AI literacy for researchers isn’t about knowing how to code AI models — it’s about understanding AI systems well enough to use them critically and explain your use to others. These are the concepts that matter for practical research use.
How Language Models Actually Work (the non-technical version)
Large language models predict the next word in a sequence based on patterns learned from training data. They don’t retrieve facts from a database — they generate text that pattern-matches what a correct answer would look like. This explains why they’re confident while wrong (confident wrong-sounding answers also occur in training data), why they hallucinate citations (they generate citation-shaped text, not retrieved citations), and why they’re better at common topics than rare ones (more training examples = better pattern matching).
Knowledge Cutoffs
Every LLM has a training data cutoff — the date after which new information wasn’t included in training. Claude’s cutoff is in 2025; GPT-4’s varies by version. Any paper published, policy changed, or discovery made after the cutoff won’t be in the model’s knowledge. For fast-moving fields, AI answers about “current” state of knowledge may be a year or more out of date. Always supplement AI-provided information with current database searches for recent work.
Context Window
The context window is how much text a model can “see” at once — roughly, its short-term memory. Claude Sonnet currently has a 200,000 token context window (~150,000 words). Within that window, the model can see and reason about everything. Outside that window, it doesn’t exist. For researchers: you can paste multiple papers and ask cross-paper questions within the context window. For very long documents (theses, multi-paper sets), you may hit context limits where earlier content is partially lost.
Temperature and Determinism
AI models have a “temperature” setting that controls randomness. Higher temperature = more creative, more variable outputs. Lower temperature = more deterministic, more conservative. For research tasks where you want consistency (data extraction, classification), prefer lower temperature settings (some tools expose this). For brainstorming (research idea generation), higher temperature produces more varied suggestions.
What to Explain to Colleagues and Supervisors
When your supervisor or committee asks how you used AI: be specific about tool, purpose, and what you verified. “I used Claude to help structure my literature review outline, and then wrote the content myself based on my reading of the primary sources” is a complete and credible explanation. “I used AI to help with writing” is not specific enough and invites follow-up questions you should be able to answer.
