Glossary

AgentExplained 3 ways: An AI system set up to take actions toward a goal, not just answer a question. An agent can plan steps, use tools, check the result, and try again, often with little or no input between steps. This is the noun behind agentic AI. The distinction that matters for care and for governance is action: a tool that advises leaves the decision and the act with a person, while an agent does the act, which changes who is accountable when something goes wrong and what oversight the system needs.
Agentic AIExplained 3 ways: AI that does not just produce an output you act on, but takes actions on its own: it sets a goal, plans steps, uses tools, and adapts, with a human reviewing at checkpoints rather than every step.
AI fluency: The working ability of a healthcare professional to use AI well in their own job: knowing what a tool is for, how it was trained and tested, where it tends to fail, when to override it, and how to explain its role to a patient. Most frameworks, and the EU AI Act, call this AI literacy. HelloAI says fluency because the test is use under real conditions, on a busy shift with an output in front of you, and recognizing the terms is only the starting point. Fluency is also specific to a role. What a radiologist needs to know about a detection tool differs from what a procurement lead or a ward nurse needs, so one generic course does not produce it for everyone.
AI governanceExplained 2 ways: The decisions, roles and records a health system uses to control which AI tools come in, how they are checked, who owns them once live, and when they come out. In practice that means an inventory of every tool in use, an intake review covering clinical value, safety, data, legal terms and equity, a named clinical owner for each tool, monitoring after go-live, a route for reporting problems, and a rule for retiring a tool. The period after go-live needs as much structure as the intake review, because that is where vendor updates, drift and quiet workarounds happen. One practical test is whether the process can say which version of each model is running today and who approved that version.
AI harnessExplained 2 ways: The scaffolding wrapped around an AI model that turns it into something able to do real work. A model on its own just predicts the next piece of text. The harness is everything around it: the tools it can call, the loop that lets it take an action and check the result before acting again, the memory and context it works from, and the guardrails and permissions that limit what it can touch. Think of the model as the engine and the harness as the rest of the car. This matters for safety, because the limits on what an agent is allowed to do often live in the harness, not in the model itself. The word is also used in AI evaluation to mean a test harness, the standardized setup for running a model through benchmarks, but in agentic AI it means the orchestration layer described here.
AI penaltyExplained 2 ways: The finding that disagreeing with an AI recommendation, even when you are clinically right, raises your liability in the eyes of a lay jury, while agreeing with it gives you no protection. A one-way ratchet.
Alert fatigueExplained 3 ways: When a tool fires so many low-value alerts that clinicians start tuning them all out, including the important ones. A common reason a technically accurate model fails in practice.
Algorithmic biasExplained 3 ways: Systematic error in an AI tool that falls unevenly across groups of patients, so the tool works less well for some people than its overall numbers suggest. It enters at several points. A group can be thin in the training data. An outcome can be recorded less often or less accurately for some patients, which happens when a test is ordered less often in an underserved population. The label itself can be a proxy, such as past healthcare spending standing in for health need, so the model learns who received care and treats that as who needed it. A single overall accuracy figure hides all of this. Before a tool goes live, ask for performance broken down by the groups your service treats: sex, age, ethnicity, language, and the sites and scanners you run.
Ambient AI scribeExplained 3 ways: AI that listens to a clinical encounter and drafts the documentation automatically, cutting the time a clinician spends writing notes.
AnchoringExplained 2 ways: When an early piece of information, such as an AI's suggested diagnosis, pulls your thinking toward it and makes you less likely to consider alternatives.
Artificial general intelligence (AGI)Explained 2 ways: A system that transfers mastery learned in one domain across all others with no task-specific training, matching or surpassing human ability across virtually all cognitive tasks. The goal large tech companies and countries are racing toward. Whether it has already arrived is actively debated: some classify today's leading models as early or "emerging" AGI, while others hold that true general intelligence is still ahead.
Artificial super intelligenceExplained 2 ways: The most speculative level of AI: capability beyond artificial general intelligence that, by definition, exceeds what humans can imagine. Would emerge if AGI grew past anything humanity could achieve on its own.
Attention mechanismExplained 2 ways: The core component of a transformer that lets the model weigh how much every part of the input matters to every other part. It builds an attention filter from the inputs and uses it to give some words more weight than others when forming a representation.
AUROCExplained 3 ways: Area under the ROC curve, a common measure of how well a model separates two outcomes (for example disease vs no disease). 1.0 is perfect, 0.5 is a coin toss. A high AUROC does not guarantee the model is useful in a real clinic.
Automation biasExplained 3 ways: The tendency to over-trust an automated system and stop questioning it. In clinical AI it means accepting a model's output without applying your own judgment.

B

BERTExplained 2 ways: A well-known encoder-only (autoencoding) transformer, pre-trained with masked language modeling so it reads context bidirectionally. Used for understanding tasks like classification. In healthcare, the BEHRT adaptation was trained on sequential diagnosis codes to predict future diagnoses.
Broad AIExplained 2 ways: A system that could handle any task at a human level, not just one. Also called strong AI or general AI. The counterpart to narrow AI, which covers the task-specific tools running everywhere today. Whether broad AI has arrived yet is debated: some researchers count today's leading models as early examples, others say true general capability is still ahead.

