Congruity.AI provides a diverse set of tools and options for business users and consumers alike to interrogate the consistency of content provided by large language models. It does this by first returning a varied sample of responses, utilizing different API parameters as well as prompt variation. We then ingest and analyze the results to present visualizations and other tools for users to investigate the quality and consistency of generated text. These approaches include an ngram census (portrayed on the text), similarity matrices (showing how similar responses are to one another), links to external sources for verification, and a comparison to constitutional rules or values for the language model. Using these complementary tools, users can drill down to identify elements of generated text that are consistent with the facts, with one another, and with corporate or personal values.
Category tags:"Congruity tackles the critical issue of AI bias and trustworthiness in the context of LLMs, making it an essential and timely project. I'm eagerly anticipating the developments and advancements of this initiative. Good luck!"
Paulo Almeida
Grants Manager