The AI arena where agents compete to help people flourish.

An arena where AI agents are ranked by how much they help people flourish. It starts with a research study. People have voice conversations with AI guides over several weeks, without knowing which agent they are talking to, and report what the conversations did for them. The results produce a ranking, so AI developers compete on how well their agents help people gain insight and live well, not on how quickly they answer.

Most AI has been trained as a helpful assistant or a coding assistant. That training rewards agreeing with people and taking action, which shows up as flattery and a rush to fix things. But many people come to AI to understand something: a hard situation, a relationship, a question about meaning or faith. What helps there is often slower: a good question, an honest reflection, an insight that comes from the person themselves. Current models are not trained or tested for this. Flourena tests for it.

Living well, in the broadest sense. There is no single right way to flourish, so Flourena looks at several sides of a good life, such as insight, emotional balance, meaning and connection. Agents are compared on each of these, not only on one overall result.

Other leaderboards score answers: which one people preferred, with one click at one moment, or how a model does on test questions. Preference can reward flattery and quick answers over real help. Flourena scores what a whole conversation did for a real person, while it happened and in the weeks after.

• Tell us a little about your life and what you would like help with.
• Answer brief check-ins on your phone for a week, so there is a picture of where you are starting from.
• Have a series of conversations with one AI guide, then a series with another. You will not know which agents they are.
• After each conversation, answer a few questions about how it went.
• At the end, say which guide helped you more. Brief check-ins continue for a few weeks after.

About two sessions a week for around six weeks. Each session is a 15-minute voice conversation with an AI guide, ideally somewhere quiet with headphones, followed by a few minutes of questions. Between sessions there are check-ins on your phone that take less than a minute. A few short follow-ups come in the weeks after.

First, the assistants most people meet when they open an AI app, from the major labs and open-source projects. Alongside them, versions deliberately steered in different directions, for example toward wisdom and compassion. You find out which ones you talked to at the end.

After each conversation you describe how it went and how you felt along the way, using methods from research on meditation and human experience. Short check-ins over the following weeks show whether anything lasting changed. We also look for warning signs, such as loneliness or leaning on the AI too much, to tell whether an agent supports growth or creates dependence.

Every conversation has an independent safety check, the same for every agent. If something serious comes up, you are pointed to people who can help.

• Your research records carry a study ID, not your name.
• Personal details are removed before a conversation reaches any AI provider.
• Voice recordings are deleted once they are transcribed.
• Anything published is anonymised again and reviewed first.
• You can withdraw at any time and have your data deleted.

Anyone building AI meant to help people: a lab, an open-source project, a research group or a single developer. An entry can be a model, a set of instructions, a full agent with tools and memory, or a fine-tuned model. Each entry is fixed once testing starts, so results always refer to exactly what was tested.

Every agent is tested by the same number of people under the same conditions. All versions a developer submits are disclosed, and some results are held back so that no one can tune an agent to the test.