The rapid integration of artificial intelligence (AI) into education has increased interest in how digital tools can support higher-order thinking rather than simply automate tasks. Within STEAM learning, this issue is especially important because students must combine evidence, reasoning, and design decisions in authentic project work. Generative AI tools based on large language models were used in this study as learning scaffolds to assist students with keyword expansion, scientific query construction, journal abstract summarization, hypothesis development, and revision of written explanations. The study examined whether such AI-supported learning improved students’ critical thinking during a STEAM microcontroller media-design project. Two intact classes completed the same project workflow: one class used AI support, while the other completed the project without AI. Student performance was evaluated using a five-dimension critical-thinking rubric covering information and keywords, concepts and logic, journal abstract interpretation, hypothesis and reasoning, and academic formatting and compliance. Results showed that the AI-supported class performed better overall, scoring about five points higher on average and showing a large overall advantage. Improvements appeared across all rubric dimensions, with the strongest gains found in journal abstract interpretation and hypothesis and reasoning, suggesting that AI was most helpful when students had to interpret evidence, connect ideas, and justify decisions. The weakest-link bottleneck analysis, defined as the rubric dimension in which each student performed worst, showed that journal abstract interpretation remained the main constraint in both groups. These findings indicate that AI can serve as a productive scaffold for critical thinking and decision-making in STEAM projects, especially by supporting evidence use and iterative reasoning, while also highlighting the need for explicit instruction in reading and interpreting scientific abstracts.