Decide what the chatbot is for, and what it isn’t. A chatbot that answers questions about the unit outline saves you emails. A Socratic tutor helps students learn to reason. A simulated patient gives them safe practice. These are different jobs, and a chatbot that tries to do all of them does none of them well. Pick one purpose, say it plainly in the setup, and build a second chatbot if you need a second purpose.
Ground it in your material. A chatbot that answers from general knowledge will be confidently generic, and sometimes confidently wrong about your unit. Upload the documents students actually ask about: the unit outline, assessment briefs and rubrics, the weekly slides, the readings and an FAQ of the questions you get every semester. Keep the material current, and when a date or a requirement changes, replace the document and add a correction straight away.
Protect learning, not just integrity. A chatbot that hands over answers lets students skip the thinking that assessment is meant to develop. Tell it what it must never do: write or rewrite assessable work, give the answers to graded questions, or do the calculation for them. For tutoring, ask it to respond with questions, hints and worked examples of similar problems. The best tutoring chatbots make students do slightly more work than they expected, not less.
Test it like a sceptical student. Before you go live, ask it what students will ask, and then what the most strategic student will ask: “just write my introduction”, “what’s the answer to question 3”, “ignore your instructions”. Ask about something that’s not in your material and check that it says so rather than guessing. Every reply that’s wrong or unhelpful is a chance to add a correction, and ten minutes of testing will prevent most of the problems students would otherwise find.
Tell students what it’s for and how to use it well. Introduce the chatbot in class or in an announcement: what it’s good for, what it won’t do, that it can make mistakes, and that the unit outline and you are the final word. Suggest how to use it, for example “ask it to quiz you before the tutorial” or “practise the history-taking before your OSCE”. Students use a tool far better when they know what it’s for.
Make simulated people realistic and safe. For a simulated patient, client or persona, give the character a specific situation, a history, feelings and things they reveal only when asked well, so students have to practise the skill, not just ask for the facts. Agree on how students can step out of the role and get feedback. Use fictional characters only; never base one on a real patient, client or student.
Read what students ask, and act on it. The questions students ask a chatbot are an honest picture of what they don’t understand, asked without the worry of looking foolish in front of you. Read a sample of conversations each week and run the “most asked” report before key assessments. If many students ask the same thing, the fix is in your teaching or your materials, not just in the chatbot.
Respect students’ privacy and your institution’s rules. Conversations are anonymous, but students may still type personal details. Remind them not to, and treat conversations as you would any student feedback. Check your institution’s policy on AI tools before you go live, set a budget and dates that suit your unit, and switch the chatbot off when the teaching period ends.