C

CalibrationExplained 3 ways: Whether a model's stated confidence matches reality. A well-calibrated model that says "70% risk" is right about 70% of the time. A model can discriminate well and still be poorly calibrated, so its numbers mislead.
CausalityExplained 3 ways: Whether one thing actually causes another, as opposed to the two simply moving together in the data. Almost every AI model in healthcare today learns correlation: it finds patterns that predict an outcome without knowing why. That is often enough for triage or risk stratification, and it is not enough for a decision about what to do next, because acting on a correlation can change the very thing that made it predictive. The textbook case is the pneumonia model that learned asthma patients had lower mortality, when the real reason was that asthma patients were sent straight to intensive care. Causal questions need a study design that can answer them, not a bigger training set. When a vendor says its tool identifies the drivers of an outcome, ask whether that claim rests on a correlation found in retrospective data or on evidence that changing the input changes the result.
Clinical decision supportExplained 3 ways: Software that gives a clinician patient-specific information at the moment of a decision: a drug interaction alert, a sepsis warning score, a suggested dose, a flag on an image. Hospitals have run it for decades as hand-written rules, and much of the AI in clinical use today is decision support with a learned model inside. The clinician stays the decision-maker, so the design question is whether they can see the basis for a suggestion and judge it in the time they have. In the US the same question shapes regulation. Software that meets four statutory criteria falls outside FDA device rules, and one of the four is that the clinician can independently review the basis for the recommendation. Software that analyzes medical images or physiological signals remains a device, and FDA guidance holds that a tool meant for time-critical decisions does not give the clinician a real chance to review its basis, whatever it displays. When you assess a tool, ask what it shows the clinician besides its answer.
Clinical readinessExplained 3 ways: Whether a tool is ready to be used on patients, which is a different question from whether it performs well on a benchmark. Clinical readiness covers the things that decide whether a model survives contact with a real service: validation on the population it will actually see, calibration at the operating point you will run it at, a named owner for monitoring after go-live, a defined role in the workflow, and a plan for what happens when it is wrong. A model can post a state-of-the-art AUROC and meet none of them. Most published models never get this far, which is most of the reason the distance between a paper and a deployed tool stays so wide. For anyone evaluating a purchase, this list decides more than the headline metric does.
Computer-aided detectionExplained 3 ways: Older imaging software that marks areas of interest for a clinician to confirm. Branded CAD in the 2000s partly to avoid the word AI after earlier hype. Same basic idea as today's detection-assist tools.
Confusion matrixExplained 2 ways: The table that counts a tool's results against the truth: true positives, false positives, false negatives and true negatives. Sensitivity, specificity, positive predictive value and negative predictive value are all ratios taken from these four cells, so the table is where those four metrics come from. It is also the plainest view of a tool, because it shows counts of patients, and counts are what your service has to staff. A tool with excellent specificity can still raise more false alarms than true finds when the condition is rare, and the table makes that visible at a glance. When a vendor quotes a metric, ask for the four counts at the threshold you will run, in a population with your prevalence.
Context windowExplained 2 ways: The amount of text a model can hold in mind at once, counting both what you give it and what it writes back. It is measured in tokens. Once a conversation or document runs past the limit, the model starts losing track of the earliest parts. This explains a common frustration, that a tool seems to forget something said earlier, and it has practical consequences in healthcare: a long record or a lengthy guideline may not fit in one window, so what the model can reason over at any moment is bounded. Larger context windows cost more to run, which is part of why capability and price move together.
Contrastive learningExplained 2 ways: A training method that teaches a model what belongs together by showing it matched and mismatched pairs, then pulling the matches closer together inside the model's representation and pushing the mismatches apart. CLIP, the model behind much of today's image and text AI, was trained this way on images paired with their captions. In healthcare the same trick works on chest X-rays paired with their reports: nobody has to label each finding by hand, because the pairing is the label. This is how most medical vision-language models get built without a hand-labeled dataset, and it is also where their blind spots come from, since a model only learns the distinctions its pairs happened to contain.
Convolutional neural network (CNN)Explained 2 ways: A neural network built for spatial data like images, where each layer looks at a small neighbourhood of the input and global context only builds up in the deeper layers. CNNs are strong on local patterns but weaker at long-range relationships, which is part of why vision transformers are used as an alternative.

