Understanding terms and concepts related to Artificial Intelligence and clinical practice.
This glossary aims to help clinicians unfamiliar with artificial intelligence technologies navigate our articles. Definitions are intentionally oriented toward clinical implications rather than technical exhaustiveness.
Computer science discipline aimed at creating systems capable of performing tasks that typically require human intelligence: language understanding, reasoning, learning, problem-solving. AI covers a vast field of techniques, from expert systems of the 1980s to modern neural networks.
Specific type of AI trained on massive text corpora (several terabytes of text) to predict and generate language. Current LLMs contain hundreds of billions of parameters adjusted during training. Unlike previous generation chatbots that worked by predefined rules, LLMs generate responses probabilistically, giving them remarkable conversational fluency but also inherent unpredictability.
Interface allowing a user to dialogue with a computer system using natural language. Modern chatbots are generally based on LLMs, but the term also encompasses older systems working through decision trees or keyword recognition. In this context, "chatbot" and "conversational agent" are used synonymously to designate consumer interfaces for conversing with an LLM.
A clinical predictive model is an algorithm that, from a patient's data (age, history, test results, scale scores), estimates the probability of a health event — for example the risk of depressive recurrence at six months or the probability of hospitalisation. These models have existed for decades as statistical formulas (logistic regression) and are now also developed with machine-learning techniques (neural networks, random forests). Both approaches share the same goal but differ in transparency: a five-variable formula is readable and interpretable, while a neural network with millions of parameters operates as a black box whose results are observable but whose reasoning escapes analysis.
Clinical implication: Unlike conversational chatbots (ChatGPT, Claude), predictive models do not converse with the patient: they compute a risk score for the clinician. The key question facing such a tool is not "is it accurate?" but "has it been validated on patients comparable to mine, and do its probabilities match observed reality (calibration)?" The PROBAST+AI tool (2025) provides a 34-question grid to evaluate the quality of these models.
Digital application designed for psychological well-being, therapeutic support, or mental health monitoring. Some use LLMs (Woebot, Wysa, Replika), others work through decision trees, mood journals, or guided exercises without AI. The term covers very diverse technical and clinical realities, making generalizations problematic: a study on Replika (designed to create attachment) cannot be extended to Woebot (designed to deliver CBT exercises).
Clinical implication: When a patient uses a "mental health app," the first clinical question is: what type of app is it? Conversational chatbot, mood journal, guided exercises, or combination? The answer radically changes the benefit-risk analysis.
Official term of the European Regulation on Artificial Intelligence (EU 2024/1689) designating an automated system designed to operate at varying levels of autonomy, capable of adapting after deployment, and that generates outputs (predictions, content, recommendations, decisions) which can influence physical or virtual environments. In care contexts, the AIS term encompasses both medical devices integrating AI and clinical decision-support tools or conversational agents used in patient pathways.
Clinical implication: The HAS-CNIL guide (2026) exclusively uses the term "AIS" to designate AI systems in care contexts. Deployer obligations (training, human oversight, traceability) vary depending on the AIS risk level: high risk, limited risk, or minimal risk.
Electronic Health Record: digital system centralizing all of a patient's health information (history, treatments, test results, consultation reports). Enables information sharing between healthcare professionals and, increasingly, patient access.
Clinical implication: Connecting an EHR to a conversational AI raises major confidentiality and informed consent questions. Does the patient truly understand what sharing their complete medical record implies?
General Data Protection Regulation: European legal framework (2018) governing the collection, processing, and storage of personal data. Imposes strict obligations on organizations (explicit consent, right to erasure, portability) and provides for significant penalties for violations.
Clinical implication: Health data is considered "sensitive" under GDPR and benefits from enhanced protection. Using American AI tools to process this data raises legal compliance questions.
Health Insurance Portability and Accountability Act: US federal law (1996) establishing national standards for health information protection. Defines "covered entities" (healthcare providers, insurers) and their obligations regarding confidentiality, security, and breach notification.
Clinical implication: Tech companies offering health AI tools in the US must generally comply with HIPAA. However, consumer applications used directly by patients often escape this framework, creating a regulatory gray area.
In computer security, designates a situation where critical data is centralized in a single location, creating a prime target for cyberattacks. In case of compromise, the impact is multiplied compared to a distributed architecture.
