A 2026 survey of more than 8,000 students across four Australian universities found that over 80% had used generative AI for study-related tasks, with nearly half using it regularly. Large language models have turned studying into a dialogue, but the article examines the architecture underneath AI study partners, where they fail, and habits that separate genuine learning from mimicry. Machine tutoring dates to the 1980s with rule-based systems like Carnegie Learning's Algebra Tutor, which modeled student knowledge with step-by-step hints but suffered from a knowledge-engineering bottleneck. Transformer-based LLMs like GPT-3.5, GPT-4, Claude, and Llama inverted that trade-off by prompting a general-purpose model instead of hand-coding pedagogical logic, but they lack determinism and can hallucinate or forget constraints. Tools like Khan Academy's Khanmigo enforce a Socratic style through a system prompt, while Quizlet's Q-Chat applies similar guardrails, and developer tools like GitHub Copilot Chat and Cursor act as coding mentors.