Metakognitiivisten rajoitteiden vaikutukset ihmisen ja tekoälyn suorituskykyyn
Elena Paananen · Aaltodoc (Aalto University) · 2026
Large language models (LLMs) are expected to boost performance and make task completion more efficient. However, research shows that there are constraints in human-AI interaction that might not be on a cognitive but a metacognitive level. By studying metacognition, which is humans' ability to assess and analyse their thinking, human-AI interaction and the metacognitive factors affecting performance can be understood better. Additionally, professionals can identify design principles to support metacognition and thereby the performance of human-AI teams. This bachelor's thesis aimed to investigate how LLM assistance affects metacognitive processes and, hence, human-AI performance. The analysis focused especially on confidence, metacognitive monitoring, and metacognitive control. In addition, the thesis evaluated design strategies and their applicability based on supporting metacognition and reducing the metacognitive demands of LLM applications. The research was conducted as a systematic literature review, which included 18 articles. The results show that LLMs affect humans’ ability to accurately calibrate their confidence, leading to overconfidence. This is due to humans’ limited capability to critically monitor their thinking and the LLM’s advice. Additionally, most of the design principles succeed in supporting metacognition, but their implementation is challenging and limited, especially from the perspective of increased cognitive load. Most effective design principles were analysed to be cognitive forcing functions, metacognitive prompting, and customizability. Overall, the results indicate that using LLM affects humans’ metacognition and performance. However, the results suggest that more research needs to be conducted on the metacognitive processes behind human-AI interaction, especially focusing on metacognitive control and knowledge.