Can AI Fact-Check Another AI? DeepMind Thinks It Can

Can AI Fact-Check Another AI? DeepMind Thinks It Can

artificial intelligence. image provided by pixabay

This post is also available in: heעברית (Hebrew)

A research team of artificial intelligence experts from Google’s company DeepMind developed SAFE – an artificial intelligence system meant to fact-check the results of large language models (like ChatGPT or Bard).

Large language models and chatbots have gained immense popularity in the last few years. They are being used for anything, from planning a trip, writing school papers, solving math problems, and answering various questions. They do, however, have one major problem, and that is accuracy – all results provided by an LLM have to be manually checked in order to ensure they are correct, which greatly reduces the value of these tools.

According to Techxplore, the researchers at DeepMind created an artificial intelligence application that can automatically check answers given by LLMs and point out inaccuracies. One major way LLM users fact-check AI responses is by simply using a search engine like Google to find appropriate sources for verification, and the researchers took the same approach – they created an LLM that breaks down claims or facts in an answer provided by the original LLM, uses Google to find sites that could be used for verification, and then compared the two answers to determine accuracy. This new system is called Search-Augmented Factuality Evaluator, or SAFE, for short.

To test this system, the team used it to verify 16,000 facts from answers given by several LLMs. They then compared their results against human fact-checkers and found that SAFE matched the findings of the humans in 72% of cases. When looking at instances when SAFE and the human checkers disagreed, the researchers found SAFE to be the one that was correct 76% of the time.

DeepMind and its team posted the code for SAFE on the open-source site GitHub, making it available for anyone interested in using its capabilities.

The team published a paper about SAFE on the arXiv preprint server.