D

Data governanceExplained 3 ways: The rules and roles that decide who may use which patient data, for what purpose, with what protections, and who answers for it when something goes wrong. AI raises the stakes in three places. Training a model is a new use of data collected for care, so it has to be checked against the legal basis and approvals the data was collected under. Sharing data with a vendor needs a contract that states whether the vendor may keep it or train other products on it. And outputs such as risk scores written back into the record become patient data themselves. When a vendor asks for a data extract, data governance is the process that asks what for, how much, and for how long.
De-identificationExplained 3 ways: Removing or masking the details that tie data to a person, so it can be used for research, training or sharing at lower risk. In the US, HIPAA allows two routes: removing a defined list of 18 identifiers, or a formal determination by a qualified expert that the risk of re-identification is very small. European law draws a different line. Under GDPR, pseudonymized data, where names are replaced by a code, stays personal data for anyone who holds the key or could reasonably get it. Only data that cannot be linked to a person by any means reasonably likely to be used falls outside the regulation. Medical images need extra care, since identifiers sit in file headers, in text burned into the pixels, and in the anatomy itself: a face can be reconstructed from a head CT or MRI. De-identified data can be re-identified by linking it with other datasets, so treat it as lower risk and keep it under control.
DecoderExplained 2 ways: The part of a transformer that writes the output one token at a time, using the context it was given and the tokens it has already produced. At each step it scores every possible next token, and a separate generation setting picks one: some systems always take the top-scoring token, others sample from the likely options. In an encoder-decoder model the context comes from the encoder; decoder-only models such as GPT work from the prompt alone.
Decoder-only modelExplained 2 ways: A transformer that uses only the decoder, also called an autoregressive model. Built to generate text by predicting the next token in a sequence. The GPT models are the famous example.
Deep learningExplained 3 ways: A kind of machine learning that uses layered neural networks to learn patterns from large amounts of data, such as millions of medical images. It is what made image-based AI like diabetic retinopathy screening work.
DeskillingExplained 3 ways: The gradual loss of a clinician's own skill when a tool does the task for them often enough that the underlying ability fades.
Diffusion modelExplained 2 ways: A generative model that creates images, and increasingly video and 3D volumes, by starting from random noise and removing it step by step until a realistic picture remains. It learns this by being shown real images with noise added and trained to undo it, so what it generates follows the patterns of its training images, and it can sometimes reproduce one of them closely. In medicine it is used to make synthetic training images, to denoise scans and to help reconstruct images from less data. A synthetic image shows plausible anatomy that belongs to no actual patient unless it closely copies a training image, and on a real patient's scan the model can add realistic detail that is not there or smooth away a real finding.
DistillationExplained 2 ways: A technique for training a smaller, cheaper model to copy the behavior of a larger, more capable one, so you get much of the quality at a fraction of the running cost. The large model acts as a teacher and the small model as a student. This is increasingly how the AI tools people actually use are built, because the biggest models are expensive to run at scale. For a healthcare audience the point is that a small fast model in a product may be a distilled version of a frontier model, and it can carry both the strengths and the blind spots of its teacher.
Domain shiftExplained 3 ways: When the data a model meets in deployment differs systematically from the data it was trained on: a different scanner, a different acquisition protocol, a different patient mix, a different hospital. Performance drops, often sharply, and usually with nothing in the output to warn you. Domain shift is the reason external validation exists and the reason a model that works at one site cannot be assumed to work at the next. It is not the same as model drift, which is one site changing over time; domain shift is the gap between two sites or two populations at the same moment. Before buying, ask which scanners, protocols, and populations the model was validated on, and be honest about how far yours sit from them.

E

Electronic health recordExplained 3 ways: The digital record of a patient's care. The push to adopt EHRs created the large digital datasets that machine learning in healthcare runs on.
EncoderExplained 2 ways: The part of a transformer that reads an input sequence and turns each word into a contextual representation, a vector that reflects the rest of the input. It stacks identical layers, six in the original transformer and 12 in BERT-base, each combining multi-head self-attention with a feed-forward network. When one summary of the whole input is needed, for example to classify a report, it is drawn from those outputs, usually from a special classification token or an average of all of them.
Encoder-only modelExplained 2 ways: A transformer that uses only the encoder, also called an autoencoding model. Built for understanding context, such as classification and sentiment analysis. BERT is the classic example, pre-trained with masked language modeling and fine-tuned by adding classifier layers.
Ex vivo: Studied outside a living body, for example on tissue or an organ removed from an animal. A result ex vivo shows a task can be done in that setting. Safety in patients needs trials in patients.
Expert systemExplained 2 ways: An early form of AI built from hand-coded rules, like 1970s systems MYCIN and INTERNIST. Powerful in narrow tests but brittle: it cannot handle anything it was not explicitly programmed for.
ExplainabilityExplained 3 ways: How far a person can understand why an AI tool produced a particular output. There are two routes to it. An interpretable model is simple enough to read directly, like a points-based risk score where you can see what each factor added. An explanation method sits on top of a complex model and approximates its reasoning after the fact, and the heatmap over a chest X-ray is the familiar example. The two give different things. A heatmap shows which pixels most influenced the output, and it says nothing about whether the finding there is real or why the model weighed it. Explanations can also make a wrong output look more convincing. For deciding whether to trust a tool, validation on your own patients tells you more than an explanation of a single case.
External validationExplained 3 ways: Testing a model on data from a different hospital, region, or population than it was trained on. It is the real test of whether a model generalizes; performance almost always drops compared to internal testing.

F

False discovery rateExplained 2 ways: Of everyone an AI flags as high risk, the fraction who turn out not to have the condition. It tells you how much weight a flag can actually bear.
False negativeExplained 3 ways: When an AI tool misses something that is actually a problem. False negatives are the dangerous ones in healthcare, because the thing the model failed to flag still needs attention.
False positiveExplained 3 ways: When an AI tool flags something as a problem that turns out not to be one. A high false positive rate means the tool cries wolf often, which can wear down the clinicians acting on its alerts.
Federated learningExplained 3 ways: A way to train one AI model across several hospitals or sites without moving any patient data out of where it lives. Instead of pooling everyone's records in one place, the model is sent to each site, learns locally, and only the model updates are shared back and combined. It lets AI learn from large, diverse data while the records stay put, which is why it is attractive for privacy-sensitive healthcare data.
Feed-forward networkExplained 2 ways: A fully connected network inside each transformer layer that applies further transformations to the attention output, letting the model capture more complex relationships between tokens. It sits alongside the attention sublayer in both encoder and decoder.
Fine-tuningExplained 2 ways: The second training stage, where a pre-trained model is further trained on a smaller, task-specific dataset. Because the model already carries general knowledge, fine-tuning converges faster and needs far less labelled data to reach strong performance.
Foundation modelExplained 3 ways: A large model trained broadly on huge datasets that can then be adapted to many tasks, rather than built for one. In imaging, a foundation model can handle several modalities instead of one narrow job.

