Summary

  • Companies using AI are frequently troubled by the fabricated responses the technology produces, including legal sanctions and firms being required to abide by fictional policies.
  • Oumi, an AI platform, has come up with a system called HallOumi, an open-source model for detecting inaccuracies.
  • HallOumi determines whether claims made by AI have supporting evidence by analysing sentences.
  • It produces three outcomes for each sentence: a score to highlight the likelihood of fabrications, references to supporting evidence, and a human-readable explanation.
  • The model looks for nuanced indicators and can spot both unintentional and deliberate misinformation.
  • Oumi has made the tool open-source, meaning businesses can incorporate it into their workflows, while an API is being developed to make the process more seamless.
  • The platform is not just a checker and works with RAG (retrieval augmented generation) but can also work independently of it, making it a flexible tool.
  • Potential applications include ensuring legal documents are factually accurate and reliable, or verifying weather forecasts are as accurate as possible.

By Sean Michael Kerner

Original Article