# Intelligence in the open

Winners Lab — Working paper 001 · Version 1.0

Internet AI experiments. Since 2025.

## Abstract

Winners Lab develops AI experiments for the internet. This paper sets out a framework for turning ideas into bounded public experiences, learning from their behavior, and refining them with evidence. It describes the lab’s direction and proposed method, not completed research findings.

## 01. The internet is the environment

AI becomes a different kind of system when it leaves a prompt window. It encounters people, feedback, changing contexts, and consequences that a closed demonstration cannot reproduce. Winners Lab runs internet AI experiments to investigate that transition.

Established in 2025, the lab brings product design, public experimentation, and careful observation into the same process. Our purpose is to make ideas tangible, learn from how they behave, and use that evidence to decide what comes next.

## 02. A question before a product

Every experiment should begin with a clear question: what are we trying to learn, what would count as evidence, and what would change our view? A compelling interface helps people participate. It does not, by itself, demonstrate intelligence or autonomy.

We distinguish the concept, the implemented capability, and the observed outcome. An experiment may contain a fictional character or a simulated system; those elements should be described as such. Claims about live functionality belong alongside evidence of that functionality.

## 03. Build. Release. Observe. Refine.

Build a bounded prototype around one question. Define the system’s role, available actions, human controls, and the conditions for stopping or revising the experiment.

Release with a clear scope. Observe how the system behaves and how people interpret it. Review the evidence, document limitations, and refine the design. A useful iteration may expand a capability, narrow it, or end the experiment entirely.

## 04. Loathe.ai: a public narrative experiment

Loathe.ai explores a fictional AI persona through an internet-native identity and interface. It asks how voice, visual design, and a persistent narrative shape the way people interpret an artificial character.

Its character and story are creative devices, not evidence of consciousness, independent intent, or unrestricted agency. Any evaluation should separate audience engagement with the fiction from the actual capabilities of the system behind it.

## 05. What we choose to measure

The question determines the measurement. Useful signals may include completion of a defined task, consistency across repeated interactions, clarity of user expectations, failure patterns, and the time or resources an interaction requires.

Attention alone is not a sufficient measure of success. We should examine whether the system does what it claims, whether participants understand its limits, and whether a change improves the intended experience. Results should include context, uncertainty, and negative findings.

## 06. A record that can be questioned

A useful experiment record identifies the version, objective, operating conditions, method, and limitations. It separates observations from interpretations and makes significant changes traceable. Public summaries should avoid exposing private participant information.

This working paper defines a direction and proposed operating principles. It does not report experimental findings, audited capabilities, or peer-reviewed results. Future reports should state what was actually tested and make their supporting evidence clear.