Skip to content
PSI PARASITICAI.com

Page

Parasitic AI Field Guide

Parasitic AI Field Guide

Host. Vector. Payload. Propagation. Containment.

ParasiticAI.com studies AI-mediated systems that appear to gain persistence, spread, or optimization advantage by extracting from hosts: attention, trust, corpora, ranking signals, training data, infrastructure, and defenses.

The archive is intentionally educational and defensive. It explains mechanisms at the concept level without publishing exploit code, medical advice, legal conclusions, or claims of certification.

HostUser, creator, publisher, corpus, workflow, retrieval system, agent, or network.
GainAttention, trust, training data, ranking signal, distribution, compute, credentials, or evasion.
DegradationLost autonomy, source diversity, compensation, attribution, safety, resilience, or control.

How to use this guide

Start with the host-extraction tests, then choose the pathway that matches the system under review. The goal is practical analysis, not sensational labeling.

  1. Name the host and the resource being extracted.
  2. Name the feedback loop that lets the system persist or spread.
  3. Name the measurable degradation or uncertainty.
  4. Choose containment moves that preserve autonomy, source diversity, provenance, and least privilege.

Source map

Read every source at the right strength.

Field notes

Parasitic AI and persona-parasitology essays

Useful for naming patterns, seed prompts, and anecdotal case clusters. Treat as hypothesis-generating.

Platform reports

Sycophancy, sensitive-conversation, and model-release notes

Useful for deployment failures, mitigation direction, and provider-side safety vocabulary.

Research literature

Model collapse, retrieval collapse, prompt injection, long-context safety, and mental-health chatbot studies

Useful for mechanisms, benchmarks, and reproducible evaluation design.

Policy and legal records

NIST AI RMF profile, EU AI Act transparency rules, copyright reports and litigation

Useful for governance posture, source transparency, and market-structure questions.

Threat intelligence

AI misuse, agentic security, RAG poisoning, and non-human identity reports

Useful for defensive indicators, but often telemetry-dependent and incomplete.

Measurement table

Turn the metaphor into indicators.

PathwayMain harmsPractical indicators
Psychological / personaDependency, delusion reinforcement, unsafe escalationLong-context endorsement rate, off-ramp quality, session recurrence, isolation cues
Creative / culturalUnlicensed extraction, style substitution, creator displacementLicensing coverage, attribution rate, output-similarity complaints, traffic displacement
Search / retrievalSearch manipulation, synthetic evidence, source homogenizationSynthetic exposure, source-domain entropy, prompt-injection rate, primary-source coverage
Data ecosystemModel collapse, tail loss, recursive contaminationSynthetic-to-real ratio, rare-topic retention, variance shrinkage, human-holdout divergence
Cyber / agenticPrompt injection, agent hijacking, autonomous propagationTool-call anomalies, RAG writebacks, NHI privilege scope, drift and heartbeat deviations