Responsible Prompting Recommendation: Fostering Responsible AI Practices in Prompting-Time

Research Background and Issues

• What problems or challenges did the authors identify?

  • With the widespread application of Generative Artificial Intelligence (GenAI) and Large Language Models (LLMs), there is a lack of guidance in the "Prompt Engineering" process to coordinate generative AI outputs with Responsible AI (RAI) practices.
  • Prompts may lead to harm, such as exhibiting social biases, misinformation, or deviating from intended design purposes. These risks necessitate integrating responsible AI principles into prompt engineering during interactions.
  • Currently, there is no systematic method based on responsible practices to design prompts, nor are there sufficient tools to support real-time prompt adjustments.

• Why is this issue important?

  • Prompt engineering is a critical aspect of interacting with generative AI, and its effectiveness directly impacts the quality and safety of generative AI outputs.
  • If prompt engineering does not incorporate responsible AI principles, the generated results may cause social harm or misleading information, leading to technology misuse or a decline in public trust.
  • Businesses and industries need more direct and effective methods to ensure that generative AI supports societal values and adheres to ethical standards.

• Research motivation and related work

  • The primary motivation of this research is to infuse the concept of responsible AI into the prompt design process, helping prompt engineers and users better avoid potential harm during interactions while promoting more diverse value inputs.
  • A significant body of literature discusses heuristic prompt design for generative AI, diverse generation methods, and prompt attack detection. However, few studies address tools that synchronize content recommendations to achieve responsible AI.

Solution

• What methods or solutions did the authors propose?

  • The authors developed an open-source recommendation system to support real-time prompt generation. This system can add "positive sentences" that promote social values or suggest removing "potentially harmful sentences" as users input prompts.
  • The approach includes a semi-automated prompt analysis and recommendation mechanism that combines algorithmic rules with a carefully curated corpus.

• What is innovative about this solution?

  • The system introduces Responsible AI (RAI) guidance in real-time, rather than relying solely on post-hoc review.
  • It provides instant sentence recommendations to incorporate social values and detects and removes potentially harmful content.
  • It offers sentence embedding and similarity analysis, which can be extended to multiple languages and scenarios, supporting dynamic adjustments.
  • By emphasizing a combination of "adding and removing content," the system enables users to validate ethical and value considerations before generating content.

• What are the implementation steps and key technologies used?

  1. Dataset Design: Create a JSON file containing over 2,000 "positive sentences" (related to social values) and "negative sentences" (harmful language in prompts).
  2. Recommendation Algorithm: Use sentence embedding and similarity comparison to extract similar sentences from user input prompts and recommend relevant values.
  3. Open-Source API: Develop a REST API to provide an interface, compatible with lightweight models supporting sentence embedding (e.g., all-minilm-l6-v2).
  4. User Interface Integration: Provide a user-friendly interface, including "sentence recommendations" and a comparison view of prompt content and generated content.

Research Outcomes

• What specific results were achieved?

  • Experiments showed that users could significantly enhance the responsibility of prompts through the recommendation system while optimizing generative AI content.
  • The research results indicated that the recommended sentences successfully embedded social values and helped users avoid negative outcomes. The recommendation algorithm performed well in real-time prompt optimization scenarios.
  • User feedback suggested that the system increased transparency and controllability in real-time interactions with generative AI and reduced the risk of generating harmful content.

• What are its advantages compared to existing solutions?

  • It combines positive and negative content recommendations, thereby promoting the value of generative AI outputs while removing inputs that could trigger potential harm.
  • It provides a lightweight, fast-response solution that does not require large-scale model fine-tuning, making it easy to scale and customize.
  • Near-instantaneous interaction supports users in optimizing inputs during prompt engineering, rather than relying on post-hoc monitoring or intervention.

• What are the experimental or evaluation results?

  • In user studies, over 80% of participants believed the recommended content improved the quality of generative AI outputs.
  • The average System Usability Scale (SUS) score reached 81.94, indicating strong usability and the system's ability to meet user needs.
  • Users reported that prompts with recommendations produced more targeted, clear, and socially responsible results compared to original prompts.

• Limitations and future directions

  • Limitations: The current study has a relatively small sample size, with participants primarily from the IT industry, which may not fully reflect the needs of different fields and industries. It is recommended to expand to more scenarios and user groups to verify generalizability.
  • If input prompts contain new terms unfamiliar to the model, recommendations based on sentence similarity may become inaccurate. Future work should focus on expanding model training or optimizing boundary detection.
  • Future Directions: Enhance UI features to better support personalized needs, such as providing user-defined prompt templates and industry-specific recommendations; explore alternative word suggestion mechanisms to avoid directly deleting potentially harmful sentences.

In summary, this research proposes an innovative real-time prompt recommendation method, providing robust support for responsible generative AI. It addresses the shortcomings of existing tools in terms of real-time capabilities, embedding social values, and guiding users.