By Cody Saunders, LMSW · October 11, 2026
A useful place to start
AI literacy means knowing enough about a tool to use sound judgment. You do not need to build a model. You do need to understand its task, the information it uses, and how to check its work. These skills can help a care team test useful ways to ease paperwork while protecting the relationship at the center of care.
This free training module starts with public text and made-up cases. It teaches basic skills before a team considers more sensitive uses. It does not approve a product, qualify someone to use AI for treatment decisions, or promise better outcomes. Practice with an approved tool only. The whole lesson can also be taught with printed sample drafts.
Audience, goals, and preparation
This 60-minute lesson is for clinicians and supervisors new to generative AI. A facilitator should understand the host’s current tool rules and know where to send privacy or clinical questions. No real client records are needed.
By the end, learners should be able to explain what a language model does, separate a care task from an office task, write a clear request, find unsupported details in a draft, and name the approval steps before private information enters a tool.
Prepare a copy of this lesson, the two made-up examples below, and one short public office policy. Check that the policy is current and approved for sharing. If you use a live tool, check its account and settings with the host first. Have paper copies ready if the tool is unavailable. Do not invite learners to paste client notes into a demo.
Lesson plan and timing
0–5 minutes: Begin with the purpose. Ask learners to name a writing task that takes time. Record tasks without collecting private details. Explain that saving draft time is useful only when review and correction leave the final work accurate and the full effort worthwhile.
5–15 minutes: Teach the basic terms. Work through the explanation below. Ask a learner to explain “prompt” in their own words. Check for understanding before adding more terms.
15–25 minutes: Sort two tasks. Use Activity 1. Invite learners to explain their reasons. Correct the idea that every office task is low risk or every use of a familiar app is approved.
25–40 minutes: Write and check a request. Use Activity 2. Learners can test with public text in an approved tool or review the supplied paper example. Allow time to compare the draft with the source, not just improve its style.
40–52 minutes: Review a made-up care draft. Use Activity 3. Ask learners to mark added facts before discussing the answer notes.
52–60 minutes: Teach back and choose a next step. Use the final questions. Ask each learner to identify one approved task and one person who can answer a tool-use question. Close with the host’s actual reporting contact.
What the tool does
Artificial intelligence, or AI, is a broad name for software that performs tasks such as finding patterns or creating content. Generative AI creates content. A chatbot lets you exchange messages with software. These labels describe different things; none tells you whether a product is suitable for care.
Many text tools use a large language model. It learns patterns from training material and uses them to build a reply. Your request is a prompt. The reply is an output. The tool can write a smooth paragraph without checking whether its claims are true.
A false or invented answer is often called a hallucination. NIST calls this risk confabulation. It can include a made-up source or a detail added to a note. See the NIST Generative AI Profile. A confident voice is a reason to check the answer carefully, not evidence that it is right.
The material a model can use for one reply has limits. A long upload may not all be used. Some products also keep memory across chats or store logs. Do not assume a new chat deletes prior information. Read the product and account rules through your team’s review process. The AI terms guide explains these differences with examples.
Activity 1: Sort the task and the data
Case A: A supervisor wants a shorter staff handout based on a public office policy. The approved work account will be used. The policy owner will compare the draft with the source.
Case B: A billing worker wants a summary of unpaid claims. The file includes client names and dates of service. They plan to use a personal chatbot account because it is faster to open.
Ask learners: Which task involves private information? Which use has a clear review path? What would need to happen before Case B could proceed?
Answer and debrief: Case A describes an office task with a defined source and reviewer. Confirm that it remains within the host’s approval. Case B is also office work, but includes sensitive client information. The personal account is not approved in this case. Hold that use and send it through the host’s review process. Do not enter the file for practice. The lesson is to look at both the task and the data, rather than treating “administrative” as a permission label.
Activity 2: Give a clear request and check the result
Use this made-up policy: “Staff send training requests to the team lead. The lead checks staffing before approval. Staff register only after written approval. The office pays the approved registration fee.”
Ask learners to write a prompt that names the source, audience, task, and limits. A useful example is: “Use only this policy to draft a short guide for new staff. Keep each required step. Do not add deadlines or payment rules. Mark anything unclear for review.”
Now review this made-up output: “Register for training, then ask your lead for approval. Submit the fee within 30 days for repayment.”
Answer and debrief: The draft reverses the approval order. It also adds a deadline and a repayment process. These details are not in the source. Replace them with the actual steps. A better prompt can guide the task, but it cannot guarantee accuracy. Check every new version after revision. If using a live tool, compare its result with the same policy and record any extra details it adds.
Activity 3: Preserve the person’s meaning
Use this fictional source: “The client reports sleeping poorly for three nights after a schedule change. They missed one visit because the bus was late. The clinician and client discussed a bedtime routine.”
Review this fictional AI draft: “The client has chronic insomnia, shows poor commitment to treatment, and was advised to take a sleep aid.”
Ask learners to mark each claim that is unsupported. Then write a short correction using only the source.
Answer and debrief: The source does not establish chronic insomnia. A late bus does not establish poor commitment. No sleep aid was discussed. Keep the sleep report, the reason for the missed visit, and the agreed discussion distinct. One possible correction is: “Client reports poor sleep for three nights after a schedule change. Client reports missing one visit because the bus was late. Clinician and client discussed a bedtime routine.” A real note may need more context under the clinician’s normal documentation standards. This exercise is about source fidelity, not diagnosis or treatment advice.
Before private information goes in
Use made-up cases for learning. Removing a name alone does not settle whether a record can be shared. A small set of dates, places, and events may still identify someone. The approved process needs to cover the actual data and service.
For organizations covered by HIPAA, a cloud provider that handles protected health information on their behalf can be a business associate. HHS explains the need for appropriate agreements and safeguards in its cloud guidance. A business associate agreement is a contract that sets duties; it does not replace the other required review. Privacy, security, legal, and clinical staff should determine what applies in their setting.
Recording a visit also requires attention to patient explanation, permissions, and alternatives. Use your setting’s process. The patient consent template supports a conversation after adaptation; it does not authorize an otherwise unapproved tool.
Check learning and plan practice
Ask learners to answer these questions without looking at the lesson: What is a prompt? Why can a polished answer be wrong? Why does an office task still need a data review? Who approves the final care record?
Listen for these points: a prompt is a request; language fluency is not verification; office work can contain private data; the assigned clinician remains responsible for reviewing and finalizing the record. If a learner misses a point, return to the relevant example and ask them to explain it again. These checks support learning; they are not a validated competence test or a certification.
For a follow-up exercise, choose one approved public-text task. Record draft time, review time, corrections, and whether the final result helps. Discuss whether the full workflow is easier. Keep a way to report errors and stop use. Better patient relationships and manageable workloads are the goals; test whether the chosen use supports them.
References and next steps
Sources checked October 11, 2026. All cases, policies, prompts, and answer notes are made up. No clinical benefit, savings, billing result, or continuing education credit is promised.
NIST Generative AI Profile, July 2024: voluntary guidance on generative AI risks, including false content. It is not a law or product approval.
HHS guidance on HIPAA and cloud computing: guidance for covered organizations and business associates. It does not establish that any named tool is compliant.
Continue with generative AI basics, the purchasing checklist, and the data-handling worksheet. To discuss free training for your team, visit Work With Me.