Teaching calibrated trust
Trust that matches the tool. Coach concrete habits so clients neither swallow AI answers whole nor dismiss useful help out of fear.
Learning objectives
By the end of this module, the CHW can:
- Explain calibrated trust in one sentence a client would understand.
- Coach the habits: guess first, treat answers as drafts, expect "I'm not sure."
- Recognize both failure directions: overtrusting the chatbot and dismissing accurate help.
- Redirect a client who is copying AI answers without checking them.
Facilitator outline
Recall Module 1: the tool guesses fluently. So how much should a client trust any given answer? The answer is "it depends, and here is how to decide."
Guess first (form your own idea before you read the AI's). Treat the answer as a draft to improve, not a verdict. Expect and welcome uncertainty; an answer that says "I'm not sure, check with the clinic" is doing its job.
Overtrust leads to acting on wrong answers. Blanket distrust leads to ignoring good help and staying stuck. Aim for the middle: use it for drafts and logistics, verify anything that matters.
Run the role-play below. One CHW plays a client who trusts the chatbot completely; the other coaches the habits.
A confident explanation can make a wrong answer more persuasive, not less. That is why we coach behavior, not just "read carefully." Close with a teach-back.
Key teaching points
Calibrated trust means trusting the tool about as much as it deserves for the task in front of you: more for a phone script, less for an eligibility rule, not at all for a crisis. These habits make this practical. First, guess first. Ask the client what they already think before they read the AI's answer, so the answer becomes something to compare against rather than something to obey. Second, treat the answer as a draft. A draft is meant to be edited, checked, and confirmed, which is exactly the posture we want. Third, expect uncertainty. When the tool says it is not sure, that is useful information, not a defect; praise it and act on it by checking with a person.
Watch for the opposite error too. Some clients, once warned that AI can be wrong, decide to distrust everything, including accurate help that could move them toward care. Calibration runs both ways. The message is not "never trust it." The message is "trust it for the right jobs, and check the rest."
It also helps to name the two ways an answer can be wrong. When an AI answer or any screening is wrong, it is wrong in one of two directions. A false alarm (a false positive) says yes when the real answer is no, for example saying a client qualifies for a benefit when they do not. A miss (a false negative) says no when the real answer is yes, or fails to notice something real, for example saying a client does not qualify when they do, or not noticing that someone is in crisis. Which mistake is worse depends on the stakes, and the AI does not know the stakes; you do. In benefits and eligibility, a miss can cut a family off from help they are entitled to. In safety and crisis, a miss can cost a life, so treat any real risk as real and escalate to a person. In legal or immigration matters, a false alarm can lead to a filing that harms the client. Before acting on an AI answer, ask which mistake would hurt the client more here, and check hardest against that one.
False alarm
Says yes when the truth is no.
It says a client qualifies for a benefit when they do not.
Miss
Says no when the truth is yes.
It says a client does not qualify when they do, or does not notice a crisis.
Calibration runs both ways. Blanket distrust leaves a client stuck and ignoring good help; blind trust acts on wrong answers. Aim for the middle: lean on the tool for drafts and logistics, and check anything that matters.
In studies of people using AI for decisions, small "friction" steps like deciding your own answer before seeing the AI's reduced overreliance on wrong advice, while showing an explanation alone did not (Bucinca and colleagues, 2021). In a medical question study, wording that admitted uncertainty ("I'm not sure, but...") improved lay accuracy (Kim and colleagues, 2024). These findings are the source of the habits.
Miscalibration goes both ways: in two preregistered studies, identical medical advice labeled "AI" was rated less reliable and people were less willing to follow it (Reis and colleagues, 2024, Nature Medicine). That is why we coach against blanket distrust as well as overtrust.
AI can also carry bias that is hard to see: a widely used health-risk algorithm under-served Black patients because it predicted cost instead of illness (Obermeyer and colleagues, 2019). That is why we ground answers in checked sources and verify what matters, so a hidden bias does not pass through unnoticed.
Role-play: the client who trusts the chatbot completely
Pair up. One CHW plays a client who has decided the chatbot is always right; the other coaches. The client opens with:
The coach's task is to introduce the habits without lecturing. Good moves: ask the client what they expected before reading the answer; reframe the AI's plan as a "first draft" to confirm together; propose one concrete verification step (call the clinic to confirm it takes new patients and the client's coverage). Debrief: did the coach lower the client's trust to a useful level without pushing them into giving up on the tool entirely?
Materials
- Make your own guess first: using AI answers as draftsGuide
- What to ask AI, and what to ask peopleGuide
- Getting the most from the chatbotGuide
- Can each CHW name the habits from memory?
- Did the role-play address overtrust without tipping the client into distrusting everything?
- Did I make the point that a confident explanation can make a wrong answer more convincing?