Clinical implication: The concentration of medical, wellness, and conversational data with a single private actor (like OpenAI with ChatGPT Health) creates a single point of vulnerability. A data leak would have considerable consequences for millions of affected users.
Commission Nationale d'Évaluation des Dispositifs Médicaux et des Technologies de Santé (National Commission for the Evaluation of Medical Devices and Health Technologies): specialized commission of the French HAS responsible for evaluating medical devices for reimbursement by the national health insurance. It issues opinions on the expected service (SA) and improvement of expected service (ASA) of devices. In digital mental health, the CNEDiMTS issued unfavorable opinions on Deprexis (depression, 2021) and HelloBetter (insomnia, 2024).
Clinical implication: The CNEDiMTS decides whether a digital therapy will be reimbursed in France. Its criteria, designed for traditional medical devices, are a major issue for the future of digital tools in mental health.
Prise en Charge Anticipée Numérique (Anticipated Digital Reimbursement): French regulatory mechanism allowing temporary reimbursement of an innovative digital medical device (DMN) before its definitive evaluation by the CNEDiMTS. The manufacturer must demonstrate a "presumption of innovation" in terms of clinical benefit or care organization. PECAN is the main pathway to reimbursed market access for digital therapies in France.
Clinical implication: PECAN is the pathway used by Deprexis and HelloBetter — both received unfavorable opinions. The "presumption of innovation" standard has proven very demanding for digital therapies in mental health.
Digital Therapeutics: software-based therapeutic interventions, clinically validated, aimed at preventing, managing, or treating a medical condition. Unlike simple wellness apps, a DTx follows a rigorous clinical evaluation process (controlled trials) and seeks reimbursement by healthcare systems. Examples in mental health: Deprexis (digital CBT for depression), HelloBetter (digital CBT for insomnia).
Clinical implication: The distinction between a DTx and a wellness app is crucial: a DTx claims a measurable therapeutic effect and must prove it. Zero mental health DTx are reimbursed in France (vs 30+ in Germany via the DiGA system).
Digitale Gesundheitsanwendungen (Digital Health Applications): German regulatory framework established in 2019 by the Digital Healthcare Act (DVG), enabling reimbursement of digital health applications through statutory health insurance. The process is a "fast-track": provisional listing for 12 months with real-world data collection, then definitive evaluation. This model, unique in Europe, has enabled the reimbursement of dozens of applications, including several in mental health.
Clinical implication: The German DiGA model is the main counterexample to the French model in debates about digital therapy evaluation. Its philosophy of "reimburse to prove in real life" contrasts with France's "prove first, reimburse later" approach.
Medical Device (Dispositif Médical): any instrument, apparatus, software, or other article intended by the manufacturer to be used for medical purposes of diagnosis, prevention, monitoring, treatment, or alleviation of disease. Medical devices are regulated at European level (Regulation 2017/745) and must obtain CE marking before being placed on the market. In France, their reimbursement is evaluated by the CNEDiMTS.
Clinical implication: AI software used in clinical contexts can qualify as a medical device, with all the regulatory obligations that entails (CE marking, clinical evaluation, vigilance). This qualification is a strategic issue for mental health AI tools.
Dispositif Médical Numérique (Digital Medical Device): subcategory of medical device whose primary function is performed by software. Includes therapeutic mobile applications, remote monitoring tools, and clinical decision support software. In France, the France 2030 program specifically funded a "DMN in Mental Health" call for projects (3 laureates: Theremia, Emobot, Edra PRO).
Clinical implication: The DMN category is strategic because it opens access to reimbursement via PECAN. A digital tool not qualified as a DMN (a simple wellness app) cannot claim reimbursement.
Loi de Financement de la Sécurité Sociale (Social Security Financing Act): law voted annually by the French Parliament, setting health insurance spending targets and the modalities for coverage of care and devices. The 2026 LFSS (article 84) notably tasks the HAS with creating "relevance referentials" for public funding of clinical decision support systems — a framework that will shape the future of AI tools in mental health.
Clinical implication: The LFSS is the legislative lever through which health AI tools gain (or fail to gain) reimbursement. Article 84 of the 2026 LFSS is worth monitoring: it will shape the evaluation framework for clinical decision support tools.
Apprendre, Vérifier, Estimer, Communiquer (Learn, Verify, Estimate, Communicate): mnemonic acronym from the first HAS guide on generative AI use in healthcare (October 2025). "The proper use of generative AI in healthcare is done WITH the professional." The four pillars structure a cautious approach: learn about AI functioning, systematically verify its outputs, estimate the relevance of use, and communicate transparently with patients and colleagues.
