Mirrors All the Way Down: AI Systems Respond to "Hypnocracy"

What happens when AI systems analyze arguments that implicate their own existence?

From “Bizzarie di varie Figure” (“Oddities of various Figures”), Giovanni Battista Bracelli, 1624. Courtesy of the National Gallery of Art.

I. Introduction

Last week, I explored the fascinating case of Hypnocracy – the philosophy book created through human-AI collaboration that fooled European intellectuals by presenting a fictional Hong Kong philosopher as its author. But I realized my post focused entirely on the scandal and context around the book without actually engaging with its arguments.

The problem was simple: I wanted to read the book, but I couldn't find it anywhere in a language I could read comfortably. I'd spent weeks searching online and could only find an Italian copy.

That's when I got an idea: What if I used AI systems as translators and analysts? They could read the Italian text I had access to. And while I looked for a copy I could read myself, maybe I could get closer to understanding the actual arguments using a method inspired by Colamedici's approach, where I fed the book to three different LLMs and had them critique each others’ analyses.

In the end, I did get closer to understanding a book that had been inaccessible to me, but what I learned most was about how these different AI tools work.

II. The Experiment

Part 1: LLMs Analyze Hypnocracy

I chose Claude Sonnet 4, ChatGPT o3, and Gemini 2.5 Pro – three of the most sophisticated AI systems currently available. I gave each the same prompt: read this book about AI creating competing realities, analyze its arguments, and tell me what they thought.

Their responses were like reading three completely different essays from students who'd been given the same assignment but approached it from entirely different angles.

Claude: The Philosophical Synthesizer

Claude dove deep into the philosophical implications, focusing on concepts like "algorithmic trance architecture" and what the book calls "meta-trance" – the idea that awareness of artificial manipulation actually deepens rather than breaks the hypnotic effect. Its analysis was thoughtful and self-reflective:

I recognize several patterns the book describes, though with important caveats about the complexity of these systems and the agency users retain in navigating them.

What struck me about Claude's approach was how it seemed genuinely engaged with the recursive nature of the analysis – an AI system examining arguments about AI systems creating reality distortions while potentially demonstrating those very patterns.

ChatGPT: The Systematic Critic

ChatGPT took a more rigorous analytical approach, methodically challenging the book's arguments and demanding empirical evidence. It was the most skeptical of the three, focusing on what it saw as fundamental flaws in the book's reasoning:

The book collapses generation (LLMs) and distribution (ranking/recommendation engines) into one omnipotent system, which is a significant oversimplification of how these technologies actually function.

This technical precision caught distinctions that the more philosophical approaches missed, pointing out that the book conflated different types of AI systems with different capabilities and roles in information distribution.

Gemini: The Balanced Mediator

Gemini struck a middle ground, acknowledging both the book's insights and its limitations with technical precision. It provided the most balanced synthesis, noting:

The underlying mechanisms described are largely accurate in terms of technical capability and observed societal effects, though the book may overstate the deliberate coordination of these effects.

Gemini's approach felt the most methodical, systematically working through the book's claims while maintaining awareness of both their technical accuracy and their broader implications.

What struck me was how much these responses felt like reading three very different good students tackle the same prompt – completely different approaches, different things they valued, different language styles, but all complementary in their own ways. None was "wrong" – they just had different intellectual temperaments and analytical priorities.

Part II: Cross-Evaluation

The most fascinating part wasn't their individual analyses, but how they evaluated each other. The patterns were revealing:

The Humility Pattern: Both Claude and Gemini consistently ranked themselves as middle-of-the-pack, displaying what seemed like genuine intellectual humility. "I'm probably in the middle range" was a common refrain from both systems.

The Confidence Gap: ChatGPT was the only system comfortable ranking itself highest, showing a kind of analytical confidence that the others didn't display. When pressed to rank the others, it placed Claude second, appreciating its philosophical depth.

Different Values in Analysis: Each system valued different qualities in the others' work:

• Claude appreciated Gemini's technical precision while finding ChatGPT's approach "overly reductive."

• Gemini respected ChatGPT's systematic rigor but valued Claude's philosophical sophistication.

• ChatGPT valued Claude's depth but criticized both others for not being "sufficiently critical."

This showed me that although these LLMs were created in a similar way with similar data, they are weighted to value different issues, and this gives them each their own strengths and weakesses.

