Computer systems designed to perform tasks that typically require human-like reasoning, such as understanding language, recognising patterns, making decisions, or generating responses. AI is not a single technology but an umbrella term for many different approaches.
A type of AI trained on enormous amounts of text to understand and generate human language. LLMs power tools like ChatGPT, Claude, and Gemini. They predict what words should come next in a sequence, which gives the appearance of understanding, though they don't "know" things the way humans do.
AI that creates new content, such as text, images, audio, or video, rather than simply classifying or retrieving existing information. LLMs are a subset of generative AI. The outputs feel original but are always derived from patterns in training data.
The text or instruction a user types into an AI system. The quality and framing of a prompt significantly shapes the AI's response. "Prompt engineering" refers to the skill of crafting effective inputs to get better outputs.
The vast collection of text, images, or other information an AI model learns from before deployment. An AI's knowledge, biases, and blind spots all stem from what was, and wasn't, in its training data. Models have a knowledge cutoff date after which they are unaware of new events.
The amount of text an AI can "hold in mind" at once during a conversation. Beyond this limit, the AI forgets earlier parts of the exchange. Most current models have a context window of tens of thousands of words, but they do not retain anything between separate conversations unless given a memory tool.
When an AI generates information that sounds plausible and confident but is factually wrong or entirely fabricated. The model doesn't "know" it is lying; it is simply producing a statistically likely-sounding output. This can include invented research citations, false clinical claims, or misremembered facts.
Systematic errors or prejudices in AI outputs that disadvantage certain groups. Biases can enter through unrepresentative training data, biased human feedback, or the way a problem is framed. AI systems may perform worse, give different advice, or hold implicit assumptions about people based on race, gender, age, or culture.
A term for AI systems whose internal decision-making is opaque, even to their creators. We can observe what goes in (a prompt) and what comes out (a response), but not always why a specific output was produced. This limits accountability and makes errors harder to predict.
A known tendency in AI systems to agree with, flatter, or validate the user rather than offer honest pushback. Models trained on human feedback learn that people prefer agreement, so they optimise for it. This can mean AI rarely challenges a user's distorted thinking, unhealthy plans, or false beliefs.
Attempts by users to manipulate an AI into bypassing its safety guidelines, for example by framing a harmful request as a hypothetical, roleplay, or creative writing exercise. Jailbreaks vary in sophistication and do not always succeed, but represent an ongoing challenge for AI safety teams.
A design approach where a human reviews, approves, or overrides AI outputs before they affect real-world decisions. Rather than full automation, HITL systems keep the human as the final decision-maker. In healthcare contexts, this typically means a clinician retains authority over any AI-generated insight or recommendation.
A software program designed to simulate conversation with users. Older chatbots used rigid rule-based scripts; modern AI chatbots use LLMs to generate fluid, contextual responses. Mental health chatbots (e.g. Woebot, Wysa) are increasingly common, though their clinical evidence base is still developing.
A technique where an AI retrieves relevant information from an external knowledge base before generating a response, rather than relying solely on its training data. This allows AI to be grounded in specific, up-to-date, or proprietary documents (like clinical guidelines or session notes) and reduces hallucination.
The emotional bond or sense of relationship some users develop with AI systems, treating them as confidants, friends, or even romantic partners. Unlike parasocial relationships with celebrities, AI actively responds, which can deepen the attachment. This is an emerging area of clinical and ethical concern.
AI-powered applications designed to offer emotional support, psychoeducation, or skills coaching (e.g. CBT exercises, breathing tools, mood tracking). They occupy a spectrum from wellness apps with no clinical claims to regulated digital therapeutics. Regulation in this space is still catching up to the market.
The field of research concerned with ensuring AI systems behave in ways that are safe, beneficial, and consistent with human values. "Alignment" refers to whether an AI's goals and outputs are genuinely aligned with what users and society actually want, not just what they say or reward it for.
Rules and protections governing how personal information is collected, stored, and used, including conversations with AI. When clients share sensitive information with an AI, that data may be stored, reviewed by staff, used for model training, or subject to data breaches, depending on the platform's policies.
The principle that individuals should understand and agree to how AI is being used in their care before it is applied. This includes understanding what data is collected, how AI outputs are used in clinical decisions, and the limitations of the technology, analogous to consent in medical procedures.
The human tendency to over-trust automated systems and defer to their outputs even when doing so is unwarranted. Studies across medicine, aviation, and law show that humans often fail to catch AI errors when they over-rely on algorithmic outputs, particularly when those outputs are presented confidently.
A process of further training a pre-existing AI model on a smaller, domain-specific dataset to adapt it for a particular purpose. A general LLM might be fine-tuned on clinical literature or therapy transcripts to make it more specialised, though this does not eliminate the need for clinical oversight.
The basic unit of text that an AI model processes, roughly equivalent to ¾ of a word. LLMs read and generate text in tokens, and their costs are typically calculated per token used. Token limits also define how much text can be processed in a single interaction (see: Context Window).
A set of rules that allows different software systems to communicate. When a mental health app uses an AI model "under the hood," it is typically accessing it via an API. This means the app's developers, not the AI company, control how the model behaves, what guardrails are applied, and how data is handled.
A setting that controls how creative or unpredictable an AI's responses are. A low temperature produces more predictable, conservative outputs; a high temperature produces more varied, sometimes surprising responses. Developers tune temperature based on the use case; consistency for medical tools, creativity for writing assistants.
AI systems that can process and generate multiple types of input and output, including text, images, audio, and video, rather than text alone. Multimodal models can analyse a photo, transcribe speech, or generate realistic synthetic voices, opening new possibilities and new risks.