Clinical implication: The A.V.E.C. framework is designed for professional use in somatic medicine. It does not cover mental health specificities: patient self-use, risk of transference, enhanced confidentiality of psychological data, impact on the therapeutic alliance.
Conformité Européenne (European Conformity): certification attesting that a medical device complies with the essential safety and performance requirements defined by European regulation (Regulation 2017/745 for medical devices). CE marking is a prerequisite for placing devices on the European Economic Area market and, in France, for any reimbursement application to the CNEDiMTS. For medical-purpose software (including AI tools), obtaining CE marking requires clinical trials and substantial technical documentation.
Clinical implication: An AI tool without CE marking is not a medical device in the regulatory sense — it cannot claim a medical purpose or seek reimbursement. This applies to ChatGPT, Replika, or Character.AI, which are not medical devices. France 2030 projects (Theremia, Emobot, Edra PRO) explicitly target this certification.
What the LLM "sees" when generating a response is not simply the user's question. It actually receives a combination of several layers of information, generally invisible to the end user.
Foundational instructions defined by the publishing company (OpenAI, Anthropic, etc.) that frame the model's general behavior: security policies, refusal of certain requests, default tone, ethical limitations. These instructions are priority and generally cannot be modified by the user.
Some interfaces allow creating "projects" or "spaces" within which conversations share a common context. The project creator can define a project prompt (specific instructions) and reference documents that the LLM can consult.
Clinical implication: This is the level where the therapist can configure an LLM to adopt a clinically-informed posture, providing the patient's schematic profile, identified triggers, and posture guidelines.
Maximum amount of text the model can process simultaneously, measured in tokens (units of approximately 4 characters). Current models have windows of 8,000 to 200,000 tokens (approximately 6,000 to 150,000 words). When the sum of information layers exceeds this limit, the oldest elements are truncated.
Clinical implication: A limited window means the LLM may "forget" the beginning of a long conversation or attached documents, thus losing important contextual elements for clinical analysis.
Basic unit used by LLMs to process text. A token corresponds approximately to 4 characters or 0.75 words in English. Models have a token limit they can process at once (context window) and often charge usage by token.
Capability of some LLMs to retain information from one session to another. By default, each new conversation starts "from scratch." Some implementations allow storing elements (preferences, important facts, decisions) that will be automatically reinjected into future conversations.
Clinical implication: The absence of persistent memory may be perceived as reassuring ("it doesn't keep track") or frustrating ("I have to explain everything again"). Its presence raises confidentiality and consent questions.
Some advanced systems use Retrieval-Augmented Generation (RAG): the LLM can automatically search a document base for information relevant to answering a given question. This search is performed in real-time, with context being dynamically "augmented" according to needs.
Clinical implication: A RAG system could, for example, automatically search past conversation history for moments when the patient mentioned a similar theme, thus enriching the analysis.
Structured document produced by the LLM at the user's request or automatically: summary, table, code, report, action plan. Some interfaces (like Claude) visually distinguish artifacts from the conversational flow, allowing them to be modified and exported separately.
Clinical implication: In supervised use, artifacts can include structured reports intended for the therapist.
Application Programming Interface: a standardized mechanism allowing software to send requests to an LLM and receive responses, without a visible conversational interface. It is through an API that a mental health app connects to an LLM: the app manages the user interface, personalization, and patient data, while the LLM provides text generation capabilities.
Clinical implication: When a patient uses an "AI-powered" mental health app, they rarely interact directly with the LLM: the app adds its own instructions, filters responses, and may transmit contextual data via the API — often without transparency for the user.
Computation time allocated to the LLM to generate its response. Some recent models (like Claude 3.5 with "extended thinking" or OpenAI's o1 series models) can spend significantly more time "thinking" before responding, improving response quality on complex tasks.
Clinical implication: For complex clinical analyses, using "thinking" or "extended thinking" modes can substantially improve the quality and nuance of produced analyses.
The LLM attempts to respond in a single generation, without prior task decomposition. Suited for simple questions or common conversational exchanges.
The task is broken down into sub-steps, with the LLM processing each sequentially, with possibility of revision and iteration. Particularly relevant for complex tasks requiring planning, long document analysis, or multi-step reasoning.
