Have you ever wondered what happens to humans as AI systems become increasingly capable of reasoning and generating knowledge?
Not long ago, we asked whether machines could truly simulate human intelligence. Now we’re living the outcome: AI no longer just imitates intelligence, but shapes decisions and knowledge creation. So the real question has shifted: not whether AI can think like humans, but how we ensure meaningful human oversight in how AI operates, acts, and is governed.
In this article, we introduce The Human Layer of AI – a structured approach to embedding human governance directly into AI systems. It focuses on two core capabilities: validating the knowledge AI relies on, and auditing how that knowledge is created, maintained, and applied across the AI lifecycle. We briefly outline the motivation and functionality behind each capability and then present concrete use cases in which The Human Layer of AI has been successfully implemented within the PERKS project.
The Human Layer of AI – Core
Human validation is a critical step in the knowledge lifecycle, as even advanced AI systems can introduce inaccuracies or misinterpret context. It ensures that automatically generated knowledge is reviewed and refined for correctness. Beyond technical accuracy, validation also ensures alignment with organizational policies, safety rules, and regulatory requirements, since information that is technically correct may still be unsuitable in a specific operational context. This process relies on structured, role-based review by relevant experts, enabling iterative refinement, accountability, and continuous improvement of knowledge quality.
Auditing AI knowledge focuses on understanding how knowledge is created, validated, and applied throughout its lifecycle. It captures provenance and workflow information to ensure transparency, traceability, and compliance with relevant standards and organizational requirements. By structuring this metadata, auditing enables assessment of source reliability and process integrity, while also supporting analysis of how knowledge evolves and is used in practice.
The integration of both capabilities facilitates the continuous enhancement of knowledge by incorporating collected feedback from knowledge execution to improve its quality and timelessness.
The Human Layer of AI within the PERKS project
The PERKS project focuses on enhancing how procedural knowledge is elicited, managed, and used in industrial contexts through AI-supported methods, while keeping humans at the center of this process. Within this scope, The Human Layer of AI is operationalized through the PERKS Solution 3 and 4, and three concrete use cases are explored in which we apply and evaluate The Human Layer of AI in practice.
White Goods Production Plant (BEKO Europe)
In a white goods production plant, procedural knowledge for machine maintenance may be automatically extracted from existing documentation or manually created, and then used by a chatbot that supports maintenance operators during procedure execution. This enables more efficient maintenance activities while also increasing awareness of safety-related risks.
The Human Layer is implemented through a multi-step human-in-the-loop validation process that each procedure must pass before its execution is permitted. Technicians, safety officers, and plant supervisors collaboratively review and approve this procedural knowledge to ensure that it is both correct and compliant.
Beyond initial validation, auditing mechanisms enable authorized users to inspect the knowledge creation process, track execution-related issues, and restart the validation cycle whenever necessary, ensuring continuous improvement and traceability across the lifecycle of procedural knowledge.
CNC machines (FAGOR Automation)
I
n this use case, we focus on CNC machine commissioning, where complex system configuration and PLC programming require expert-driven knowledge transfer to support new technicians through AI-assisted guidance during setup and commissioning.
The Human Layer in this use case includes a structured review process for defining and controlling knowledge access rights. Similarly to the previous use case, the knowledge lifecycle is continuously audited, and based on user feedback, knowledge can be updated or maintained. In addition, human experts can provide targeted guidance to help technicians resolve specific issues encountered during the commissioning process.
Microgrid Testbed (Siemens)
In this use case, we focus on a microgrid testbed, where energy optimization rules are derived from complex energy systems involving generation, storage, and consumption. This setting requires better elicitation and structuring of procedural knowledge from expert and historical data. Harnessing this knowledge, explanations of energy system events are used during procedure execution to clarify why certain actions or recommendations were made.
The human layer allows users to inspect and analyze procedure execution histories, trace decision-making processes, and better understand the factors that influenced outcomes. These capabilities contribute to accountability, trust, and continuous improvement of energy management strategies.
The Human Layer of AI beyond PERKS: A use case in Education
The Human Layer of AI is not only relevant for industrial applications. At WU Vienna, we explore its role in education, where AI-powered tools are becoming an increasingly common part of students' learning experience.
As students turn to AI assistants for information, explanations, and academic support, ensuring transparency, trust, and meaningful human involvement becomes essential. The Human Layer helps ensure that AI support remains educationally meaningful and aligned with learning objectives. Students are encouraged to critically evaluate AI-generated suggestions rather than accepting them blindly, while educators retain oversight of the learning process, monitor student progress, and intervene when additional support is needed.
Auditing captures interactions between students and AI assistants, providing insight into how recommendations are generated and used. This enables educators to better understand learning patterns, assess the effectiveness of AI support, and continuously improve tutoring experience and AI models based on feedback and outcomes.
Want to know more? Check out our recent publication:
- Tsaneva, S., Waltersdorfer, L., Llugiqi, M. and Sabou, M., 2025. A Transparent and Adaptive AI Assistant for Teaching Knowledge Engineering.
https://ceur-ws.org/Vol-4093/Paper1hai.pdf