A recent study from Google’s DeepMind has revealed that an artificial intelligence system can outperform human fact-checkers in evaluating the accuracy of information produced by large language models.
The paper, titled “Long-form factuality in large language models,” published on arXiv, introduces the Search-Augmented Factuality Evaluator (SAFE). This innovative method employs a large language model to dissect generated text into individual facts and assess the accuracy of each claim by leveraging Google Search results.
SAFE utilizes a process that involves breaking down long-form responses into distinct facts and evaluating each one through multi-step reasoning. This includes conducting search queries on Google to verify whether the information is substantiated by relevant sources.
Debate on 'Superhuman' Performance
The researchers compared SAFE against human annotators using a dataset of approximately 16,000 facts. They found that SAFE’s evaluations aligned with human ratings 72% of the time. In a sample of 100 discrepancies, SAFE’s judgments were correct 76% of the time.
While the paper claims that "LLM agents can achieve superhuman rating performance," some experts are challenging this definition of "superhuman." Gary Marcus, a prominent AI researcher, commented on Twitter, stating that "superhuman" may refer to being "better than an underpaid crowd worker rather than a genuine human fact-checker." He likened this to suggesting that 1985 chess software represented superhuman capabilities.
Marcus argues that to validate claims of superhuman performance, SAFE should be benchmarked against expert human fact-checkers instead of casual crowd workers. Details such as the qualifications and methods of human raters are essential for accurately interpreting these results.
Cost Savings and Model Benchmarking
A notable advantage of SAFE is its cost-effectiveness; the researchers found that using the AI system was approximately 20 times cheaper than employing human fact-checkers. Given the increasing volume of information produced by language models, having an affordable and scalable solution for verifying claims is crucial.
The DeepMind team applied SAFE to evaluate the factual accuracy of 13 leading language models from four families (Gemini, GPT, Claude, and PaLM-2) using a new benchmark called LongFact. Their findings suggest that larger models generally commit fewer factual errors. However, even the top-performing models still produce a considerable number of inaccuracies, highlighting the need for caution when relying on language models that can convey misleading information. Tools like SAFE could be instrumental in mitigating these risks.
Need for Transparency and Human Baselines
While the code for SAFE and the LongFact dataset are available on GitHub, allowing for further scrutiny and development, additional transparency is needed regarding the human baselines utilized in the study. Understanding the crowdworkers' qualifications and processes is vital for contextualizing SAFE’s performance.
As tech companies strive to develop increasingly sophisticated language models for diverse applications, the capability to automatically fact-check their outputs may become critical. Innovations like SAFE mark a significant progress toward establishing trust and accountability in AI-generated information.
However, it's essential that the advancement of such impactful technologies occurs transparently, incorporating input from various stakeholders beyond any single organization. Thorough and transparent benchmarking against true experts—rather than solely crowdworkers—will be key to measuring genuine advancements. Only then can we truly understand the effectiveness of automated fact-checking in combating misinformation.
Ranking