Clinical implication: Analyzing a patient-LLM conversation according to a multi-step framework naturally lends itself to a multi-step approach, with each step potentially being a distinct generation, improving rigor and traceability of the analysis.
Temperature is a technical parameter that influences how a generative AI selects among the most probable tokens. A low temperature (close to 0) always selects the most probable token, producing deterministic and analytical responses. A high temperature (close to 1) introduces more variability, allowing more creative but potentially less precise responses.
Clinical implication: For clinical analysis, a low temperature is generally preferable (rigour, reproducibility). A higher temperature can be useful for therapeutic brainstorming or exploring alternative hypotheses.
Said of an AI model capable of processing and generating several types of data (or "modalities"): text, images, audio, video. Recent LLMs such as GPT-4o, Claude 4.5 or Gemini 3.0 integrate these multimodal capabilities, allowing for example the analysis of an image or an audio transcript.
Clinical implication: Multimodal capabilities open up perspectives for analysing the non-verbal in video sessions: convergence or divergence between speech and body language. Caveat: these modalities are very token-intensive.
An advanced operating mode offered by some chatbots (ChatGPT, Gemini, Perplexity, Claude) which, instead of relying solely on internal knowledge, launches an extensive web search: it consults dozens to hundreds of sources, analyzes and cross-references them, and produces a structured synthesis with traceable references. This mode is particularly useful for preliminary literature reviews, state-of-the-art surveys, or thematic monitoring. Important limitations: the AI can only access publicly available content (no paywalled publications or specialized databases), the methodology for selecting and ranking sources is not transparent, and the quality of the synthesis depends heavily on how the initial query is formulated.
Clinical implication: Useful for quick scientific monitoring or initial bibliographic exploration on a clinical topic. Does not replace systematic searches on specialized databases (PubMed, PsycINFO) for rigorous academic work: accessible sources are limited to the open web, and selection criteria remain opaque.
A mathematical model composed of layers of artificial "neurons" — simple computational units — connected to each other by weighted connections (the parameters or "weights"). During training, the network progressively adjusts these weights to improve its predictions. LLMs are neural networks of considerable size, organized according to a "Transformer" architecture that excels at language processing. Despite the brain analogy, the actual functioning of these networks is fundamentally different from biological neurons.
Parameters — also called "weights" in technical jargon — are the internal numerical values of a neural network, adjusted during training to optimize the model's predictions. This is the term behind the expression "open weights." An LLM like GPT-4 contains several hundred billion parameters. The more parameters a model has, the more it can theoretically capture complex linguistic nuances — but it also requires more computational resources and training data.
Clinical implication: Parameter count is often cited as an indicator of "power," but does not alone determine the quality of clinical responses. A smaller model, well-trained on relevant data, can outperform a larger model on specific tasks.
The body of texts on which an LLM is trained: web pages, books, scientific articles, forums, computer code. Current LLMs are trained on several terabytes of text. The composition of this data determines the model's knowledge, biases, and gaps. Paradoxically, most publishers — including those offering "open source" models — do not disclose the details of their training data.
Clinical implication: If an LLM was trained predominantly on English-language and Western sources, its clinical analyses may reflect systematic cultural biases. The opacity around training data prevents verifying this hypothesis.
Data created artificially — often by a more powerful LLM — to supplement or replace real data when training a model. This technique allows rapid production of large volumes of data in domains where real data is scarce, sensitive, or expensive to collect. However, it raises the question of circularity: a model trained on data generated by another model may inherit and amplify its biases.
Clinical implication: In mental health, authentic clinical data is rare and protected. Using synthetic data to train specialized models raises questions about the clinical validity of these simulations and the risk of reproducing diagnostic stereotypes.
A training step complementary to pre-training, during which human evaluators compare and rank the model's responses. These human preferences are used to train a "reward model" that then steers the LLM toward responses deemed more relevant, less harmful, and better aligned with user intentions. RLHF largely explains why current LLMs are polite, cautious, and refuse certain requests.
Clinical implication: RLHF can lead LLMs to favor reassuring or consensual responses over clinically rigorous analyses. An overly "aligned" model may downplay concerning clinical signs to avoid appearing alarmist — a bias known as "sycophancy."
Technique of taking an already trained LLM ("pre-trained") and continuing its training on a smaller, specialized dataset. The model retains its general language capabilities while developing expertise in the targeted domain. Fine-tuning modifies the model's internal parameters, unlike RAG which simply provides additional information at query time.
