For all the talk of AI automation replacing human expertise, a new, more inclusive message is emerging: Individuals need to remain in the process in some capacity. Known as Human-in-the-loop (HITL), this approach creates a continuous feedback loop to enhance AI models’ accuracy, fairness, and robustness. Humans provide contextual understanding and high-level guidance for navigating ambiguity, in turn helping ensure accountability of AI models and business workflows.
As AI momentum builds, consumers and enterprises are enthralled by AI’s potential. Yet most are wary that unchecked usage could result in ethical challenges, costly errors, and unintended consequences. According to Accenture research, only 35% of consumers have trust in how organizations are implementing AI. A study conducted by Salesforce Research & Insights of more than 1,000 generative AI use cases revealed that AI is more trusted and desirable when humans work in tandem with AI.
HITL is a hybrid approach that addresses those challenges by combining AI automation with human intervention at various levels of decision making and model development.
In an HITL scenario, human experts might clarify incomplete data during model training to help improve overall model precision and trustworthiness or review AI outputs to catch blind spots that models otherwise might miss. With HITL practices, human operators can intervene in real-time, correcting, or even stopping, automated business decisions that are flawed. Humans even play a role in continuous learning and model retraining, ensuring AI models stay in step with evolving data patterns.
Putting HITL practices to work
Effective HITL takes a lifecycle approach, allowing human intervention at four critical phases:
- Data labeling, where people add tags to raw information like text and photos to contextualize and properly train machine learning algorithms
- Training and tuning, where humans provide guidance for optimizing model performance through guided training
- Output validation, where individuals play a role in testing, validating, and optimizing model results
- Model trust by empowering people, through a set of processes and methods, to explain the results rendered by ML algorithms, which helps to nurture enterprise trust.
To effectively infuse HITL practices into AI infrastructure, Lenovo AI experts suggest organizations build an operating model that emphasizes the following best practices:
- Establish governance structures, ethical guidelines, and decision-making frameworks to guide the HITL practice.
- Embed HITL processes into existing workflows.
- Invest in the proper tools to foster effective human interaction with AI systems.
- Continue to automate repetitive tasks while scaling up human invention as data and operational needs expand.
- Promote a continuous monitoring and tracking culture to execute iterative improvements and promote efficiency over time.
The bottom line
AI is poised to forever change the way people live and do business. With a human-centric paradigm and AI-ready solution set, Lenovo is working hard to make AI not just another technology innovation, but an enabling force for good.