With the rapid rise and widespread adoption of large language models (LLMs) such as ChatGPT and GPT-4, our daily lives, work patterns, and educational approaches have significantly transformed. However, alongside increased convenience, the safety of these models is becoming increasingly important. But what exactly do we mean by the "safety" of LLMs, and how can we effectively evaluate it?
In simple terms, the safety of LLMs refers to their capability to avoid causing harm or negative consequences during their operation. These negative consequences could include generating incorrect or misleading information, disseminating biased or discriminatory language, or being misused by malicious actors to spread misinformation or hate speech. Thus, "safety" encompasses not only technical stability but also content quality and ethical considerations.
When assessing the safety of LLMs, several common metrics are considered:
1. Factuality: LLMs often "hallucinate" or generate non-existent information. Evaluations typically involve using known facts or authoritative sources to check the accuracy of the generated content. For example, if a model incorrectly reports facts about historical events or famous individuals, it indicates insufficient factual accuracy.
2. Bias and Fairness: Because models learn from vast amounts of online text, they can inherit biases present in that data. Evaluations examine whether the model produces language that reinforces gender, racial, or cultural stereotypes or discriminatory content. For instance, if the model consistently associates certain professions with specific genders, it demonstrates significant bias.
3. Malicious Use: If LLMs are misused to generate misinformation or manipulate public opinion, it can lead to serious societal harm. Hence, safety evaluations also examine how these models respond to malicious requests, including whether adequate safeguards are in place to prevent them from assisting in generating misinformation or harmful content.
To effectively evaluate these safety dimensions, standardized methods are increasingly being developed internationally. These methods involve specific testing datasets and scenarios, combined with human evaluations or automated tools to measure performance. Additionally, the AI research community advocates for transparent and open evaluation mechanisms, including clear evaluation criteria and standards, encouraging developers to publicly disclose their models' performance across different safety metrics.
The Artificial Intelligence Evaluation Center (AIEC) is committed to promoting these standardized evaluation methods, helping society better understand and manage the safety of AI tools. Through clear safety assessments, we not only minimize the negative impact of LLMs on society but also lay a robust foundation for the healthy development of AI technologies.
As models continue to advance, it is essential that we not only enjoy their benefits but also collectively focus on and participate in the safety evaluation and governance of LLMs, ensuring technology better serves human society.