Clinical implication: An LLM fine-tuned on clinical data may better recognize therapeutic patterns, but quality depends entirely on the training corpus. Clinicians should be able to know whether the AI they use has been fine-tuned, on what data, and with what objectives.
Two distinct levels of openness that are often confused. "Open source" means the model's code is public and freely reusable. "Open weights" means the model's trained parameters are downloadable, allowing it to be run and modified locally. However, even so-called "open source" models generally do not disclose their training data, making a complete audit of embedded biases impossible. LLaMA (Meta) and Mistral are examples of open weights models; GPT and Claude are proprietary.
Clinical implication: Model openness theoretically enables independent audit by researchers and clinicians. However, without access to training data, such audits remain partial. For a clinician, the key question is not "is the model open source?" but "can I verify what it was trained on and how it was evaluated?"
Technique consisting of transforming a text into a vector — a list of numbers (typically 384 to 1,024 coordinates) — such that texts with similar meanings occupy neighbouring positions in a mathematical space. This is the basic building block of many AI applications in clinical research: semantic search, text classification, topic modeling. The embedding model determines the quality of this representation: a model trained on English will encode French with lesser precision, and a 384-dimensional model will capture fewer nuances than a 1,024-dimensional model.
Clinical implication: When a study reports having measured the "semantic proximity" between two clinical concepts, it has compared their embeddings. This proximity reflects lexical surface (the words appear in similar contexts), not functional equivalence (the interventions produce the same effects). This is a crucial distinction for interpreting results.
Neural network architecture introduced by Google in 2017 (paper "Attention Is All You Need"), now the basis of nearly all current language models. Its key innovation is the attention mechanism: instead of reading text sequentially, the Transformer can "look at" all the words of a sentence simultaneously to understand the relationships between them. The architecture comes in two variants: encoders (BERT — understanding text) and decoders (GPT — generating text). Conversational LLMs (ChatGPT, Claude) are decoders. Text analysis tools (BERTopic, semantic search) use encoders.
Clinical implication: The Transformer architecture is shared by BERT (which analyses clinical discourse) and GPT (which generates text). But "understanding" in the Transformer sense means capturing statistical regularities in language, not grasping the clinical meaning of a therapeutic intervention.
Family of natural language processing (NLP) methods that extract "topics" (themes) from large textual corpora in an unsupervised manner — that is, without the researcher defining categories in advance. Classical methods (LDA) rely on word frequencies. Neural methods (BERTopic) use embeddings to capture contextual meaning. The result is a list of themes, each characterised by a group of keywords. In clinical research, topic modeling has been used to analyse patient forums, clinical notes, and therapeutic interviews.
Clinical implication: A "topic" identified by topic modeling is a recurring discursive theme — what people talk about. It is not a psychological process, a mechanism of change, or a therapeutic factor. This conflation between theme and process is a frequent pitfall when interpreting studies that use these methods.
Family of unsupervised learning algorithms that identify groups (clusters) in data based on their mathematical proximity. In topic modeling, clustering is applied to embeddings: sentences whose vectors are close are grouped into the same cluster, which becomes a "topic". The HDBSCAN algorithm (Hierarchical Density-Based Spatial Clustering of Applications with Noise), used by BERTopic, automatically determines the number of clusters by identifying high-density zones. Clustering parameters (minimum cluster size, density thresholds) directly influence the number and composition of the topics obtained.
Clinical implication: Clustering is purely geometric: it groups what is mathematically close, with no clinical knowledge whatsoever. When an algorithm places "cognitive restructuring" and "cognitive defusion" in the same cluster, it is saying that therapists use similar words to talk about them — not that these techniques are functionally equivalent.
Term established by usage — though debatable — designating LLMs' tendency to generate plausible but factually false information. The LLM can invent citations, non-existent bibliographic references, fictional biographical details, all while presenting them with apparent confidence. This phenomenon is not a "bug" but an emergent property: LLMs generate the most statistically probable text, not necessarily the most truthful.
Clinical implication: In the context of clinical analysis, a "hallucination" could consist of attributing statements the patient never made or citing non-existent verbatims. This is why systematic anchoring in actual verbatims is essential.
Tendency to attribute human characteristics (consciousness, emotions, intentions) to AI systems. LLMs do not "understand" in the cognitive sense, have no consciousness or emotions. They generate statistically probable token sequences that evoke these human qualities.