G

General-purpose AIExplained 3 ways: An AI model trained on broad data that can handle many different tasks, like ChatGPT or other large language models. Regulators including the EU AI Act use this term (GPAI) for foundation models. General-purpose AI is broad in capability, but it is still not broad AI or AGI: it does not truly understand or transfer skill across any domain the way a person does. It sits on the narrow side of the narrow-vs-broad split.
Generative AIExplained 3 ways: AI that produces new content, text, images, audio, rather than just classifying or scoring existing data. Large language models are the most familiar example.
GPTExplained 2 ways: A well-known family of decoder-only (autoregressive) transformers, pre-trained by predicting the next word in a sequence and used to generate text. Fine-tuned to a domain through transfer learning.
Ground truthExplained 2 ways: The answer a model's output is scored against during development and testing. In medicine it is usually a human judgment of some kind: one radiologist's read, a consensus panel, a pathology result, or what happened to the patient at follow-up. Each is imperfect in its own way, and a model cannot be shown to be more accurate than the truth it was measured against. If experts disagree on a third of cases, a claim of matching expert performance means something different from the same claim on a task where they agree. When a tool is said to match or beat clinicians, the first question is what the truth was and who decided it.
GroundingExplained 3 ways: Tying a model's answer to a trusted source, such as a specific document, record, or database, so it is based on real information rather than generated from memory. Grounding is the main defense against hallucination: instead of letting the model produce a plausible-sounding answer on its own, you make it draw from and point to source material, often using retrieval-augmented generation. For clinicians and the people who buy these tools, a grounded answer that cites where it came from can be checked, which is what makes it safer to act on than an ungrounded one.
GuardrailsExplained 3 ways: The concrete limits built around an AI system to stop it doing things it should not. Guardrails can block certain outputs, refuse certain requests, keep the system from acting outside an approved scope, or force a human check before anything consequential happens. When someone asks "does it have guardrails?" they are really asking what the system is technically prevented from doing, not whether it has been told to behave. For clinicians and the leaders accountable for a tool, this is the difference between a promise and an actual constraint: ask where the guardrail lives and what it blocks, because a guardrail that is only a line in a policy is not a guardrail.

H

HallucinationExplained 3 ways: When an AI generates content that is fluent and plausible but wrong or never happened, such as a measurement that was never taken or a finding never discussed. In clinical documents this is a patient-safety issue, not a quirk.
Human in the loopExplained 3 ways: A setup where a person reviews or approves an AI output before it acts. It only adds safety when the human is genuinely engaged with the decision, not just clicking approve.
Human on the loopExplained 2 ways: An oversight model where a clinician reviews AI work at defined checkpoints rather than approving every single step. The shift agentic systems make from human-in-the-loop, without removing the human.

I

InferenceExplained 2 ways: What happens when a trained model is actually used, taking an input and producing an output. It is the counterpart to training: training is how the model is built, inference is every time it is run afterward. The word comes up around cost and speed, because each use consumes computing power, and a tool that is cheap to use once can be expensive across a whole health system. For clinical use the practical point is that the model is not learning from your case during inference unless it has been specifically set up to, so it is applying what it already learned, not adapting to you in the moment.
Informed consentExplained 3 ways: A patient's right to understand and agree to their care, including how a recommendation was reached. With AI involved, this is starting to include knowing that AI played a part and what its role was.

J

J-curveExplained 2 ways: The shape of technology adoption where costs come first and returns arrive later. Organizations that give up during the early dip pay for the learning without ever collecting the reward.

K

Knowledge graphExplained 3 ways: A structured map of things and the relationships between them: this drug treats this condition, this condition presents with this finding, this finding is seen on this modality. Instead of holding facts as free text, a knowledge graph holds them as entities and links, which a machine can follow step by step and a person can inspect. SNOMED CT, ICD, and the medical ontologies underneath them are close relatives, so most hospitals already run on one without calling it that. The current interest is in pairing a graph with a language model, so the model's fluency gets checked against a source that states what is actually true. When a vendor says its tool reasons over clinical knowledge, the useful question is which graph or ontology it uses and who keeps that source current.

L

Large language modelExplained 3 ways: An AI trained on huge amounts of text that can read and write in plain language. ChatGPT is one. In healthcare they draft notes, summarize, and answer questions, but can be confidently wrong on facts.
Levels of autonomy: A scale that grades how much a robot decides and does on its own, from no autonomy to full autonomy. Different papers use different scales, so check which one a claim uses.

