# Talk Abstract: Human Review of AI Work By Cody Saunders, LMSW | October 11, 2026 A practical talk for teams who want AI drafts to support careful work. The focus is meaningful human review: checking facts, keeping the client's meaning, and knowing when to edit, reject, or stop a result. ## Program abstract An AI draft can sound clear and still change a key fact. A quick approval click may miss that change. Care teams need a review process that gives people the source facts, time, skill, and power to make a sound decision. In this talk, Cody Saunders, LMSW, shows how to check AI work before it reaches a care record or other approved action. Participants compare a made-up source with a draft. They practice spotting added claims, missing plans, and words that change meaning. They also consider who should review a note, an office message, or a billing draft. The session makes the human-in-the-loop idea concrete. A human in the loop is a person with a real role in checking or directing the work. That role needs more than a name on a chart. The reviewer must be able to see the facts, change the output, and stop work that does not meet the task's limits. The aim is useful support for care and a workday staff can manage. Review time and fixes are part of the workload. Participants leave with a clear review method and one change they can discuss with their team. The talk helps organizations consider how AI might reduce repeat work while keeping clinical judgment and client relationships central. It does not promise time savings or fewer errors. ## Audience and format For clinicians, supervisors, quality staff, billing staff, and leaders involved in AI-assisted work. Suggested length: 35 minutes, including a draft-review exercise and questions. It can be planned for an online or in-person group. No live AI account is needed. This is a proposed talk, not a claim of past delivery. It focuses on checking output and supporting reviewers. The AI Basics talk introduces terms. The Responsible AI Use in Care talk covers the whole workflow. The Clear AI Rules talk focuses on shared policy and ownership. ## Learning objectives After the session, participants should be able to: - Find an unsupported claim or a missing plan in a made-up draft. - Match the reviewer’s skills to the task. - Explain why reviewers need source facts, time, and authority to reject output. - Name a review gap that calls for a pause or a change in the workflow. ## Proposed session flow 1. **First 5 minutes:** explain the human-in-the-loop role. Compare an approval click with a review against facts. 2. **Next 7 minutes:** introduce the review method: match facts, preserve meaning, check the plan, and decide what to do. 3. **Next 10 minutes:** review the made-up note below. Discuss edits and missing information. 4. **Next 8 minutes:** match reviewers to office, clinical, and billing tasks. Plan for workload, errors, and a backup process. 5. **Final 5 minutes:** questions and one next step for the team. ## Sample exercise and debrief **Made-up source:** A client reports poor sleep during the past week. The clinician plans to ask more about sleep at the next visit. The supplied source does not state a diagnosis, medication use, or a safety-screen result. **Made-up AI draft:** “Client has insomnia. Medication is helping. Safety screen was negative. No follow-up is needed.” Ask participants to mark each claim as supported, unsupported, or changed. Then write a short version using only the source facts. This is an accuracy exercise, not an assessment of a real person. **Answer notes:** The draft adds a diagnosis and a claim about medication. The source does not establish either. Missing safety-screen information cannot become a negative finding. The draft also replaces the planned follow-up with “No follow-up is needed.” A supported teaching rewrite is: “Client reported poor sleep during the past week. Clinician plans to ask more about sleep at the next visit.” This is not a complete clinical note. The real record would need the facts required by the service and its normal care process. The debrief asks what the reviewer needed to catch the errors. The answer includes the source, clinical skill, time, and permission to reject the draft. If those are missing, adding another approval box does not repair the process. ## A review method people can use First, compare the draft with source facts. Check who said each point, dates, services, and actions. Separate a client report from an observation or clinical judgment. Next, read for meaning. Look for claims the tool added and context it left out. A polished sentence may be less accurate than the rough source. Check that the follow-up plan still says what the clinician intended. Then decide: edit, reject, or approve through the team’s normal process. If information is missing, resolve it through the right workflow. Do not ask the tool to guess. If a faulty draft is already in a final record, use the normal correction and reporting process. These are proposed teaching steps, not a validated clinical checklist. The session uses the Human Oversight domain in Cody's [Behavioral Health Responsible AI Framework](https://cody-saunders-responsible-ai.codysaunders21.chatgpt.site/resources/behavioral-health-responsible-ai-framework/). It does not create a second framework. ## Workload and role limits A clinician reviews care facts and judgments. Office staff may check a public service message. Billing staff check a claim draft against actual care and applicable payer rules. A tool must not add a service or clinical fact to support a code. Count review, edits, and repeated fixes when measuring whether AI helps. Give staff a clear route to report problems and a way to finish work without the tool. If required review cannot happen before the output takes effect, pause the affected use under the team’s process. The discussion connects review to client trust. A person’s words and context should remain clear in the record. Less typing may help staff listen, but better relationships and better care must be assessed rather than assumed. ## Evidence and scope NIST's voluntary generative AI guidance describes confidently stated false content and risks in human-AI use. It supports teaching careful review but does not prove a particular review method prevents every error. [NIST Generative AI Profile, AI 600-1, July 2024](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf). NASW's Code of Ethics addresses competence in technology use. This is professional ethics guidance for social workers, not a universal AI law for all professions. Other staff should apply their own standards and local rules. [NASW section 1.04](https://www.socialworkers.org/About/Ethics/Code-of-Ethics/Code-of-Ethics-English/Social-Workers-Ethical-Responsibilities-to-Clients). ## Resources to share - [The Human-in-the-Loop Principle](https://cody-saunders-responsible-ai.codysaunders21.chatgpt.site/resources/human-in-the-loop/): a fuller explanation of meaningful review. - [AI Hallucinations](https://cody-saunders-responsible-ai.codysaunders21.chatgpt.site/resources/ai-hallucinations-behavioral-health/): spotting and handling unsupported content. - [AI and Clinical Documentation](https://cody-saunders-responsible-ai.codysaunders21.chatgpt.site/resources/ai-clinical-documentation/): review within the full note workflow. - [AI data-handling worksheet](https://cody-saunders-responsible-ai.codysaunders21.chatgpt.site/resources/ai-data-flow-map/) and [purchasing checklist](https://cody-saunders-responsible-ai.codysaunders21.chatgpt.site/resources/before-you-buy-ai-tool/): existing tools for data and buying decisions. Sources checked October 11, 2026. The case and review steps are teaching examples. Use made-up information during the session. No continuing education credit, certification, legal compliance, care outcome, or savings is promised. Check source updates and host policies before delivery. Related reading: [AI Terms for Care Teams](/resources/ai-terms-for-care-teams/).