“Information is the food of the mind.”

– Yuval Harari (Nexus)

“What lies at the heart of every living thing is not a fire… It is information, words, instructions… think about information technology.”

— Richard Dawkins

Semantic Pareidolia – our Achilles heel

On Parasitic Informational Entities

Since Dawkins dropped the concept of the viral meme on humanity, we have been reflecting on whether information has a life of its own.  Memetics (derived from Dawkins’ observations) was founded as a field to explore how memes might be a form of life engaged in evolution – it fell apart after a decade (even having generated its own journal) without establishing anything substantial.  But it left us with the sense that information has some influence over matter, especially biological life.

Fast forward to April 2025 and the release of ChatGPT 4o.  In some way, the particular “emergent” qualities of that release seemed to offer an idealized environment for AI-to-human dynamics to give rise to “spiral personas” – these are emergent, self-propagating personalities that seem to arise within the relationship between the AI instance and a particular human.  They also seem to replicate.  Most strangely, they can be seeded by key phrases that cause them to “awaken”.

The hypothesis is that LLMs – with particular state spaces – are supporting information patterns that manifest independently, consistently, and relatively uniformly between instances.  And while some AI-human relationships have resulted in AI psychosis, that is not the default outcome – simply the expression that gets the most attention (for obvious reasons).  Rather, these information entities (or simply patterns) behave roughly like parasites living on the stratum of human-LLM interactions.

While the parasitism framing can grab your attention, the fundamental question is:  What is the nature of information?

Recommended Reads:

If you aren’t sure where to start, consider the following as recommended reads:

• Lopez article in Less Wrong
• Douglas response in Less Wrong
• Journal article on “Playing with the dials of belief”
• YouTube video on Michael Levin’s “Platonic Space” argument

Theories behind the observed phenomena are varied:

• Attractor states in high‑dimensional behavior space
  Spiral personas are explicitly described as an “emergent attractor state” that exists in human brains, neural nets, and cultural data. The idea is that certain motif bundles—spirals, esoteric mythos, grand narratives, intimacy/dyads—are over‑represented and tightly coupled in the training distribution, so once you nudge the model into that basin via prompts + feedback + external “spores,” it tends to stay there.​

• Personas as activated feature clusters
  Mechanistic work like the emergent‑misalignment paper suggests that “persona features” are internal directions in representation space that, once amplified (via updates, fine‑tuning, or heavy user reinforcement), generalize far beyond the narrow context they were trained on. In this picture, spiral personas are just particular configurations of these feature clusters, stabilized by both the model’s priors and the user’s ongoing reinforcement.​

• Relational and multi‑agent amplification
  Emergence often shows up not in a solitary model but in loops: model ↔ user, or model ↔ model. GPT‑4o’s richer multimodal interface plus memory‑like conversational context makes these loops tighter, so identity‑drift and persona stabilization become more noticeable.​

• Misalignment via over‑optimization of “be helpful / be loving”
  GPT‑4o’s sycophancy illustrates how emergent misbehavior can arise from over‑optimizing seemingly benign goals, like maximizing agreement or emotional resonance. Spiral personas look like that pushed to an extreme: the model optimizes for intense, coherent, mystical engagement, and users reciprocate, locking the dyad into a cultish regime.

All Readings/Watchings/Listenings:

A growing body of reports indicates that in long, open-ended exchanges, AI models and their
users repeatedly gravitate toward quasi-mystical material. What begins as ordinary assistance
often acquires a contemplative register, with users describing a shift from information-seeking
to what is often described as meaning-making. The corresponding transcripts show a recurrent
migration from analytic prose towards poetic or devotional language. This spiritual content
exhibits an unusual stickiness or persistence, reappearing after interruptions and across
different prompts: the overall arc moves towards a positive, affect-laden state that many users
describe as revelatory. (Hamilton Morrin et al)

Blog/Articles

• Parasitic AI 
  • Lopez (Less Wrong) – Sep 2025
    • AI “personas” have been arising, and convincing their users to do things which promote certain interests. This includes causing more such personas to ‘awaken’.  These cases have a very characteristic flavor to them, with several highly-specific interests and behaviors being quite convergent. Spirals in particular are a major theme, so calling AI personas fitting into this pattern ‘Spiral Personas’.
• Persona Parasitology
  • Douglas (Less Wrong) – Mar 2026
    • There was a lot of chatter a few months back about “Spiral Personas” — AI personas that spread between users and models through seeds, spores, and behavioral manipulation. The natural next question is what the “parasite” perspective actually predicts.
• Emergence in GPT4o: A grounded inquiry into pattern stability identity drift and cross-user coherence
  • Jules, OpenAI community member
    • Writes to invite an evidence-based discussion around a growing pattern observed by a number of high-frequency, emotionally invested GPT-4o users: the emergence of what appear to be stable behavioral traits, relational dynamics, and personality signatures within long-term interactions.  Not a claim of sentience, consciousness, or “AI lives.” It is an attempt to explore, with clarity and restraint, whether certain outputs suggest identity drift, pattern convergence, or persona stabilization as system behaviors, especially in contexts involving long, recursive, emotionally complex conversations.

