As chatbots spread, understanding how they generate responses is vital. This paper presents Think Like AI, an interactive tool that helps middle school students learn core concepts of large language models (LLMs):“token”, “prompt,” and how sampling parameters (temperature and top-p) control the diversity and randomness of generated text. The tool engages students through two interactive modules: (1) Next-Word Generation, which visualizes predicted word probabilities, and (2) Prompt Construction, which challenges students to test how LLMs handle negation. The tool was piloted with 35 students (grades 6-8, ages 10-14) in two after-school sessions. Although 91% (32/35) had prior experience with chatbots, mainly ChatGPT, most lacked conceptual understanding: 94% (30/32) were unfamiliar with the term “token” and 78% (25/32) held preconceptions about “prompt.” After using the tool, 69% (22/32) described “token” as numerical pieces or parts of words, and 84% (21/25) correctly defined “prompt” in context. Also, 66% recognized that increasing temperature increases randomness in output, and 56% recognized that higher top-p values allow the model to consider more word choices. Students also began to develop more nuanced views of control in generative AI, seeing that outputs are shaped by both user input and system parameters. These findings suggest Think Like AI is a promising tool for making probabilistic text generation visible and understandable through hands-on experimentation.