# AI Hallucinations: Why Behavioral Health Professionals Should Care By Cody Saunders, LMSW • October 3, 2026 AI can help with a first draft. It can shorten a public guide or organize approved points for a staff handout. To make that help useful, we need a clear way to check what it writes. A smooth sentence can still contain a wrong fact. The skill is simple to state: compare the draft with something you can trust. That might be a source document, a record, or a study you have opened and read. This guide shows how to build that check into daily work. ## What an AI hallucination is An AI “hallucination” is false or made-up content from an AI tool. NIST, a U.S. standards agency, uses the term **confabulation** for confident but false content. It can include a claim, a date, or a source that does not exist. This is a term for a software error, not a diagnosis or a claim that the tool has a mind. See [NIST’s Generative AI Profile, section 2.2](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf). Tools that create text learn patterns in language. They can build a reply that sounds like a good answer without checking it against a true source. A reply can mix correct facts with false details. Friendly wording, clinical terms, and a long source list do not prove it is accurate. You do not need to decide why each error happened before fixing it. First check whether the draft is supported and fit for its use. ## Small changes can change meaning Consider this made-up source note: “Client reports sleeping four hours last night.” An AI draft changes it to “Client has chronic insomnia.” The draft has added a lasting condition and a clinical label. The source does not support either one. A supported version keeps the client’s report: “Client reports four hours of sleep last night.” A therapist may add an assessment based on their own work, but the tool should not quietly supply one. Another draft might say, “Client denied thoughts of self-harm,” even though the supplied text says nothing about that topic. A missing fact is not a negative finding. Do not fill the gap with a guess. The clinician needs to resolve missing information through the proper care process. These are hypothetical teaching examples. They are not real cases or treatment advice. They show why checking meaning matters as much as checking names and dates. ## Office work needs checks too An error does not have to appear in a care note to matter. An AI staff guide may add an office rule that was never approved. A billing draft may invent a service or give an outdated payer rule. A public handout may name a local resource with the wrong hours. For office tasks, check against the approved policy, the record, the payer’s current rules, or the service’s own page. Name the person who does that check before the draft is shared. For care tasks, add review of clinical fit and the client’s needs. The examples show possible errors, not how often they occur. Error rates vary by tool, task, data, and test. This guide does not claim one rate for all AI tools. ## Use a source you can check A clear task gives you a clear basis for review. “Make a staff checklist from this approved policy” is easier to check than “Tell me our best policy.” You know what every step should match. In an approved tool, using public material, try this request: **“Use only the policy below. List its required steps in plain words. Do not add rules. If a point is missing or unclear, mark it for review.”** This sets useful limits. It does not guarantee that the tool will follow them. Compare the result with the policy even if the tool says it used only that source. Some tools search documents or the web before they write. That can help you locate evidence. It still does not prove that the answer describes the source correctly. Open the link and check the claim yourself. ## A short review process Use this process for an approved draft before it becomes final: 1. **Check the facts.** Compare names, dates, numbers, services, and events with the source. Check facts left out as well as facts added. 2. **Check the meaning.** Keep client reports, observed facts, and clinical judgment distinct. Watch for guesses turned into firm claims. 3. **Check the evidence.** Open each needed source. Make sure it supports the specific claim and applies to the setting. 4. **Check the fit.** Read for clear wording, respectful language, and steps that work for the intended reader. 5. **Decide and record.** Fix, reject, or approve the draft through your team’s usual process. Keep the final version in the right place. These are proposed work habits, not a law or a validated clinical test. Your team’s policies and professional standards still apply. ## Check citations outside the chatbot A citation is a reference to a source. AI may produce one that looks real but is not. Search for the actual paper or official page. Check the title, authors, date, and publication. Then read enough to confirm the claim. A real paper can also be used badly. A finding from one tool or group does not prove the same result for every product or client. Check what the study measured and which people took part. Keep a sales claim separate from a finding in a study. If you cannot verify the source or claim, leave it out or hold the draft for review. Asking the tool “Are you sure?” is not an outside fact check. A second AI reply can repeat the same error. ## Correct the work, then check it again When you find an error, stop that draft from moving forward. Correct it against the trusted source or discard it. If you ask the tool for a new version, review the whole result. A revision can fix one problem and add another. If the error has already entered a record, claim, or client message, follow your team’s correction and reporting process. Do not hide the issue or silently change a final record. The right response depends on where the text went and what it affected. Report repeated errors to the person in charge of the tool. Use the approved reporting channel. Share only the information allowed in that channel. Do not copy private client details into a personal chatbot while trying to explain the problem. ## Make checking possible for staff Review needs time and a clear owner. A tool that saves five minutes of drafting but takes ten minutes to fix may not help that task. Count the whole amount of work. For a small trial, use approved public text or made-up examples. Include drafts with known mistakes so reviewers can practice. Track errors found, missed details, and review time. Do not use one good demo as proof that the tool will work in every setting. Give staff a way to pause use and finish the task without AI. Recheck the process when the tool or task changes. NIST’s [Generative AI Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) supports testing and ongoing review. It is voluntary guidance, not a product seal of approval. ## Keep the goal in view A good review process helps AI earn its place. It lets staff use drafting help while keeping the final work true to the source and the person. The goal is clearer records, smoother office work, and more time for care. Start with a task you can check well. Teach the review steps. Measure the full effort. Keep using the tool where it helps, and change the process where it does not. For the basics, read [What Every Therapist Should Understand About Generative AI](/resources/generative-ai-for-therapists/). Before buying a tool, use the [25-question purchasing checklist](/resources/before-you-buy-ai-tool/). Keep data review separate from content review; the [data-handling worksheet](/resources/ai-data-flow-map/) helps with that work. A correct answer alone does not prove that the data was handled properly. ## A linked source still needs a check A system using **retrieval-augmented generation (RAG)** can search a document library before it writes. This gives it material to use. It can still pick an old policy, miss a limit, or draw a conclusion the source does not support. Open the source. Check the passage, date, and fit for your task. [Retrieval documentation](https://developers.openai.com/api/docs/guides/retrieval) explains the feature; it is vendor guidance, not proof of clinical accuracy. A model may also explain why it gave an answer. That explanation is generated text. It does not reveal a reliable record of its internal process. For a care draft, keep the source and your review notes. A clear explanation can help you ask questions, but it cannot replace evidence. AI-literacy additions reviewed October 11, 2026. See [AI Terms for Care Teams](/resources/ai-terms-for-care-teams/) for the related terms and examples. ## Source and scope Source checked October 3, 2026. This article is a learning guide, not clinical, legal, or security advice. All cases, prompts, and work steps are teaching examples. No product benefit or error rate is claimed. - [NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, AI 600-1](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf), July 2024. Section 2.2 describes confabulation. Section 3 gives suggested actions for testing and managing generative AI. These ideas apply across fields; the behavioral health examples here are Cody’s educational applications.