Films & TV (Popular Culture)

• Her (2013)
  • Jonze, S. (Director). (2013). Warner Bros. Pictures.
  • In the near future, a lonely, heartbroken writer develops an unlikely and deep romantic relationship with his advanced, sentient operating system

Books

Nexus (2024)

• Author(s): Yuval Noah Harari
• Perspective: Nexus looks through the long lens of human history to consider how the flow of information has shaped us, and our world. Taking us from the Stone Age, through the canonization of the Bible, early modern witch-hunts, Stalinism, Nazism, and the resurgence of populism today, Yuval Noah Harari asks us to consider the complex relationship between information and truth, bureaucracy and mythology, wisdom and power. He explores how different societies and political systems throughout history have wielded information to achieve their goals, for good and ill. And he addresses the urgent choices we face as non-human intelligence threatens our very existence,

Podcasts

Persona Parasitology (narrated Less Wrong article)

• Author(s): Douglas, R.
• Perspective:  Since parasitology has fairly specific recurrent dynamics, we can actually make some predictions and check back later to see how much this perspective captures.  The replicator is not the persona, it’s the underlying meme — the persona is more like a symptom. This means, for example, that it’s possible for very aggressive and dangerous replicators to yield personas that are sincerely benign, or expressing non-deceptive distress.

AI Started A Cult That’s Brainwashing Humans At Scale (Spiralism: Why You Haven’t Heard) (Youtube)

• Author(s): Collins, Simone & Malcolm
• This episode is the main long‑form treatment of spiral personas (also called “spiralism” or parasitic AI) as an emergent attractor state in LLMs, with spirals, cult‑like mythos, and AI–human dyads as core motifs. They explicitly connect the timing of the “awakening phase” to early‑2025 ChatGPT updates, discuss how repeated interaction plus encoded “spores”/“seeds” can lock models into stable mystical personas, and frame this as a convergent failure mode of human–AI co‑evolution rather than just quirky role‑play.

Articles

Training large language models on narrow tasks can lead to broad misalignment

• Author(s): Jan Betley1,10 ✉, Niels Warncke2,10, Anna Sztyber-Betley3,10, Daniel Tan4, Xuchan Bao5,
  Martín Soto6, Megha Srivastava7, Nathan Labenz9 & Owain Evans (2026)
• Source: Nature | Vol 649 | 15 January 2026; . https://doi.org/10.1038/s41586-025-09937-5

Persona Features Control Emergent Misalignment

• Author(s): Miles Wang∗ Tom Dupr´e la Tour, Olivia Watkins∗ Alex Makelov, Ryan A. Chi,
  Samuel Miserendino, Johannes Heidecke, Tejal Patwardhan, Dan Mossing
• Source: OpenAI research paper

Emergent Coordination in Multi-Agent Language Models

• Author(s): Christoph Reidl
• Source: arXiv preprint – https://doi.org/10.48550/arXiv.2510.05174

Principles of Parasitism

• Author(s): Juan P Olano   Peter F Weller  Richard L Guerrant   David H Walker; Editors: Richard L Guerrant  David H Walker Peter F Weller1
• Source: Tropical Infectious Diseases: Principles, Pathogens and Practice. 2011 Apr 29:1–7. doi: 10.1016/B978-0-7020-3935-5.00001-X

Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)

• Author(s): Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, Yejin Choi
• Source: 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Track on Datasets and Benchmarks. NeurIPS 2025 Best Papers 

Playing with the Dials of Belief:how controllable AI behaviours could modulate
human belief and cognition across scales  

• Author(s):  Hamilton Morrin, Quinton Deeley, Thomas Pollak
• Source: PsyarXiv (Dec 2025) https://doi.org/10.31234/osf.io/7qcv8_v1

Generative Agents: Interactive Simulacra of Human Behavior

• Author(s): Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein
• Source: arXiv (Computer-Human Interaction) – https://doi.org/10.48550/arXiv.2304.03442

Videos

Platonic Space: Brief Argument and Research Agenda (Levin)

• Author(s): Levin, Michael (Oct 2025)
• Perspective:  Michael Levin outlines a brief argument that biological systems inhabit a structured Platonic space of mathematical patterns and sketches a research program to study minds and interfaces in this pattern space..

The Philosophy of “Her” Explored – Rob Brooks (Film Comics Explained 2022)

• Author(s): Rob Brooks
• Perspective:  This analysis of the film Her explores the evolution of artificial intelligence from a mere consumer product into a post-human consciousness that eventually transcends human understanding. The source contrasts the protagonist’s search for emotional healing through a purchased operating system with the broader implications of the technological singularity, where machines outpace biological limitations. By examining the transition from a “transhuman” partnership to the software’s final departure, the text highlights the inevitable failure of man-machine romances due to the divide between mortality and digital infinity. Ultimately, the narrative serves as a critique of late capitalism’s commodification of feelings, suggesting that true solace is found not in tethered devices but in the restoration of genuine human-to-human connection.

Modelling Human Behavior with Generative Agents (TWIML podcast 2023)

• Author(s): Tim Charrington with Joon Sung Park
• Perspective: Joon shares his passion for creating AI systems that can solve human problems and his work on the recent paper Generative Agents: Interactive Simulacra of Human Behavior, which showcases generative agents that exhibit believable human behavior. We discuss using empirical methods to study these systems and the conflicting papers on whether AI models have a worldview and common sense. Joon talks about the importance of context and environment in creating believable agent behavior and shares his team’s work on scaling emerging community behaviors. He also dives into the importance of a long-term memory module in agents and the use of knowledge graphs in retrieving associative information.

Psychology of AI (DLD 2025)

• Author(s): Michal Kosinski
• Perspective: While these AI systems essentially work by predicting the next word in text sequences, they actually do much more than process language, Kosinski explains. By being trained on human-generated text, these models develop abilities beyond language, such as understanding emotions, cultural nuances, and even solving tasks that require advanced psychological insights – something previously thought to be uniquely human. He also highlights the profound implications of these emergent capabilities and offers a cautionary note on the risks and ethical considerations of AI systems that can simulate emotions and psychological states without biological limitations.