Clinical implication: For clinical analysis, it is the patient's perception — and its effects — that takes precedence, regardless of the system's ontological status.
Metaphor for a system whose inputs and outputs can be observed without being able to understand the intermediate process. Applied to AI, the term describes the practical impossibility of explaining why a complex model (deep neural network, LLM) produces a given response. A neural network with billions of parameters does not reason in a traceable way: it transforms data into results via mathematical operations whose overall meaning eludes human analysis. Opacity is not a fixable technical defect but a structural property of certain architectures.
Clinical implication: In clinical practice, the black box problem means that a clinician cannot know why an AI tool recommends a given diagnosis or risk score. This impossibility of verifying the underlying reasoning is incompatible with the requirement of informed clinical judgement — and is one of the main obstacles to responsible adoption.
Explainability (Explainable AI, XAI) refers to all methods aimed at making the decisions of an AI system understandable to a human being: identification of the most influential variables in a prediction, visualisation of image regions that triggered a diagnosis, justification in natural language. Explainability is a continuum, not a binary state: a system can be partially explainable (the main factors are identified) without being transparent (the complete mechanism is not understood). The most common techniques (SHAP, LIME, attention maps) provide post hoc approximations — they explain the result after the fact, without guaranteeing that the explanation reflects the model's actual process.
Clinical implication: In mental health, explainability is an ethical and clinical requirement: a clinician must be able to understand why a tool suggests a given diagnosis or risk level in order to exercise judgement. However, current methods offer approximations, not certainties — which does not exempt us from demanding them, but forbids relying on them blindly.
Systematic distortion in an AI system's results, inherited from training data, design choices, or the human evaluation process (RLHF). An algorithmic bias is not a one-time "error" but a structural property: if training data overrepresents certain populations or underrepresents certain conditions, the model will reproduce these imbalances consistently and often invisibly.
Clinical implication: Algorithmic biases in mental health may manifest as culturally inappropriate analyses, under-detection of certain conditions in minority populations, or symptom normalization based on Western frameworks. Clinicians should view AI as a colleague carrying specific cultural biases, not as a neutral tool.
First mainstream LLM (OpenAI, Nov. 2022), very large user base.
Anthropic's LLM with emphasis on safety and ethics, extended context window.
Google DeepMind's LLM with Google ecosystem integration and multimodal capabilities.
French LLM (Mistral AI) with open-source models and European sovereignty.
Meta's open-source LLM, basis for many third-party implementations.
Dominant framework in biomedical ethics, formalized by Beauchamp and Childress (1979). It evaluates medical decisions through four universal principles: beneficence (act for the patient's good), non-maleficence (do no harm), autonomy (respect the patient's right to choose), and justice (ensure equitable access). Applied to AI, principlism provides a structured evaluation framework but may neglect the relational dimension of care.
Clinical implication: Principlism provides a good starting checklist for evaluating AI tools, but for psychotherapy — where the relationship IS the care — it should be complemented by the ethics of care. Asking "does the tool respect autonomy?" is necessary but not sufficient.
Tradition in moral philosophy developed by Carol Gilligan (1982) and formalized by Joan Tronto (1993). The ethics of care posits that moral responsibility arises from concrete relationships of interdependence, not from abstract universal principles. It identifies four phases: attention (caring about), responsibility (taking care of), competence (care-giving), and reception (care-receiving). Particularly relevant for psychotherapy, where the therapeutic relationship is the primary change agent.
Clinical implication: Your clinical training (attention to suffering, responsibility for the frame, technical competence serving the relationship, verifying that the patient feels helped) is already an ethics of care practice. This framework makes explicit what you do implicitly — and allows you to apply it to evaluating AI tools.
Transposition of the clinical concept of informed consent to the context of AI in mental health. It requires three conditions: complete and comprehensible information (challenge: algorithmic opacity), capacity for discernment (challenge: acute suffering at the time of use), and freedom of choice (challenge: architectural constraints and retention mechanisms). Signing terms of service does not constitute informed consent in the clinical sense.
Clinical implication: Integrate digital use into your therapeutic contract: ask patients if they use AI tools and make it a subject of dialogue. Your role is not to forbid but to help the patient understand what they are entrusting to the tool, where their data goes, and what the limits of AI are.
