Preventing Bots from Engaging with Unethical or Inappropriate Queries

AI for Process
1

Hello Everyone,

I’m Namrata Hinduja Geneva, Switzerland looking for effective ways to ensure bots do not respond to unethical or inappropriate questions. Are there built-in settings, filters, or best practices in AI design that help prevent such interactions? I’d appreciate any guidance on implementing safeguards to maintain ethical standards and responsible bot behavior.

  Thanks and Regards

Namrata Hinduja Geneva, Switzerland

2

Hi Namrata,

I can certainly provide information on how to prevent bots from engaging with unethical or inappropriate queries within versions 10 and 11. Here’s a detailed breakdown of the built-in settings, filters, and best practices.

Safeguards in Kore.ai XO Platform

The Kore.ai XO Platform provides a multi-layered approach to ensure bots behave ethically and responsibly. These safeguards have evolved, with more explicit and configurable features introduced in version 11.

Kore.ai XO 10: Foundational Safeguards

Version 10 of the Kore.ai XO Platform laid the groundwork for responsible AI, primarily through the integration of large language models (LLMs) and the introduction of the XO GPT models. The key to preventing undesirable bot responses in this version lies in a combination of training, specific model features, and data management.

  • XO GPT Model Guardrails: The introduction of Kore.ai’s own XO GPT models for functionalities like User Query Paraphrasing and Conversation Summarization came with built-in safety measures. These are not exposed as configurable settings in the UI but are integral to the model’s architecture. They include:

    • Content Moderation: The models are designed to detect and block harmful or inappropriate content automatically.
    • Behavioral Guidelines: These are embedded to maintain a professional and appropriate tone in the bot’s responses.
    • Input Validation: The model assesses user inputs to ensure they align with usage guidelines.

  • Knowledge AI and Content Curation: With the introduction of Knowledge AI, which leverages generative AI to answer questions from documents, the primary method of control is the content you provide it. Ensuring that your knowledge sources (like PDFs and FAQs) are thoroughly vetted and free of biased, unethical, or inappropriate information is crucial.

  • Data Security and PII Anonymization: XO 10 includes robust features for data security and the anonymization of Personally Identifiable Information (PII). This ensures that sensitive user data is not exposed in logs or used in model training, which is a core aspect of responsible AI.

  • Training and Testing: A fundamental best practice in XO 10 is the rigorous training and testing of your bot. By defining intents and entities for sensitive topics, you can control the conversation flow and prevent the bot from generating open-ended responses in areas where it shouldn’t. Using the platform’s testing tools, you can simulate conversations to identify and rectify any potential for inappropriate responses.

Kore.ai XO 11: Advanced and Configurable Controls

Version 11, now known as AI for Service, significantly enhances the responsible AI capabilities of the platform by introducing more explicit and configurable “Guardrails.” This gives developers more direct control over the bot’s behavior.

  • The Guardrail Framework: This is a major addition in XO 11. It allows you to validate LLM requests and responses to enforce safety and appropriateness. This framework includes:

    • Configurable Policies: You can define and apply policies to both the user input (prompt) and the bot’s generated response.
    • Built-in Scanners: The framework offers pre-built scanners for common issues like toxicity, prompt injection, and bias.
    • Custom Rules: You can create your own rules using regex patterns to ban specific topics or keywords, providing a tailored approach to content moderation.

  • Enhanced PII Detection: In XO 11, PII detection and redaction can be configured at the individual “Agent Node” level. This gives you granular control over what data is sent to the LLM, enhancing privacy and preventing the model from processing sensitive information.

  • Improved Monitoring and Logging: XO 11 features “Enhanced Usage Logs for Guardrails.” When a guardrail is triggered, the system logs the event with an identifier and an explanation. This transparency is vital for understanding why a bot responded in a certain way and for refining your safety policies.

  • DialogGPT and Agentic Orchestration: The introduction of DialogGPT for more natural, agentic conversations also comes with inherent safety features. The orchestration engine is designed to manage conversations in a more controlled manner, reducing the likelihood of unexpected and inappropriate deviations in the conversation.

Best Practices for Both Versions

Regardless of the version, the following best practices are essential for maintaining ethical bot behavior:

  • Human-in-the-Loop: Implement workflows where the bot can seamlessly hand over the conversation to a human agent when it encounters a sensitive topic or when the user expresses significant frustration.
  • Clear Bot Persona and Limitations: Define a clear persona for your bot and be transparent with users that they are interacting with an AI. Clearly state the bot’s capabilities and limitations.
  • Continuous Feedback and Iteration: Use the analytics and conversation logs to continuously monitor your bot’s interactions. Collect user feedback to identify and address any instances of inappropriate or unhelpful responses.
  • Ethical Design: From the outset, design your bot’s conversation flows with ethical considerations in mind. Anticipate potential misuse and build in safeguards from the beginning.

By leveraging the built-in features of the Kore.ai XO Platform and adhering to these best practices, you can effectively minimize the risk of your bot engaging in unethical or inappropriate interactions.