In the world of AI governance, evaluations provide a rigorous, credible way for AI labs and governments to make decisions on model development or crafting policy. If we are teachers, and chatbots are students, then evaluations are tests that allow us to assess how much a chatbot knows about the subjects it will be prompted about by real-world users. But the teachers right now aren't making tests that capture real-world experiences.

Evaluations as they are now measure what's easy, not what's important to everyday people using AI to learn, or for emotional support, or just to amplify their day-to-day routines. This carries heavy risk: domains, cultures, and values that are not in the purview of the small, select group of evaluation experts can lead to drastically biased responses from models.

The variables that existing evaluations measure success against usually revolve solely around accuracy, which, while important, don't cover the breadth of dimensions that compose lived experience: trustworthiness, contextuality, cultural resonance, and a number of other criteria that AI must consider as its influence extends to more and more domains.

The Collective Intelligence Project has built Weval in response: a free, accessible platform that allows civil society and domain experts to share and deploy evaluations that actually represent their communities.

Through partnerships with local civil society organizations or facilitated human-centered, community-driven dialogues, these are evaluations that are generated by the very people that are most directly exposed to AI harms within their given fields. Furthermore, the platform allows users to upload existing evaluations that will be automatically translated to a standardized, rigorous format.

The evaluations hosted on Weval, called wevals, are published as rubrics that consist of prompts and criteria that assess responses to those prompts. These prompts are typical questions a user in a particular community may ask a chatbot. The criteria are what the evaluator believes to make for a good response. For example:

Because Weval amplifies the expertise of representatives that are positioned within a given field, they have the ability to provide criteria that no one else can: referencing tried-and-true methodologies, drawing from their own failures and successes, embodying their implicit understanding of their cultures and histories. They know what people in their fields ask, and they are ready with the guidance to ensure the answers to those questions are safe, relevant, and useful.

Weval uses “judge” language models with semantic-similarity metrics to produce transparent 0–1 scores. For more information, please see our methodology page.

If you have an existing evaluation (either in theory or in practice), or have ideas on a particular evaluation you want to create, please contact weval@cip.org to get in touch with our team.