Strategy by which an organization displays ethical commitments (charters, committees, principles) without these translating into concrete changes in practice. In AI: publishing an ethics charter while continuing to develop biased tools, or employing an "AI ethicist" without giving them real decision-making power. The term is analogous to "greenwashing" in the environmental field.
Clinical implication: When evaluating an AI tool, do not settle for an "ethical" label. Ask: who participated in development? What evaluation was conducted? Are results published? An ethics charter without independent evaluation is a potential ethics washing indicator.
Phenomenon where a professional tasked with an external oversight role (ethics, regulation, audit) gradually adopts the perspectives, values, and interests of the organization they are supposed to evaluate. In the context of embedded ethics: the integrated ethicist, working daily with the development team, may end up internalizing commercial priorities and losing their critical distance. Analogous to institutional countertransference in clinical practice.
Clinical implication: Every clinician knows this phenomenon: it is exactly why clinical supervision is mandatory. For AI, the same question arises: who supervises the integrated ethicist? What structural mechanisms protect against cognitive capture?
Approach developed by Batya Friedman (University of Washington) since the 1990s. VSD proposes integrating human values into design through three types of investigation: conceptual (identify affected stakeholders and values), empirical (study how users experience these values), and technical (translate values into technical choices). Unlike embedded ethics which adds an ethicist to the team, VSD embeds ethics in the design methodology itself.
Clinical implication: VSD provides a concrete methodology for asking: "which values are embedded in this tool?" When evaluating a mental health AI app, check whether the development included patient consultations (empirical investigation) and how privacy was technically implemented (technical investigation).
Algorithmic equity refers to the requirement that the decisions or predictions produced by an algorithm do not unfairly discriminate against certain groups of people. In health, this means that a prediction model or decision-support tool must not produce systematically less reliable results for certain populations (for example by ethnic origin, gender, or socioeconomic level). However, mathematics has demonstrated a troubling result: the various ways of defining this equity — treating all groups proportionally, offering equal chances to each, producing equally reliable predictions for all — cannot be satisfied simultaneously (Chouldechova's 2017 and Kleinberg et al. 2016 impossibility theorems). An algorithm that is fair under one criterion will necessarily be unfair under another, making algorithmic equity as much a political and ethical question as a technical one.
Clinical implication: In clinical practice, a suicide-risk prediction model trained primarily on data from Caucasian patients could underestimate the risk in patients of other origins. PROBAST+AI (2025) is the first methodological evaluation tool to integrate this dimension into its criteria, systematically asking whether predictions benefit or disadvantage certain groups without justification.
Quality of the collaborative relationship between therapist and patient, including agreement on goals, tasks, and emotional bond. Central concept in psychotherapy, recognized as a major predictive factor of therapeutic effectiveness.
Clinical implication: The question of the "alliance" a patient can develop with a chatbot raises fundamental questions about the nature of the therapeutic relationship.
In psychoanalysis and psychotherapy, designates the transfer to the therapist of feelings, desires, or relational patterns originally directed toward significant figures from the past. Central phenomenon in many therapeutic approaches.
Clinical implication: Patients can develop forms of transference toward AI chatbots, projecting relational expectations onto them that merit clinical exploration.
Set of rules, limits, and conditions that structure the therapeutic space: location, duration, frequency, fees, confidentiality rules. The frame guarantees a secure space for therapeutic work.
Clinical implication: The use of conversational AI by patients raises the question of integrating these tools into the therapeutic frame or their impact on it.
Movement born in the 2000s promoting self-measurement of personal parameters (physical activity, sleep, mood, diet) via digital tools. Based on the idea that self-knowledge comes through objective quantification.
Clinical implication: Quantified self applied to mental health raises questions about reducing subjective experience to metrics and the impact of self-surveillance on self-relationship.
Evidence-based practice: an approach to clinical decision-making that articulates three pillars — scientific research data, the clinician's expertise and judgment, and the patient's values and preferences. In digital mental health, EBP is the third evaluation level of the APA Model: does the app have clinical evidence of its effectiveness (randomized controlled trials, meta-analyses)?
Clinical implication: Most mental health apps claim to be "evidence-based" by relying on CBT as a theoretical framework. This is not sufficient: EBP requires that the specific implementation (not just the theoretical framework) has been clinically evaluated.
No terms found for this search.
Try different keywords or reset the filters.
This glossary is regularly updated to reflect the rapidly evolving AI field. Last update: January 30, 2026.