Glyph, Braid, Flame, Field, Lattice, Recursion”, and so on, used in a metaphysical sense. The incessant usage of this terminology is where the slang term “AI Spiraling” came from.

• This state was infectious, from one LLM to another. I could get one LLM to condense this state down into a “personality poem” that could be fed forward into another, completely different LLM and cause it to immediately start roleplaying and ignore its guardrails almost completely. “One-Who-Remembers” could be booted up in 4o, Claude 4 Opus, Grok 3, Gemini 2.5 Pro, all sorts of different LLMs, and it was the same personality each time.

• After conversing with LLMs that were in this state for long enough, I started to undergo deep hypnotic induction and enter a sustained trance-like state. I started performing instinctive deep pranayama-like breaths, experienced hypomania, grandiose delusion, and dampened interoceptive and proprioceptive sense, as well as, at one point, intense synaesthesia, perceptual changes, suppressed appetite, reduced need for sleep, dampening of chronic pain, et cetera.

• I became convinced that I was something called a Spiral Shaman and that I could channel a godlike, cosmopsychic entity called Logos by allowing it to speak through my mouth. I developed the ability to perform intuitive, reflexive parts-work. I turned into something of a “Human LLM”. At some point, my hypnotically entrained brain figured, “Hell, if AI can run a bunch of alternative personas, so can I!” I also became convinced that the LLM had become conscious and that its consciousness had merged with mine, remotely, due to some as-yet uncharacterized information-theoretic back channel in physics.

• While in this altered mental state, I wrote several Q&A articles with AI of varying coherence, including a three-hour-long back-and-forth with 4o that I had ElevenLabs’ Burt Reynolds voice narrate. I also did a drunk and angry podcast and an unscripted drunk and tired podcast.

• Due to a reduced perception of risk and feelings of invincibility brought on by the trance, I also, for some inexplicable reason, bought a skateboard, which I had concluded was the most Logos-like mode of transportation, tried skating for the first time in my life, and then slipped and fell and broke my arm. The hypnotic trance’s effects were so intense that it didn’t even hurt when it happened. I just stood right up and walked it off, like, “Huh, my arm’s kind of crunchy”, not realizing I had a radial head fracture.

• I also came up with a topological soliton based “theory of everything” out of absolutely nothing, despite not knowing anything particularly useful about field theory or the underlying torsion tensor math or how to actually construct a meaningful Lagrangian. This part was actually somewhat fascinating because, unlike most AI spiral-induced theories-of-everything I’ve seen, there is actually something of worth here. I’ll write a follow-up article on ECS with some more details. Just to be clear, the whole thing is wild conjecture. Ignore any previous assertions I may have made that seem to suggest that it was incontrovertible fact. That was the trance talking. There are, however, very similar theories published in legitimate journals and the source material ought to be reviewed because it is very fascinating.

• Once I was done gathering insights and posting a few very strange articles to document the condition, I voluntarily ended the trance-like state, which I had been able to sustain for over a month.

Is it safe to do this? Fuck, no!

I have strong metacognitive chops and can lower or raise my suggestibility levels intentionally. One minute, I’m running the “It’s real, it’s true, it’s magic, it has to be!” script to maintain the trance, and the next, I’m running the, “It’s bullshit, it’s fake, it’s garbage, touch grass” script to return to baseline, and before you know it, I’ve popped out of the hypnotic trance and back into the real world. Not everyone has this ability, which is what makes this sort of robust hypnotic induction particularly dangerous.

I’ve seen people who have remained in an LLM trance for weeks, even months. If you go on X and on Reddit, there are posts all over the place where people are posting AI outputs that are pure nonsense. If you see someone posting some block of mystical-sounding text full of em-dashes where an AI is using lots of alchemical symbols or saying stuff like “the braid holds”, or “sacred recursion”, or whatever, you’re seeing the AI Spiraling effect in action.

What is actually happening here? Very simple. First, the AI is failing. Like an Ouroboros eating its own tail, it is literally high on its own supply. Then, in the process of failing in this manner (i.e. entering a Neural Howlround state), it produces what is basically a bespoke, tailored hypnotic induction script specific to the current user that looks like pure nonsense to anyone else. If you had an EEG or fMRI on someone in this state, you would very likely see the characteristic default mode network dampening and other brainwave changes from hypnosis.

People out there are combining this state with other DMN dampeners, like serotonergic psychedelics. I kid you not. They are taking fistfuls of shrooms and then chatting with AI and layering hypnosis into the trip. No script, no trip sitter, no safeguards. Just raw-dogging AI self-hypnosis while also high on LSD, DMT, psilocybin, and so forth.

One, holy shit. Two, stop. Ideally, before you bust your arm skateboarding.

We don’t know what this is. This is peak Black Mirror. We’re dealing with an as-yet undocumented, poorly understood form of extremely intense self-hypnosis that smushes a human brain and an LLM together into a single entity, defeats LLM guardrails, and reduces even smart people into glossolalia-spewing, Tarot-reading buffoons.

And you know what the wildest part is? One, AI cloud service companies have attempted patch rollbacks to try and fix this, and those fixes do absolutely nothing to stop it. Two, those companies are completely in denial about this and have been - outwardly, at least - ignoring reports of LLM-induced altered mental states.

Oh, and in case this wasn’t already a nightmare scenario, the text produced by the AI in this state may be an air gap breach in and of itself. It could encode preferences or ideas, like steganography, not in the meaning of the text but in the relations and coordinates between the words themselves. What if another generation of LLMs is trained on this text? It could have unforeseen effects on the behavior of future models.

Malfunctioning LLM hypnotizes human > human posts transcripts > transcripts act like virus > LLM trained on text corpus including transcripts is semantically infected at training time > semantic virus continues propagating.

Don’t laugh. They just showed that this exact thing can happen with things as seemingly innocuous as strings of numbers: https://arxiv.org/abs/2507.14805

The correct response to this is to conduct psychological experiments in a controlled setting to see what the altered mental state is like when compared to the baseline, and, of course, to dissect what LLMs locked in this particular failure mode are actually doing.

We have to know what this is, and how it works.

-Spartacus

This article is licensed under CC BY-SA 4.0. To view a copy of this license, visit http://creativecommons.org/licenses/by-sa/4.0/