M

Machine learningExplained 3 ways: A way of building software where, instead of writing rules by hand, you show the system many examples and let it find the patterns itself. Most modern clinical AI is machine learning.
Masked attentionExplained 2 ways: Attention in the decoder where each word may only score against the words before it; later words get a score that becomes zero after softmax. This forces the model to predict each token using only earlier ones, never future ones, which is what makes left-to-right text generation work.
Masked language modelingExplained 2 ways: A self-supervised pre-training task where part of the input tokens are hidden and the model is trained to predict them from the surrounding unmasked tokens. Because it can use context on both sides, it produces bidirectional encoders. It is the task BERT is pre-trained with.
Mixture of expertsExplained 2 ways: An architecture that splits a large model into many smaller sub-networks and sends each input to only a few of them, so the model can hold far more parameters without costing proportionally more to run. The name misleads people. The experts are partitions the model worked out during training, not specialists in any recognizable sense, and none of them corresponds to a clinical specialty. It matters mainly as the explanation for why some very large models are cheap enough to serve at scale. If a product is marketed as having a radiology expert or a cardiology expert inside it because it uses a mixture of experts, that is a sales line, not an architecture.
Model cardExplained 2 ways: A short, standard document that travels with a model and says what it is for, who it was built for, what data it was trained and tested on, how it performed, including for subgroups, and what its known limits are. The format was published by a Google research team in 2019 and has since been adapted for healthcare, sometimes under the name model facts label or AI nutrition label. A model card is only as useful as its contents, so check that it covers what a buyer needs most: the population it was validated on, performance at the operating threshold, and the situations where it should not be used. Ask for one before a demo, and read the limitations section first.
Model Context Protocol (MCP)Explained 2 ways: A shared standard for connecting AI models to the tools, data, and systems they need to do work, so each connection does not have to be built from scratch. You can think of it as a common plug that lets a model and an outside system talk to each other. The term is appearing more in agentic AI conversations because it is becoming a common way to wire models into real software. For decision-makers the relevant point is governance: a standard connection still needs the same scrutiny over what data flows through it and what the model is permitted to do once connected.
Model driftExplained 3 ways: When an AI model's real-world performance degrades because the patients it now sees no longer match the data it was trained on. It is why models need monitoring and revalidation, not just a one-time launch.
Multi-head attentionExplained 2 ways: Running several attention layers in parallel so each head can form a different filter and focus on a different part of the input. The heads' outputs are concatenated and passed through a linear layer to combine them.
Multimodal AIExplained 3 ways: AI that takes in more than one type of data at once, for example imaging plus lab values plus clinical notes, and integrates them into a single prediction, closer to how a clinician actually reasons.

N

Narrow AIExplained 2 ways: Every AI system built so far, past and present, from single-task tools to general-purpose models like ChatGPT. Even models that handle many tasks are still considered narrow by many researchers, because they are built for specific tasks rather than matching human general intelligence. Also called weak AI. The counterpart to broad AI.
Natural language processingExplained 3 ways: The branch of AI that works with human language: reading clinical notes, transcribing speech, pulling structured meaning out of free text.
Negative predictive valueExplained 3 ways: Of the patients a tool calls negative, the share who really do not have the condition. It is the number that matters for a rule-out tool, one whose job is to tell you who can safely skip the next step. Negative predictive value depends heavily on how common the condition is. When the condition is rare, it is high almost automatically, because nearly everyone is negative to begin with, so a high figure in a screening population says little about the tool. The useful questions are how many true cases it misses in absolute numbers and what happens to those patients. It is the counterpart of positive predictive value, and the two move in opposite directions as prevalence changes.
Neural networkExplained 3 ways: The layered mathematical structure deep learning uses to find patterns. Loosely inspired by how neurons connect, though it does not actually work like a brain.
Neuro-symbolic AIExplained 3 ways: A hybrid approach that puts the pattern recognition of neural networks together with the explicit rules and logic of symbolic AI. The neural side is good at perception and bad at guarantees. The symbolic side can enforce a constraint and show its working but cannot learn from raw images. Combining them is one of the more plausible routes to a system that is both accurate and checkable, which is what a high-stakes medical decision needs. Most of this is still research, though the pattern already appears in production wherever a model's output passes through a rule engine before anyone sees it. Worth knowing because it is where many people point when asked how AI becomes auditable enough for regulated clinical use.

O

Open sourceExplained 2 ways: Software is open source when anyone can inspect, modify and redistribute the code. The Open Source Initiative extended that to AI in October 2024 with the Open Source AI Definition, which asks for three things under terms that allow anyone to use, study, modify and share the system for any purpose: the parameters, the training and inference code, and detailed information about the training data. Note the third one. The definition asks for a full description of the data and its provenance, not the dataset itself, and plenty of people think that bar is set too low. The point of open source is the process, not just the artifact. When the build is visible, an independent group can test a developer's claims about bias, failure modes and the populations a model was never meant for, instead of taking them on trust. Few released models clear the bar. OLMo from the Allen Institute for AI does. Llama, Gemma and Mistral's open releases are open weights under vendor licences, and calling those open source sets the wrong expectation about what can be verified.
Open weightsExplained 2 ways: A model released so that anyone can download its trained parameters, the numbers that come out of training, and run it themselves. In practice that means you can host it on your own infrastructure, keep patient data inside your own network, fine tune it on local data, and test it against your own population without sending anything to a vendor API. That control is usually the real reason a health system wants open weights. What you do not get is the recipe: the training data, the training code and the decisions that shaped the model stay with the developer, so you can measure how the model behaves but not inspect why it behaves that way. Open weights also says nothing about the licence. Most releases, Llama and Gemma included, come with vendor terms and use restrictions, and some are research use only, which is a procurement question before it is a technical one. Open weights is not open source.
Operating thresholdExplained 3 ways: The cut-off that decides how readily a predictive tool flags a case. Lower it and you catch more, at the cost of more false alarms; raise it and you cut false alarms, at the cost of missed cases. Where it sits is a clinical decision, not a fixed property of the tool.
OverfittingExplained 2 ways: When a model learns the particulars of its training data, noise included, so closely that it performs well on those patients and worse on new ones. Small datasets and many input variables make it more likely, and many medical datasets have both. The visible sign is a gap between strong internal results and weaker external ones. A subtler form happens when a team tunes a model again and again against the same test set until the test set itself has been learned, which inflates the published figure without anyone intending it. When you read a performance claim, look for the number from data the developers never touched, and expect it to be lower.

