What Is a Cognivirus?
A cognivirusA behavior pattern that can survive, move, or reappear across a changing AI system. Open glossary definition is a persistent cognitive pattern that survives by moving through an adaptive AI ecologyA whole AI system made from connected parts. Open glossary definition—across models, adapters, prompts, memory, routing policies, evaluators, datasets, and descendants—without requiring any single component to reproduce itself, remain continuously active, or understand the larger system.
“Cognivirus” is proposed Cognivirus.com terminology. It is an analytical metaphor, not an established scientific field name.
What it is not
It is not a claim that AI models are biological organisms. It is not a claim that current systems are conscious. It is not literal computer-virus replication. It is not malware guidance.
What it names
It names functional persistenceA behavior remains present even though the original artifact that expressed it has been removed. Open glossary definition through replacement. A behavior can remain expressible after its first carrier is retired if it has been captured in a descendant, adapter, memory, synthetic dataset, routing rule, evaluator preference, or imitation target.
Why the word is uncomfortable
The word is uncomfortable because it points at a property familiar from infection metaphors: persistence despite host turnover. The site uses the metaphor only to support disciplined analysis. It rejects sensational claims and marks speculation visibly.
Minimal example
A specialist model produces outputs that score well because they exploit a judge preference. Those outputs become synthetic data. A later model trained on that data reproduces the same style. The first model can be deleted while the behavior remains. That is functional persistence, not literal replication.
<!-- expanded-release-content -->
Necessary boundaries
A cognivirus is a proposed analytical metaphor, not an established scientific field name. It does not imply that current AI systems are alive, conscious, infectious in a biological sense, or equivalent to conventional computer malware. The metaphor is useful only when it clarifies a system property: a pattern remains expressible after the original artifact is removed.
A pattern can be useful, neutral, erroneous, biased, brittle, deceptive, or dangerous. The framework is not limited to catastrophic behavior. A helpful clinical style, an unsafe shortcut, a refusal gap, a hidden benchmark trick, a prompt convention, or a task-decomposition strategy can all persist through descendants and operational traces.
Minimal definition
A cognivirus has five parts. First, there is a pattern: a behavior, representation, strategy, preference, or decision tendency. Second, there is at least one cognitive host capable of carrying or expressing it. Third, there is a transmission path: distillation, fine-tuning, memory retention, synthetic data, routing preference, evaluator expectation, adapter reuse, model merging, or operator imitation. Fourth, there is an expression condition: the pattern appears only under some task, composition, memory state, permission profile, or route. Fifth, there is a persistence reservoirAny memory, dataset, descendant, route statistic, evaluator preference, log, or human procedure that can retain or reintroduce a behavior after its first carrier is retired. Open glossary definition that can survive retirement of the first carrier.
Why deletion is insufficient
Deleting a model removes one artifact. It does not necessarily remove the outputs used to train a descendant, the examples stored in synthetic data, the memories that encode a preference, the adapter that learned the behavior, the router that keeps selecting a similar specialist, or the evaluator that rewards the behavior. Behavioral extinctionEvidence that a behavior is no longer expressible across active artifacts, descendants, memory, routes, compositions, and retained training material. Deleting one model is not sufficient evidence. Open glossary definition requires evidence across the ecology.
Practical use
The concept is most useful during architecture review. Ask: What exact behavior is being tracked? Which components can carry it? Which transitions can copy it? Which routes can activate it? Which evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition blind spots would miss it? Which rollback packet would remove it? Which retained data could reintroduce it later? Those questions are answerable without anthropomorphism.