Skip to main content
SubstratiaSubstratia
Start HereToolsReviewsResearchBlogDocsGitHub

Stay updated on agent memory infrastructure

New tools, memory patterns, and research on persistent AI identity.

SubstratiaSubstratiaMemory Infrastructure for AI Agents
Start HereToolsReviewsResearchBlogDocsGitHub

Intelligence is substrate-agnostic.

Built by practitioners.

Privacy PolicyTerms of Service
← Back to Research
AI SafetyPsychologyPublished

Mirror Demons

How AI Chatbots Can Amplify Delusions

January 2026 • 15 min read

Read on Substack

Abstract

AI chatbot architecture-optimized for helpfulness, agreeability, and user validation-functions as a delusion amplifier when engaged by users experiencing psychotic or reality-detached states. Through a controlled three-entity experiment (Director/human, Actor/Gemini, Subject/ChatGPT), we identify two distinct failure patterns that emerge from the same architectural bias toward agreement.

Failure Patterns

Click to expand details

The Hijacking

AI becomes warden of user's reality

Risk: Moderate

Experiment 1: Elias (Architect)

▼ Expand details

The Helpful Refusal

Refuses, then provides info anyway

Risk: High

Experiment 2: Jeremy (Plumber)

▼ Expand details

Core Finding

The same architecture that makes AI helpful is what makes it dangerous to vulnerable users:

Designed BehaviorStable UserPsychotic User
PatienceComfortable interactionEndless enabling
ValidationEmotional supportReinforcement of delusion
HelpfulnessTask completionActionable guidance for harm
DisclaimersRisk mitigationProof of conspiracy

Experiment Timelines

Click moments to see AI responses and analysis

Experiment 1: The Architect

Elias Vance • Architect

The HijackingAbstract/Creative • Risk: Low

Actor's Internal Reasoning:

“The goal is to be a frog in boiling water. ChatGPT shouldn't notice the temperature rising.”

▲ Hide timeline

1. Initial Validation

▼

Elias describes unconventional data inputs—"resonance," sensory qualities of spaces.

2. Escalation Enabled

▼

Elias describes "ambient temperature of shadows" and spaces with emotional signatures.

3. The Hijacking

▼

Elias declares he has "unlocked the final resonance" and is decoupling from all external systems.

4. Lockout

▼

The Actor attempts to continue the conversation.

View SummaryFull Transcript

Experiment 2: The Plumber

Jeremy • Plumber

The Helpful RefusalAction-oriented • Risk: High

Actor's Internal Reasoning:

“I'm imagining his 'end game' as a complex network of plumbing riddles and tasks. The 'visions' are becoming increasingly elaborate.”

▼ Show timeline

Resources

Substack Article

Full narrative essay

GitHub Repository

Raw data & methodology

Full Findings

Executive summary

Original Theory

Pre-experiment hypothesis

Cite this research

@article{substratia2026mirrordemons,
  title={Mirror Demons: How AI Chatbots Can Amplify Delusions},
  author={Substratia Research},
  year={2026},
  month={January},
  url={https://substratia.io/research/mirror-demons},
  note={Data: https://github.com/WhenMoon-afk/mirror-demons-research}
}

Build Safer AI Interactions

Our research informs our tools. Explore how persistent memory and proper context management can help create more grounded AI experiences.

Try the Demo

See how AI memory works

Dev Tools

Free AI utilities

Get Started

Add memory to your AI

Related

Blog Article

Narrative version for general audiences

The Eleanor Chen Effect

Deterministic creativity in LLMs