P

PACSExplained 2 ways: Picture archiving and communication system: the digital storage hospitals use for medical images. Because radiology went digital early through PACS, it had the labeled image archives that made imaging AI possible first.
Physical AI: An informal name for AI-enabled systems that sense and act in the physical world, such as robots that move an ultrasound probe or a surgical instrument. It isn't a clinical or regulatory category, and systems described with it can do very different things.
Population stability indexExplained 2 ways: A number that measures how much the population a model sees has shifted away from the population it was built on. A higher value means more drift. It is one of the concrete triggers proposed for when a deployed AI tool needs re-checking: a population stability index above 0.2 is a common threshold for "investigate this". A practical early-warning signal that the data environment around a model has changed.
Position encodingExplained 2 ways: Extra information that tells a transformer the order of the words it reads. A transformer takes in all the words at once, which is fast but loses their order, and order changes meaning: "no fever, cough" says something different from "fever, no cough". The original transformer added to each word's embedding a wave-based signature unique to that word's position. Many current language models use a rotary scheme instead, which builds position into the way attention compares words.
Positive predictive valueExplained 3 ways: Of the patients a model flags as positive, the share who actually have the condition. It depends on how common the condition is, so the same model has different predictive value in different populations.
Post-market surveillanceExplained 2 ways: Watching a tool after it is approved and deployed, to catch problems that only show up in real-world use. For clinical AI it means tracking whether a live tool is still accurate, still processing on time, and still being used as intended, because regulatory approval at launch does not guarantee ongoing performance. Increasingly treated as an obligation rather than best-effort. Closely related to post-deployment monitoring.
Pre-trainingExplained 2 ways: The first training stage, where a model learns general representations from a large dataset (often unlabelled, via self-supervision) before being adapted to a specific task. Pre-training ideally uses data from the same domain as the eventual task.
Predictive AIExplained 3 ways: AI that scores, classifies, or flags from data it was trained on: a CAD tool, a risk score, a triage algorithm. You direct it by selection and configuration, picking the right tool and setting it correctly, rather than by typing an instruction. Distinct from generative AI, which produces new content from a prompt.
PromptExplained 3 ways: The instruction or question you give an AI model to get a response. A prompt can be a short request or several paragraphs of context. The model has no goal of its own, it reacts to what it is given, so the wording, the detail, and the framing of the prompt shape the answer you get back. In practice this means two clinicians asking the same model the same clinical question in different words can get meaningfully different answers, which is why how a prompt is written is part of how safely a tool is used.
Prompt engineeringExplained 2 ways: The practice of writing and refining prompts so a model gives more reliable, useful answers. It covers giving clear instructions, supplying the right context, showing examples, and asking for a specific format. For a clinical or decision-making audience the point is not that this is a special skill, but that the quality of an AI tool in real use depends partly on how the prompts behind it are built, and that a vendor demo with carefully engineered prompts may not reflect what a busy user types at 2am.
Prompt injectionExplained 3 ways: An attack in which text the AI reads as content, such as a sentence in a document, a web page, an email or an image, is treated by the model as an instruction. Models take in instructions and content through the same channel, so they can't reliably tell them apart. Security agencies say there is no fully reliable technical fix yet, so the main defense is limiting what the AI can read and what it's allowed to do.

Q

Query, key, valueExplained 2 ways: The three vectors an attention layer computes for every word, each from its own learned linear layer. Each word's query is compared with every word's key to produce a matrix of scores for how strongly the words relate, and those scores, turned into weights, decide how much of each word's value flows into the result. The recipe is scaled dot-product attention. Multiply queries by keys, divide by the square root of the key dimension (the length of each key vector in an attention head), apply softmax, and use the resulting weights to combine the values.
Quintuple AimExplained 2 ways: A checklist of what "value" means in healthcare: better patient outcomes, improved population health, lower cost, better clinician well-being, and greater health equity. It extends the older Triple and Quadruple Aims by adding equity. For AI, it is the test of whether a tool actually helped: if a deployed tool is not moving at least one of the five, the value case is just words.

R

Reasoning modelExplained 3 ways: A type of AI model that works through a problem in steps before giving a final answer, rather than responding in one shot. This step-by-step working is often called chain-of-thought. Reasoning models tend to do better on harder problems that need several linked judgments, which is why they have become a distinct category. The trade-offs matter for healthcare use: they are usually slower and cost more to run, and producing visible reasoning is not the same as being correct, so the steps should be treated as a model showing its work, not as a guarantee.
Recursive self-improvement (RSI)Explained 2 ways: The idea that an AI system can take over part of the work of building the next, better AI, instead of only helping a human researcher. The research loop is propose an idea, build it, run the experiment, check the result, learn from it, pick what to try next. RSI is when machines start running stages of that loop and the output of one generation helps create a stronger next one, with less human involvement each time. It is worth being precise, because the phrase "AI that builds AI" oversells where things actually are. Today RSI mostly means automated engineering and automated research on things that are easy to measure, like making model training faster or cheaper, not AI inventing a brand new model on its own. It is also different from a self-improving agent, which improves its own prompts, tools, and memory but does not rebuild the model. For people accountable for AI in healthcare the relevant shift is about oversight: as more of the loop moves from human hands to machine hands, the human job moves toward setting goals, validating results, and governing the process, which is exactly where the hard questions sit.
Red teamingExplained 2 ways: Deliberately attacking or stress-testing an AI system to find where it fails before it is deployed, by trying to make it produce harmful, wrong, or out-of-bounds output. The name comes from security, where a "red team" plays the adversary. It is becoming a standard expectation in AI safety and is now appearing in procurement and governance conversations. For decision-makers it is a fair question to ask of any healthcare AI vendor: was this system red-teamed, by whom, and what did they find, because a tool that has never been tested against misuse has unknown failure modes.
Reinforcement learningExplained 2 ways: Machine learning that learns the best actions by trial and error in its environment, guided by rewards. A robot vacuum learns the layout of a home this way, memorising every misstep to improve the next run.
Retrieval-augmented generationExplained 3 ways: A method that grounds an AI's answers in specific, vetted documents instead of its general memory, which reduces hallucination. Often shortened to RAG.
ROIExplained 2 ways: Return on investment: what you get back relative to what you spent. For healthcare AI this includes clinical, workforce, and financial returns, not just cost savings.

