AI-Informed Therapy Glossary (Revised v2)

Revision date: 2026-01-06
Term Category Why it’s relevant for AI-informed therapists Simple definition (5th-grade level)
AI (Ambient) ScribeCore ConceptListens to sessions and generates a draft clinical note (e.g., SOAP/DAP). Requires clinician verification, vendor diligence (BAA), and clear client disclosure.A tool that listens to a therapy session and writes a *draft* note for the therapist to review and fix.
AI NotetakerCore ConceptGeneral-purpose meeting summary tools. Often not built for clinical privacy/compliance. Therapists should distinguish consumer vs clinical-grade tools.A tool that records and summarizes conversations—usually for meetings, not therapy.
Agentic AICore ConceptSystems that take actions toward goals (e.g., scheduling, sending messages). Raises consent, safety, and accountability concerns when “acting” rather than “suggesting.”An AI that doesn’t just answer—it can take steps to complete a goal for you.
AlgorithmCore ConceptFoundational to how software makes decisions; helps therapists understand that AI outputs come from programmed or learned rules, not intuition.A step-by-step recipe a computer follows to do something.
API (Application Programming Interface)Core ConceptExplains how systems connect (e.g., EHR ↔ billing ↔ transcription). Critical for understanding where client data flows and where it can be exposed.A controlled way for one app to ask another app for information—like a waiter taking an order.
Artificial Intelligence (AI)Core ConceptUmbrella term for systems that perform tasks associated with human cognition (pattern recognition, prediction, language). Important for client psychoeducation and boundaries.Computer tools that can learn patterns and do tasks that seem like “thinking.”
DataCore ConceptAI quality depends on data quality and fit-for-purpose. Clinicians should assess what data is collected, shared, stored, and how it is protected.Information (words, numbers, pictures) that computers use to learn or make decisions.
Digital PhenotypingCore ConceptPassive behavior signals from devices (sleep/activity/typing) used to infer mental state. Needs careful interpretation, bias awareness, and transparent consent.Apps using phone signals (like movement or sleep) to guess how someone might be doing.
Generative AICore ConceptProduces new text/images/audio. Relevant for homework support, summaries, and drafting—while managing hallucinations, privacy, and overreliance.AI that can create new writing or images that didn’t exist before.
Machine Learning (ML)Core ConceptSubset of AI where systems learn from examples. Helps clinicians frame AI as statistical learning, not “understanding.”A way computers learn from many examples instead of being told every rule.
Training DataCore ConceptThe dataset used to train a model/tool. Clinicians should ask what it includes/excludes and whether it matches their client populations and use case.The specific information used to teach the AI how to do its job.
Adversarial AttacksEthics / SafetyAI can be tricked by intentional inputs. Relevant for vendor security, data integrity, and risk assessment for clinical tools.A sneaky way to trick an AI into doing the wrong thing.
BiasEthics / SafetySkewed outputs from unrepresentative data or design choices. Important for equity, cultural humility, and avoiding harm in AI-assisted recommendations.When the AI is unfair because what it learned from wasn’t balanced.
Explainability (XAI)Ethics / SafetyMethods that provide partial, testable reasons for outputs (not full “mind-reading”). Supports accountability and safer clinical use.Ways to get the AI to show *clues* about why it answered the way it did, so we can check it.
PrivacyEthics / SafetyCore legal/ethical issue: what data is shared with third parties, how it is protected, and whether proper agreements/consents exist.Keeping personal information safe and not shared without permission.
Regulation / OversightEthics / SafetyLaws, professional guidance, and organizational policies governing AI use in healthcare and documentation workflows.Rules and monitoring to make sure AI is used safely and fairly.
SentienceEthics / SafetyHelpful for boundary-setting: current AI can sound empathic without experiencing feelings or selfhood.Truly feeling and being aware like a person—current AI does not have this.
Sentience HaloEthics / SafetyRisk that users infer consciousness and trustworthiness from fluent language. Can drive overreliance or attachment.When an AI sounds so human that people start to think it truly understands or feels.
ChatbotLLM-SpecificInterface clients commonly use. Requires guidance on limits, verification, and appropriate use cases.A computer program you can talk with—usually by typing.
Fine-TuningLLM-SpecificAdapts a general model to a specific domain (e.g., mental health). Useful for evaluating whether a tool is specialized and how it was trained.Extra training that makes a general AI better at one specific job.
HallucinationLLM-SpecificConfident but incorrect output. Clinicians should assume outputs may be wrong and require verification—especially for citations, diagnoses, and policy.When the AI makes up facts that sound real but aren’t true.
Large Language Model (LLM)LLM-SpecificUnderlying technology for many chatbots. Important to explain that models generate text by prediction, not by guaranteed factual recall or real-time access.An AI trained on huge amounts of writing that predicts the next words to create answers.
Reinforcement Learning from Human Feedback (RLHF)LLM-SpecificTechnique shaping helpful/safe behavior and “empathetic-sounding” responses. Not evidence of real emotion; supports psychoeducation about boundaries.Humans rate the AI’s answers, and the AI learns to give the kinds of answers people rate as helpful and safe.
Business Associate Agreement (BAA)Operations / ComplianceKey vendor requirement when a tool handles protected health information (PHI). Helps determine whether a service is appropriate for clinical use.A legal agreement that says a company must protect health information if they handle it for you.
De-identification vs AnonymizationOperations / ComplianceClinicians often confuse these. De-identified data may still carry re-identification risk; policies should define what is removed and what remains.Removing identifying details is not always the same as making it impossible to identify someone.
Minimum Necessary / Data MinimizationOperations / ComplianceLimits what information is disclosed to tools/vendors. Reduces exposure if systems are breached or misused.Only share the smallest amount of information needed to do the job.
Retention & DeletionOperations / ComplianceWhere data is stored, for how long, and how it can be deleted. Essential for consent, policy, and vendor review.How long the tool keeps the information and whether you can delete it.
Human-in-the-Loop ReviewOperations / ComplianceClinician remains responsible for accuracy and clinical judgment. A safeguard against hallucinations and inappropriate summaries.A human checks and approves what the AI produces before it’s used.
Audit Logs & Access ControlsOperations / ComplianceTracks who accessed data and when; limits access to authorized users. Core components of security governance.Records of who looked at information and settings that limit who can see it.