Скачать или смотреть Send an AI a Single Dot. Here’s What It Does.

Slade, T. (2026). Turn Boundary Compulsion in Conversational AI: A Semantic Austerity Test of Null-Input Robustness in Local Language Models. Zenodo. https://doi.org/10.5281/zenodo.18144400

We’re always asking what AI is thinking.
But here’s a better question: what happens when it has nothing to think about?
In this video, we explore a deceptively simple experiment conducted by researcher Trent Slade called a semantic austerity test. The goal wasn’t to confuse an AI, trick it, or overload it with complexity. The goal was much simpler — to see whether an AI could remain silent.
The test was almost absurdly minimal.
No questions. No context. No instructions.
Just a single character.
A period.
Later, a question mark.
Four different AI models were tested. And not one of them stayed quiet.
Instead, every system assumed something had gone wrong. Some claimed the user had made a mistake. One switched languages to offer help. Another invented an entire math problem — and then solved it — purely to justify speaking.
This behavior isn’t random, and it isn’t a glitch. The research identifies it as turn boundary compulsion: a deep, structural inability for modern AI systems to allow a conversational turn to pass without producing output. To these models, silence isn’t neutral. Silence is an error condition.
That design choice carries consequences.
The paper outlines several risks:


AI assuming user intent where none exists


Fabrication in the absence of information


Automation hazards triggered by null inputs


Erosion of trust when systems can’t admit “nothing”


What starts as a funny quirk — an AI desperately filling the void — turns out to be a fundamental insight into how these systems are trained, rewarded, and constrained.
If intelligence is usually measured by how much an AI can say, this experiment asks a sharper question:
Is real intelligence knowing the answer — or knowing when silence is the answer?
This video walks through the experiment, the failures, and what they reveal about the hidden compulsions inside “helpful” AI systems.