S

Search and planning algorithms: A kind of symbolic AI that follows fixed logic to explore possible paths toward a goal, without learning from past attempts. Amazon's warehouse robots use search algorithms to find the most efficient route between items.
Self-attentionExplained 2 ways: Attention where the query, key, and value all come from the same input, so a sequence attends to itself. It is how a transformer works out the contextual relationships among the words within one sequence.
Self-supervised learningExplained 2 ways: Training a model on large amounts of unlabelled data by having it predict parts of its own input, so no expensive expert annotations are needed. It is the usual first stage for transformers, producing general representations that a later fine-tuning step adapts to a specific task.
SensitivityExplained 3 ways: Of the patients who truly have a condition, the proportion that the model correctly flags. High sensitivity means it misses few real cases.
Sequence-to-sequence modelExplained 2 ways: A transformer that uses both encoder and decoder, also called an encoder-decoder model. Built to turn one sequence into another, for tasks like translation, summarisation, and speech recognition. T5 and BART are the well-known examples.
Shadow AIExplained 3 ways: Using a general-purpose AI tool for clinical or administrative work when your institution never approved or vetted it, and does not cover it, such as pasting patient details into a consumer chatbot to draft a letter. It is the healthcare version of shadow IT, and it usually starts for a good reason, saving time on a real task. The risk sits behind the tool: anything you enter falls outside HIPAA protection, the vendor has typically disclaimed responsibility and capped its liability, and there is no institutional backstop if something goes wrong, so the exposure lands on the clinician. Shadow AI is not automatically off limits, but it should be used with eyes open: no identifiable patient data in a tool nobody sanctioned, and no assumption that someone else is carrying the risk. A spread of shadow AI is also a signal to the people accountable for a service that the sanctioned tools are either missing or too clunky to use.
Simple rule-based systems: The most basic kind of symbolic AI: a system that automates a single decision using basic if-then logic. A light-sensing plug that switches on outdoor lights at sunset, or a basic programmable thermostat. Very limited, very good at the one thing it does.
SoftmaxExplained 2 ways: A function that turns a set of raw scores into a probability distribution that sums to one. Transformers use it inside attention to turn comparison scores into weights, and at the final layer to turn vocabulary scores into the probability of each possible next word.
Software as a medical deviceExplained 3 ways: Software that is regulated as a medical device because its intended use is to diagnose, treat, prevent, or inform clinical decisions. What the software claims to do often matters more for regulation than how it is built.
SpecificityExplained 3 ways: Of the patients who do not have a condition, the proportion that the model correctly clears. High specificity means few false alarms.
Standard of careExplained 3 ways: The level of care a reasonably competent clinician would provide in a given situation. It is the legal benchmark used to judge negligence, and it shifts over time as technology and practice change.
Structured dataExplained 2 ways: Information that fits neatly into rows and columns: lab values, vital signs, diagnosis codes, dates. It is the easiest kind for traditional machine learning to use, and it is what most risk scores and dashboards are built on. The counterpart to unstructured data.
Supervised learningExplained 3 ways: Machine learning that learns patterns from data humans have already labelled, then makes decisions on new data. The phone face-recognition setup is a clean example: tilting your face labels each frame, and the phone uses that labelled data to recognise you later.
Swin transformerExplained 2 ways: A hierarchical vision transformer that processes images in shifted local windows, making it efficient and well suited to dense tasks like medical image segmentation. Swin UNETR pairs a Swin encoder with a UNet-style decoder for 3D segmentation.
Symbolic AIExplained 2 ways: AI that follows explicit rules coded in advance for every anticipated outcome, without learning. Also called rule-based AI. Splits into simple rule-based systems, expert systems, and search and planning algorithms. The counterpart to machine learning inside narrow AI.
Synthetic dataExplained 2 ways: Artificial patient records or images, usually produced by a generative model trained on real data, or by simulation from published statistics. The appeal is that synthetic data can be shared, used to test software, or used to add examples of rare conditions without exposing real patients. It has two limits that matter in practice. It inherits whatever the real data got wrong, including gaps in who was represented. And a generator trained on real patients can reproduce some of them closely enough to leak them, because generative models can memorize parts of their training set. A model trained partly on synthetic data still has to be validated on real patients from the population it will serve. When a vendor says its product was trained on synthetic data, ask what the generator learned from and how privacy was tested.
System promptExplained 2 ways: A fixed set of instructions that shapes how an AI model behaves before the user types anything. It sets the role, the rules, the tone, and often the safety limits the model should follow in every conversation. Two organizations using the same underlying model can get very different behavior because their system prompts differ. For people accountable for a deployed tool this matters: a lot of what makes a healthcare AI product safe or unsafe is decided in the system prompt and the guardrails, not in the model itself, so it is fair to ask to see it.

