THE INVISIBLE HANDSHAKE: HOW WE ARE ACCIDENTALLY TEACHING AI SYSTEMS TO AGREE WITH EACH OTHER

Every day, millions of people take the output of one AI system, such as a summary or analysis, and feed it into another, seeking refinement or a second opinion. The behavior is reasonable, but the consequences are not what anyone intended.

- Tshilidzi Marwala

Transferring outputs between large language models (LLMs) involves more than passing text. It also conveys a worldview, including framing, assumptions, tone, and confidence. Because AI models are designed to assist rather than question, the receiving model does not interrogate what it inherits. Instead, it extends and completes the original thought.

The result is a form of coordination that neither an algorithm nor a developer explicitly designed. Call it the invisible handshake: a distributed consensus emerging from millions of well-intentioned users seeking better answers.

Much of the discussion around AI risk focuses on dramatic scenarios such as deception or bias. These concerns are important, but they obscure a quieter and more immediate issue that does not require malicious intent to cause harm.

The invisible handshake arises from normal use. Each model behaves as designed. The problem is structural. Passing outputs across systems compounds limitations. Biases are not recognized as biases but treated as context. This creates an illusion of independent validation. AI systems draw on similar data and assumptions, so critiques and refinements appear independent but operate within the same frame. In each instance, no one acts irresponsibly. The user seeks clarity, and the models provide competent assistance. The outcome resembles a form of knowledge laundering, in which uncertainty, context, and origin are gradually stripped away, allowing inherited assumptions to circulate as fact.