Summary

  • Stanford researchers have created two new benchmarks that can assess the difference and contextual awareness of AI systems to reduce bias and improve overall decision-making processes.
  • Current fairness benchmarks, such as DiscrimEval, have been criticised for only assessing models on a one-size-fits-all basis, with an overly simplistic view of fairness.
  • The new benchmarks, which consist of a series of questions with objectively correct answers, have already been used to test models from Google and OpenAI, which performed poorly.
  • Techniques used to reduce bias in AI, such as instructing models to treat all equally, ultimately make them less accurate and effective.
  • The study suggests the creation of increasingly diverse datasets and greater investment in interpretability could help develop more accurate AI.
  • However, some believe AI will never be completely fair and unbiased without a human in the decision-making loop.

By Scott J Mulligan

Original Article