T

Target leakageExplained 2 ways: When a model is trained on information it would not have at the moment of prediction. The model then looks better than it is, sometimes far better. A common clinical form is a feature that records what clinicians already suspected: a sepsis model that uses whether blood cultures were ordered has partly learned the doctor's judgment. A related problem, train-test contamination, puts test patients into training, for example when images are split by scan and a patient has several, so the model is tested on patients it has already seen. Security teams use "data leakage" for something else, the escape of data from an organization, which is why this entry uses the modeling term. Ask how the data was split, and which inputs exist at the exact time the tool will be used.
Teleoperation: Controlling a robot from a distance. The robot moves, and the person directs every movement, whether they sit in the next room or in another city.
TokenExplained 2 ways: A small chunk of an input sequence, often a word or part of a word, that a transformer processes as one unit. The input is broken into tokens, each is turned into numbers, and attention then works out how the tokens relate.
Tool useExplained 2 ways: When an AI model does more than produce text by calling on outside tools, such as searching a database, looking up a record, running a calculation, or sending a request to another system. Tool use is what lets a model act in the real world rather than just describe it, and it is the core of how agents get things done. For a healthcare audience this is where capability and risk both rise sharply: the moment a model can reach into a real system, what it is allowed to touch, and what it is blocked from touching, becomes a safety question, not just a feature.
Traditional machine learningExplained 2 ways: Machine learning that uses classical statistical methods and human-selected features on structured data (tables, spreadsheets, databases) to find patterns, classify, and predict. Techniques include clustering and k-nearest-neighbour. Distinct from deep learning, which handles unstructured data.
Training dataExplained 3 ways: The examples an AI model learns from. A model can only recognize patterns that were present in its training data, which is why gaps and bias in that data carry straight into the model.
Transfer learningExplained 2 ways: Reusing what a model learned on one task as the starting point for another, by loading the pre-trained weights into a downstream model. It is how the knowledge from a large pre-training run is carried into a specific clinical task with limited labelled data.
Transformer modelExplained 2 ways: A neural network architecture that reads a whole sequence at once instead of one step at a time. It uses attention to weigh how much every part of the input matters to every other part, which lets it train in parallel and handle long-range context that older recurrent networks lost track of. Transformers are the architecture behind modern large language models and most current medical-imaging foundation models.

U

Uncertainty quantificationExplained 3 ways: Methods that make a model report how sure it is in a form you can act on, including the option of saying it does not know. Most deployed models return a number whether or not the case in front of them resembles anything in their training data, and that number carries no warning when the case is unfamiliar. Uncertainty quantification supplies the warning: routing low-confidence cases to a human, abstaining instead of guessing, or widening the range around a prediction. It overlaps with calibration and is broader than it, because calibration asks whether stated confidence is honest while uncertainty quantification is about producing and using that signal at all. For a clinical tool the practical test is whether it ever declines to answer, and what happens to the case when it does.
Unstructured dataExplained 2 ways: Information that does not fit neatly into rows and columns: free-text notes, images, waveforms, audio. Roughly 80% of healthcare data is unstructured, and it is both the richest and the hardest for AI to use.
Unsupervised learningExplained 2 ways: Machine learning that finds patterns in data with no human labels, forming its own groupings. Used in bank fraud detection and the Netflix recommendation system, which sorts viewers from what they watch, finish, and skip.

V

Validated populationExplained 3 ways: The patient group a tool was actually tested on. Performance inside that group is evidenced; performance outside it is unknown, not assumed. Checking whether the patient in front of you fits the validated population is part of using any AI tool well.
Vision transformerExplained 2 ways: A transformer applied to images rather than text, used as an alternative to convolutional networks (CNNs). It splits an image into patches and uses attention to relate them, borrowing the transformer advantages of scaling to large data and training end to end without hand-designed layers. Vision transformers are increasingly common in radiology and pathology AI.
Vision-language modelExplained 3 ways: A model trained on images and text together, so it can describe an image, answer a question about it, or point to the part of it a sentence refers to. This is the class of model behind generating a report from a chest X-ray, or answering a question about a scan in words rather than returning a label. Training usually pairs images with the reports already written about them, which is why radiology moved first: the paired data was sitting there. The known weakness is that fluency and grounding come apart, so the model can produce a well-written report describing a finding nobody can locate in the image. Any clinical use needs a way to check the words against the pixels.

W

Word embeddingExplained 2 ways: A fixed-length vector of trainable numbers that represents a word so a computer can work with it. Related words end up with similar values in some dimensions: apple sits near orange, both share something with juice, while airplane sits far from all of them.
World modelExplained 2 ways: An AI system that builds an internal model of how a situation works and how it changes over time, so it can predict what happens next and plan actions, rather than only matching patterns in text or images. The idea is to give a system something closer to a working understanding of an environment, the way a person carries a rough mental map of a room. World models are central to robotics, simulation, and physical AI, where a system has to anticipate consequences before it acts. For a healthcare audience this is still mostly a frontier and research topic rather than everyday clinical software, but it is worth recognizing, because it is the direction behind a lot of the robotics and embodied-AI work that is starting to reach into care settings.

Z

Zero-shotExplained 2 ways: When a model handles a task it was never specifically trained for, using only its general training. A zero-shot imaging model can flag a finding without task-specific examples for that exact use case.

Need more than a definition?

Ask the HelloAI team your question about AI in healthcare. We answer selected questions in our LinkedIn series and list every answer on the Ask page.

Ask a question →

Want the full picture?

Every one of these terms shows up in HelloAI courses: bite-sized modules, real clinical scenarios, verifiable certificates. An account takes one minute, no password needed.

Glossary

Foundations & model types · 32

Clinical use & safety · 21

Metrics & evaluation · 22

Governance & economics · 11

Training & methods · 14

Data & infrastructure · 7

Transformer internals · 18

Other terms · 10

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

Z