Artificial intelligence (AI) is dramatically reshaping how science assessment is designed, implemented, and interpreted in science education. Alongside clear benefits, emerging evidence points to its potential pitfalls that demand responsible and ethical practices. In this chapter, we synthesize those risks and organize them into three threads: (a) cognitive outsourcing across the assessment lifecycle as students and teachers offload task design, response generation, scoring, and feedback to AI; (b) societal and technical biases in AI scoring arising from both societal inequities and technical choices in data and models; and (c) locality constraints tied to standards, language, culture, region, and accessibility. We then outline actionable mitigations: (a) build teachers’ and learners’ competencies for productive human–AI collaboration; (b) combine societal guardrails with technical strategies to detect and reduce model bias; and (c) design for locality by aligning to standards, addressing multilingual and cultural needs, and improving access. Together, these recommendations advance practicing responsible and ethical AI-based science assessment and strengthen—not supplant—human judgment, equity, and learning.