A flexible and universal guide to designing human-centred AI products.

Data deliberately crafted to trick or confuse an AI. For example, slightly altered images that fool an image classifier into making the wrong call.

A hypothetical machine that can think and learn across any subject, just like a human. It does not exist yet.

Removing or masking personal details from data so individuals cannot be identified. A basic data protection practice in HCAI systems.

A scale describing how much the AI does on its own versus how much a human stays in control. The appropriate point depends on the stakes and reversibility of each decision.

When an AI produces unfair or skewed outputs because the data it learned from over-represented or under-represented certain groups of people or situations.

An action taken by an AI that has a meaningful, often hard to reverse impact. For example, deleting files, sending an email or making a breaking code change. These should always require confirmation first.

A type of AI task where the system assigns something to a category. For example, deciding whether an email is spam or not spam.

A label that groups AI outputs by how certain the system is. For example: high, medium, low or unknown. Each band should look and behave differently.

Making the AI's level of certainty visible to the user so that a guess looks and feels different from a near-certain answer.

An always-available mechanism embedded in the product that lets users tell the AI it was wrong, rather than a one-off survey surfaced occasionally.

A specific point in an AI workflow where the human needs to be able to act. For example, to approve, override, adjust or undo what the AI is doing or about to do.

Telling a user what would have needed to be different for the AI to reach a different outcome. For example, if your income had been £5k higher, you would have been approved.

Which people, places and situations are actually included in the data an AI learned from. Gaps in representation lead to gaps in performance for those groups.

Showing a user why the AI reached a conclusion, not just what it concluded. Tracing this reasoning path is especially important in high-stakes or regulated contexts.

An interface designed for users who are domain experts, giving them tools to label, correct or annotate AI outputs directly rather than simply accept or reject them.

A clear, low-effort way for a user to tell the system it was wrong. For example, a thumbs down or an incorrect button.

A backup mechanism that activates when an AI system goes wrong, to prevent serious harm from occurring or escalating.

A specific, predictable way an AI system can go wrong. For example, hallucinating, misclassifying or silently failing on an input it was not trained for.

Presenting an AI prediction as an exact figure when it is really an estimate. Showing a house price as £312,450 implies a level of certainty that does not exist. A range is more honest.

A type of AI task where the system creates new content from scratch. For example, text, images, code or audio based on a prompt or instruction.

A technical rule or constraint that stops an AI from producing harmful, inaccurate or out of scope outputs.

When an AI generates information that sounds plausible and confident but is factually wrong, with no indication that it is uncertain.

An approach to designing AI products and systems that keeps human needs, safety and wellbeing at the centre of every decision.

An annotated workflow diagram showing every point where a human needs to be involved in an AI process, what they need to know and what controls they need in order to act.

A structured document identifying who is affected by an AI system, how they could be harmed and what the design response to each failure mode should be.

A clear, easy to use control that lets a person immediately deactivate an AI feature or system. It should be designed as a normal product feature rather than an emergency measure.

A type of AI trained on vast amounts of text that can generate, summarise and respond to written language. ChatGPT and Claude are examples.

Recording what an AI system does, including inputs, decisions and outputs, so you can understand, audit and debug its behaviour over time.

The understanding a person has of how a system works. If this does not match how the AI actually behaves, people will misuse it, lose trust or make harmful decisions based on it.

When an AI puts something in the wrong category. For example, incorrectly flagging a legitimate email as spam.

A plain language document that explains what an AI model does, what it was trained on, where it performs well and where it does not. It should be accessible to everyone in an organisation, not just technical teams.

 The practice of an AI product flagging when a user has asked it to do something outside its training or reliable capability, rather than attempting an answer and failing silently.

A question or piece of data the AI was not trained on and cannot handle reliably. The system should flag this clearly rather than producing a confident but unreliable output.

A visible, low-friction control that lets the human reject or change an AI decision at the moment it is made, without having to dig through menus or settings.

The tendency of recommender systems to surface popular items over personally relevant ones, because popular content has more data behind it.

Revealing features and complexity gradually as a user needs them, giving people time to build an accurate mental model before encountering more advanced behaviour.

A design approach where the AI earns greater autonomy over time as the user builds an accurate and appropriate level of trust, rather than operating at full autonomy from day one.

A technique where an AI looks up relevant information from a trusted source before generating a response, reducing the likelihood of hallucination.

An AI that suggests content, products or options based on a person's past behaviour or preferences. Netflix recommendations and Spotify playlists are common examples.

A type of AI task where the system predicts a numerical value. For example, the likely price of a house or tomorrow's temperature.

An action taken by an AI that can be undone. HCAI products should make reversal easy and obvious.

A type of AI task where the system works out a series of steps to achieve a goal in the right order. Commonly used in agents and automated workflows.

A deliberate escape hatch in a recommender system that lets a user bypass the algorithm and discover something unexpected. Often surfaced as a Surprise me button.

When an AI goes wrong but gives no indication that anything has gone wrong, producing a confident looking output that is actually incorrect. This is the highest priority failure type to design against.

Showing the user where the AI's information came from, with a link or reference, so they can trace and verify the reasoning.

The process of making sure everyone with decision-making power over an AI product genuinely understands its risks and is committed to the agreed mitigations before build begins.

An AI response in a defined, predictable format such as a table or JSON that can be automatically checked for errors rather than read freeform.

A published document explaining what an AI system does, what it was trained on, its known limitations and who is responsible for it. It should be updated whenever the system changes significantly.

The information an AI learned from. It determines what the system knows, what it gets right and what biases or blind spots it may carry.

The period of time after an AI takes an action during which the user can still reverse it. It should be as generous as technically possible and made clearly visible to the user.

Letting users choose how much of a task they hand over to the AI. More human control for high-stakes decisions, less friction for routine ones. The level should be adjustable by the user.

Keeping a history of previous versions of a document, file or output so that AI changes can be reviewed and reversed without permanently overwriting earlier work.

 When the AI visibly responds to a user's correction, showing that the feedback was taken on board. This encourages people to keep providing feedback and improves the system over time.