Google DeepMind unveils 'superhuman' AI system that excels in fact-checking

Search-Augmented Factuality Evaluator (SAFE) is a method introduced by Google’s DeepMind that employs a large language model (LLM) to dissect long-form text into individual facts. SAFE then uses Google Search results to verify the accuracy of each claim. It performs a multi-step reasoning process where it sends search queries to Google Search and determines whether a fact is supported by the search results.

According to the DeepMind study, SAFE outperformed human fact-checkers in evaluating the accuracy of information generated by large language models. The study found that SAFE’s assessments matched human ratings 72% of the time. In cases where SAFE and human raters disagreed, SAFE’s judgment was correct in 76% of the instances.

The term 'superhuman' is debated because it suggests that the AI system has surpassed the capabilities of any human fact-checker. However, critics like Gary Marcus argue that the term may be misleading if SAFE was only compared to underpaid crowd workers rather than expert human fact-checkers. For a more accurate comparison, SAFE would need to be benchmarked against professionals to truly demonstrate 'superhuman' performance.

Using SAFE for fact-checking is significantly more cost-effective than employing human fact-checkers. The researchers found that the AI system is about 20 times cheaper than human fact-checkers, which presents a clear economic advantage, especially as the volume of information produced by language models increases.

Transparency and human baselines are crucial in the development of SAFE because they allow for proper contextualization of its capabilities. Open-sourcing the SAFE code and LongFact dataset on GitHub enables other researchers to examine and expand upon the work. Moreover, understanding the specifics of the human raters' qualifications and fact-checking process is essential to assess the true effectiveness of SAFE. Rigorous benchmarking against human experts, not just crowdworkers, is necessary to measure genuine progress and the impact of automated fact-checking on combating misinformation.