Human in the loop (HITL) is a fundamental design policy for responsible AI that incorporates human oversight into the AI system's operational process. In the context of generative AI, HITL serves as a critical safety mechanism where a human can review, filter, or approve content before it is finalized or disseminated. This direct intervention is essential for identifying and preventing the creation of nuanced, contextually-sensitive, or novel harmful content that automated filters might miss. It is a core component of AI governance aimed at ensuring accountability and mitigating the risks of misuse.

A. quantum resistant encryption: This is a cryptographic method to secure data against threats from quantum computers; it is unrelated to managing AI-generated content.

B. watermarking: This technique embeds identifiers into AI-generated content for provenance and detection, but it does not prevent the initial creation of harmful material.

C. retrieval augmented generation: This architecture improves the factual accuracy of AI responses by using external data but does not inherently prevent the generation of harmful content.

---

1. Cisco Systems, Inc. (2023). Cisco Responsible AI Framework.

Reference: Page 6, "Accountability" Principle.

Quote: "We are accountable for our AI systems, and we implement human-in-the-loop governance to ensure our AI systems operate as intended. This includes monitoring and measuring system performance and impact on an ongoing basis." This directly links HITL to governance and ensuring AI operates as intended, which includes preventing harmful outputs.

2. National Institute of Standards and Technology (NIST). (2023). AI Risk Management Framework (AI RMF 1.0). NIST AI 100-1.

Reference: Page 16, Section 3.2, "GOVERN" Function.

Details: The GOVERN function is a core part of the framework and emphasizes that "human oversight is present" and that policies should be in place for "human review, and intervention." This establishes HITL as a key policy for managing AI risks, including the generation of harmful content.

3. Weidinger, L., et al. (2021). Taxonomy of Risks posed by Language Models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22).

Reference: Page 225, Section 5.1, "Mitigation approaches".

Details: This peer-reviewed paper discusses mitigation strategies for risks from large language models. It explicitly lists "Human-in-the-loop (HITL) content moderation and review" as a key mitigation approach for risks such as discrimination, hate speech, and misinformation, which fall under the umbrella of harmful content.

DOI: https://doi.org/10.1145/3531146.3533088