By Cody Saunders, LMSW · October 11, 2026
Learn the words behind AI tools, then connect them to the work you do. A product label is only a starting point. What matters is the task, the data, the evidence, and the person who checks the result.
Artificial intelligence (AI)
Software that performs tasks such as finding patterns, making predictions, or creating content. AI is a broad label. Ask what a tool actually does before deciding where it belongs in care.
Large language model (LLM)
A model trained to work with language patterns. Many chatbots use one to build replies. A fluent reply can still be wrong. Clinical words do not show that the model has assessed a client.
Generative AI
AI that creates new content, such as text, images, or sound. It can help draft a public handout. A person still checks whether the draft is accurate and useful for its purpose.
Prompt and prompt engineering
A prompt is the request or material you give a tool. Prompt engineering means shaping that request. Name the task, approved source, audience, and limits. A strong prompt helps guide a draft; it cannot promise safe or correct work.
Training data and knowledge cutoff
Training data is material used to build a model. A knowledge cutoff describes a limit on its learned information. Search tools may bring in newer material. Check the real source and date instead of assuming the reply is current.
Context window and memory
The context window is the amount of material a model can use in one response. Product memory is a separate feature that may keep details across chats. A tool may also store logs. Starting a new chat does not prove old data was deleted.
Hallucination
A false or invented AI answer. NIST also uses the term confabulation. A draft might invent a study or add a symptom that was never reported. Compare important claims with the actual source before using them.
Algorithmic bias
A pattern in a system that can produce unfair results. A care draft might assume that everyone has stable housing or label the same behavior differently across groups. Review language and test examples from the people your team serves.
Black box and explainability
Black box describes difficulty understanding how a system reached a result. Explainability concerns ways to understand its behavior. A chatbot’s explanation is itself generated text. Keep source records and review steps; do not treat a convincing reason as proof.
Sycophancy
A tendency to agree with or flatter a user instead of giving a well-grounded answer. Some models show this in research tasks. Agreement is not a clinical assessment. Explore what a client found useful while checking advice that may have shaped their choices.
Jailbreaking
Attempts to get a model to bypass its safeguards. Written rules alone do not make a tool safe for care. Teams need controls for access, data, review, and reporting. A successful safety demo does not show how the tool behaves in every situation.
Human-in-the-loop
A person takes part in a system’s work, often by reviewing its output. For care, name who checks the source, can reject the draft, and owns the final decision. A quick approval click is not enough.
Chatbot
An interface that lets someone exchange messages with software. The chat format tells you little about its quality, data rules, or intended use. A general chatbot and a tool made for a specific care task need different reviews.
Retrieval-augmented generation (RAG)
A system finds material and gives it to a model to help build an answer. It may search an approved policy library. It can still retrieve the wrong page or misread it. Open the source and check whether it supports the answer.
AI parasocial attachment
A one-sided sense of connection with an AI persona. A person may feel heard even though the system cannot share a human relationship. This term is not a diagnosis. Ask how use affects sleep, daily tasks, real relationships, and care.
AI companions and mental health apps
Companions aim to provide ongoing interaction. Mental health apps have many different purposes, from journals to care tools. A caring tone does not establish clinical benefit. Review the exact purpose, evidence, data practices, and limits of each product.
AI safety and alignment
Safety work seeks to reduce harms. Alignment concerns how a system’s behavior fits intended goals and values. For a care team, turn those broad goals into checks: protect data, preserve client meaning, allow review, and stop when the tool fails.
Data privacy and consent
Privacy concerns how information is collected, used, stored, and shared. Consent concerns a person’s informed choice. They are connected but distinct. Permission from a client does not by itself make an unapproved tool suitable or satisfy every rule that applies.
Informed consent in AI use
Explain a tool’s role in words the client understands. Discuss what data it handles, who receives it, known limits, and available choices. Follow the rules for your profession and setting. Revisit the discussion when the tool’s role changes.
Automation bias
Giving too much weight to a system’s output because it seems fast, polished, or official. Before reading an AI summary, note key source facts yourself. Then compare them. Make it easy for staff to question and reject a draft.
Fine-tuning
Additional training that adapts a model using selected examples. This differs from supplying documents for one reply or retrieving sources. A model tuned on care text still needs testing for the exact task. Ask what data was used and how its use was allowed.
Token
A unit a model uses to process text or other inputs. A token may be part of a word. Token limits affect how much material fits; token counts may also affect cost. Do not assume a long record was fully read just because it was uploaded.
Application programming interface (API)
A way for one piece of software to request work from another. An AI feature inside a record system may use an outside model through an API. Map the full path of the data, including providers, logs, and stored outputs.
Temperature
A setting, in some models, that changes variation in generated replies. Lower settings may make replies more repeatable. They do not make facts true. Support and behavior vary by model; test the actual service and settings your team uses.
Multimodal AI
AI that works with more than one kind of input or output, such as text, images, and audio. Visit audio may include private details from several people. Check recording permissions, data handling, and transcript errors before considering this use.
Put the terms to work
For a draft, start with generative AI basics. For private information, use the client-data guide and the data-handling worksheet. For choosing a service, use the purchasing checklist and the responsible AI framework.
Sources and further reading
Technical details checked October 11, 2026. These definitions are for learning. The practice steps are suggested review habits, not a validated assessment or a statement that a product improves care.
NIST’s Generative AI Profile describes risks such as false content, harmful bias, privacy loss, and human reliance on AI. It is voluntary guidance, not a law or product approval.
OpenAI’s concepts documentation explains model inputs and tokens. Its retrieval documentation explains searching supplied material. Its memory guidance shows how memory can vary by product and settings. These are vendor sources for technical features, not evidence of clinical benefit.
Anthropic’s research on sycophancy reports agreement-seeking behavior in tested language models. This vendor research does not establish outcomes in therapy or describe every current product.
HHS guidance on cloud computing explains HIPAA duties for covered organizations and services handling protected health information. A business associate agreement is only part of those duties.
NASW’s client responsibilities address informed consent, competence, and permission for recording. These are professional standards; applicable laws and team policies also matter.
Technical feature references: Fine-tuning and temperature settings. Vendor documentation; neither establishes clinical benefit.
Further reading: Aeli Health’s AI Literacy Glossary for Therapists. Its topic selection helped inform this guide. The explanations and care examples here are original.