Large language models emerged as mainstream tools in 2022, but reasoning-capable models released in late 2024 through 2025 represented a step change in scientific utility, achieving PhD-level performance on domain-specific benchmarks and reliably completing autonomous tasks requiring hours of human effort. Yet the unbounded nature of these systems, combined with legitimate privacy concerns and rapidly shifting capabilities, makes it difficult for researchers to identify an access point and establish safe practices. This guide provides a practical framework for engaging with reasoning LLMs across scientific disciplines. I first describe what these models are and how they work, including core concepts (statelessness, context windows, prompt engineering) that inform productive use. I then introduce the "dual stance," a mental framework for treating LLM outputs simultaneously as high-quality intellectual contributions and as unverified claims requiring systematic checking. The guide presents a five-tier privacy framework for data protection, discusses legal constraints on using published literature with LLMs, and details verification practices including chunking, grounding through retrieval-augmented generation, and two-level fact-checking. Research applications (manuscript summarization, on-demand literature reviews, scientific writing) and education applications (interactive journal clubs, oral exam preparation) are illustrated with worked prompts. I address risks including hallucination, cognitive bypass, and degradation of scholarly outputs, proposing specific mitigations for each.
James M. Dewar (Sun,) studied this question.