AI’s impact on higher learning remains a central conversation at Washington College as at many campuses. Michael Harvey, John S. Toll Professor of Leadership Studies, has created the following resource for students interested in using AI tools that identifies the limited precision of these tools and the value of dialogue when using LLMs to improve output.
Most people treat AI, especially ChatGPT and other large language models (LLMs), like a search engine: ask a question, grab the answer, move on. This can work for many facts and for general explanations or broad overviews, but it’s a risky approach when precision matters.
LLMs are pattern-matching systems trained on vast amounts of text. They’re good at recognizing structures, generating variations, and working through problems step-by-step. They have no judgment about whether your question is stupid, no fatigue when you ask them to try again, and no ego invested in their first response. This makes them useful for exploration, drafting, and working through ideas — but only if you treat them as a collaborator rather than an oracle.
The difference shows up in how you interact. A search engine rewards precision: the right keywords get you the right link. AI rewards dialogue. You set up a problem, examine what comes back, redirect, probe weaknesses, and gradually arrive at something useful. The first response is probably not the best response, and the tool has no way to know what “better” means for your situation unless you tell it.
To get good results, do three things:
- Give context. Don’t assume the AI knows what you’re working on. A quick setup helps: what you’re doing, what kind of response you need, what constraints matter.
- Push back. Ask “what’s weak here?” or “what would critics say?” or “what would an expert in this field say?” AI defaults to agreeability. Make it earn your trust by testing its reasoning.
- Iterate. First drafts are rarely final drafts. Refine what works, discard what doesn’t, ask for alternatives. The tool has no fatigue limit.
For anything important — facts, citations, technical details — you’re going to have to add a step of your own: verify. Extra work, yes, but highly recommended.
When responses start degrading or getting repetitive, start fresh.
Why the caution? In Plato’s Apology, Socrates observes that poets “say many fine things, but do not know the meaning of them.” Twenty-four centuries later, we’ve built systems with similar limits.
LLMs generate fluent text through pattern-recognition, not understanding. They’re optimized to produce plausible, confident- sounding text because that’s what their training rewards. Turns out the Socratic method — question and follow-up question — is the best approach to working with LLM
Why the caution? In Plato’s Apology, Socrates observes that poets “say many fine things, but do not know the meaning of them.” Twenty-four centuries later, we’ve built systems with similar limits. LLMs generate fluent text through pattern-recognition, not understanding. They’re optimized to produce plausible, confident-sounding text because that’s what their training rewards. Turns out the Socratic method — question and follow-up question — is the best approach to working with LLMs.
| What LLMs Still Get Wrong Despite dramatic improvements in reasoning and mathematics, LLMs have one persistent problem: they can’t tell when they don’t know something. Instead of saying “I don’t know,” they fabricate plausible- sounding answers. This shows up most dangerously in: Fake quotations and citations. Models routinely invent nonexistent legal cases, complete with realistic names and details. Even the most advanced systems still fabricate citations at alarming rates Factual errors. Even GPT-5, with its 80% reduction in hallucinations, still gets roughly one answer in ten wrong. Why? LLMs work through pattern-matching, not truth-seeking. They generate text that sounds right based on statistical associations, not verified facts. This is a calibration problem. A well-calibrated system matches confidence to accuracy — if it’s uncertain, it says so. LLMs don’t do that. They generate plausible text whether or not they know the answer. Ask for an obscure academic’s dissertation title, and instead of saying “I don’t have that information,” an LLM will invent a realistic-sounding title that’s completely wrong. That’s because current evaluation methods reward confident guessing over honest uncertainty. As of Fall 2025, LLMs fail the Socrates test: they don’t readily admit what they don’t know. Like Socrates, you have to push them to do so. |