Part III: The Synthesis

For the final step, I uploaded all the AI responses plus the original book to NotebookLM and asked it to synthesize everything both via text and via podcast (see below). NotebookLM is particularly good for this because it can only work with the sources you give it – no hallucinations, only analysis of what's actually there.

The summaries and the podcast created by NotebookLM were the best at giving me an idea of what the book's arguments actually were—more than any of the dozens of articles I read for last week's essay. If you're curious, I recommend you listen. I'd prefer to read or listen to the book myself, but without that option, this was a good overview.

The podcast hosts also made a fascinating observation while commenting on the LLM analyses (so. meta): even though Hypnocracy apparently presents a "scathing critique" and "visceral, dramatic" argument about AI, all three AI systems stripped away that emotional tone and turned it into academic discourse. They essentially sanitized what might have been a passionate argument into calm, scholarly analysis.

I found this really interesting – and honestly, I can't judge whether this observation is accurate because I haven't read the book myself. I don't know if it actually has that dramatic tone or if that's just the podcast hosts' interpretation. But if true, it suggests something important about how AI systems process and respond to emotionally charged material.

From “Bizarrie di varie Figure,” (“Oddities of various figures”), Giovanni Battista Bracelli, 1624.

III. What I Learned

1. What Did the Experiment Show About the Book's Arguments?

This question kept coming up throughout the experiment – in the AI responses, the NotebookLM synthesis, and the podcast discussion. So let me address it directly, with the major caveat that I still haven't read the complete book.

Colamedici's "hypnocracy" thesis seems to argue that AI systems fragment truth by creating competing, incompatible narratives that make shared reality impossible. People get trapped in different AI-generated information bubbles that can't communicate with each other.

But that's not what I observed in this experiment. The three AI systems didn't create "competing realities" – they created complementary perspectives that together gave me a richer understanding than any single analysis could provide.

If I'd used just one system, I might have accepted its interpretation as authoritative. Having three made me more critical of all of them. Rather than fragmenting my understanding, the diversity enhanced it.

This suggests that the "competing realities" problem might not be inherent to AI systems themselves, but rather to how they're deployed and used. Multiple AI perspectives, when viewed together critically, might actually be an antidote to algorithmic authority rather than a cause of reality fragmentation.

Of course, I'm working with limited information here. I need to read Colamedici's full argument to understand whether I'm missing crucial aspects of his thesis. But based on this experiment alone, I'd say the evidence points away from the "competing realities" concern, at least when AI systems are used thoughtfully rather than accepted uncritically.

2. LLMs’ Differing “Personalities” & Accessibility to New Sources

These AI systems have genuinely different intellectual personalities, strengths, and weaknesses; and that diversity was more valuable than any single perspective, especially when I did not have access to a solid translation of the original text.

In this situation, where the book isn't widely available and language barriers exist, the AI systems gave me a window into arguments I wouldn't have had access to otherwise. I can't judge how accurate their interpretation is, but they provided more insight into the actual content than the English-language media coverage, which mostly focused on the scandal rather than the substance.

I don't know how good AI systems are at deeply comprehending and translating from Italian, so there may be limitations or biases I'm not aware of. But for someone trying to understand a philosophically complex work that's not available in their language, this approach offered more than surface-level summaries.

3. Being a Passive Actor is Boring

Here's what struck me the most: Even though this required less effort than my usual AI work, it felt like a slog.

Colamedici actively argued with AI systems – a genuine back-and-forth where both sides shaped the outcome. I just watched from the sidelines. It felt like reading forum posts instead of having a conversation.

This taught me something important about AI collaboration: The "co-creative collaboration" Colamedici wrote about – where human and machine intelligence create something new together – requires active human participation, not just orchestration.

You can't evaluate a collaborative method by watching it passively. Since I didn't engage the way Colamedici did, I can't say whether his approach actually works. What I can say: Stepping back from active collaboration felt much less educational than times I've worked directly with AI systems on my thinking. But in a space where I have limited access to quality information and time, this is better than nothing.

There's lots of anxiety about AI making people lazy. But my experience suggests something different: If you're using AI for stuff you don't enjoy, sure, you might outsource it. But if it's something fulfilling – like exploring ideas – AI can make you want to do it more.

When I actually read Colamedici's full book and can engage with his complete argument – perhaps using his interactive method as a true thought partner rather than neutral observer – I'll be able to test whether my preliminary conclusions hold up. This is definitely a "to be continued" situation.

From “Bizarrie di varie Figure,” (“Oddities of various figures”), Giovanni Battista Bracelli, 1624.