Why it matters
It reveals that model perception can fail in ways that differ sharply from human perception.
Connected concepts
humAIne field guide · 2026
AI, without
the fog.
A working glossary for the language shaping technology and business. Plain definitions, practical context and the connections between them.
Explore the glossary Living reference · No jargon required
Concept compass01 / orientation
Human
judgement
× AI
judgement
× AI
Data
Models
Systems
Consequences
Find
Search terms, abbreviations and ideas in one place.
Understand
See a plain definition and why the concept matters.
Connect
Follow related concepts to build a working mental model.
90 terms9 lensesPlain EnglishCross-linked
The field, indexed
Find the term.
See the system.
90 terms
A
Why it matters
Memory supports continuity and personalisation, while creating privacy, relevance and data-retention questions.
ALSO: Memory
Connected concepts
Why it matters
The term describes a system property rather than a single product: autonomy can range from tightly bounded to broad.
Connected concepts
Why it matters
Agents move AI beyond answering questions into completing parts of workflows.
Connected concepts
Why it matters
Bias can produce unfair or inaccurate outcomes, especially when affected groups are underrepresented in design and testing.
ALSO: Algorithmic bias
Connected concepts
Why it matters
Governance connects abstract principles to ownership, approval, monitoring and consequences.
Connected concepts
Why it matters
Capability in the workforce determines whether access to AI becomes leverage or merely more software spend.
Connected concepts
Why it matters
Observability makes failures diagnosable and supports quality, cost, safety and drift monitoring.
ALSO: Observability
Connected concepts
Why it matters
Moving from isolated pilots to durable capability requires clear ownership and reusable foundations.
Connected concepts
Why it matters
Model usage fees are only one input; adoption, process redesign and measurable outcomes usually decide the return.
ALSO: AI ROI · Return on investment
Connected concepts
Why it matters
Risk depends on both model capability and context; the same model can be low-risk in one use and unacceptable in another.
Connected concepts
Why it matters
A good use case names the problem, owner, data, risk and success measure—not just the technology.
ALSO: Use case
Connected concepts
Why it matters
The strongest gains often come from redesigning the operating model, not inserting a chatbot into the old one.
Connected concepts
Why it matters
An algorithm is a procedure; a trained model is the learned result of applying algorithms to data.
Connected concepts
Why it matters
Powerful capability is only useful when behaviour remains steerable, safe and appropriate to its context.
Connected concepts
Why it matters
AI is the umbrella term; machine learning, generative AI and agents are approaches within it, not synonyms for the whole field.
ALSO: AI
Connected concepts
Why it matters
Attention helps models connect information across a prompt, though it is not the same thing as human attention or awareness.
Connected concepts
Why it matters
Augmentation focuses design on the human–machine system rather than treating labour replacement as the only objective.
ALSO: Human augmentation · Copilot
Connected concepts
Why it matters
Automation creates value when the task, exception paths and accountability are clear—not merely because AI can attempt it.
ALSO: Autopilot
Connected concepts
B
Why it matters
Benchmarks aid comparison, but high scores may not transfer to a specific workflow or production environment.
Connected concepts
C
Why it matters
It can automate systems without dedicated APIs, but needs careful permissions and recovery from visual ambiguity.
Connected concepts
Why it matters
Provenance can provide evidence about origin and modification, though it does not by itself prove that content is true.
Connected concepts
Why it matters
A larger window permits more material, but relevance, attention and cost still need active management.
Connected concepts
D
Why it matters
AI governance cannot work when the underlying data has unclear provenance, permissions or accountability.
Connected concepts
Why it matters
Lineage supports auditability, rights management, quality control and incident response.
Connected concepts
Why it matters
The integrity of data pipelines is part of AI security, not only a data-quality concern.
Connected concepts
Why it matters
The coverage, quality, rights and biases of a dataset shape what a model can learn and where it may fail.
Connected concepts
Why it matters
It made modern advances in language, vision, speech and generative media possible.
Connected concepts
Why it matters
Deepfakes increase the cost of trust and make provenance, verification and media literacy more important.
Connected concepts
Why it matters
Diffusion became a leading approach for high-quality image, audio and video generation.
Connected concepts
E
Why it matters
Edge deployment can improve privacy, resilience and latency while operating under tighter hardware limits.
Connected concepts
Why it matters
Embeddings enable semantic search, recommendations, clustering and retrieval for AI applications.
Connected concepts
Why it matters
A model can sound impressive while failing the exact task, population or risk threshold that matters.
ALSO: Eval · Evals
Connected concepts
Why it matters
Useful explanations support review, challenge and accountability, especially in consequential decisions.
ALSO: Interpretability · XAI
Connected concepts
F
Why it matters
Features determine what evidence a system can use; irrelevant or leaked features can make results misleading.
Connected concepts
Why it matters
Fine-tuning can make repeated specialised behaviour more reliable, but it is not the best tool for simply adding current facts.
Connected concepts
Why it matters
One capable base model can support many products through prompting, retrieval or fine-tuning.
Connected concepts
G
Why it matters
It changes AI from a back-office prediction engine into a general-purpose interface for creating and transforming work.
ALSO: GenAI · Gen AI
Connected concepts
Why it matters
GPU availability, memory and energy use are major constraints on AI cost and scale.
ALSO: GPU · AI accelerator
Connected concepts
Why it matters
Grounding reduces unsupported answers and makes responses easier to verify.
Connected concepts
Why it matters
Reliable systems use layered controls—validation, permissions, monitoring and human escalation—not a single prompt.
Connected concepts
H
Why it matters
Confident language can hide uncertainty, so consequential outputs need grounding, verification and appropriate human review.
Connected concepts
Why it matters
Human oversight is most valuable when decision rights and escalation thresholds are explicit, not ceremonial.
ALSO: HITL · Human oversight
Connected concepts
I
Why it matters
Inference is where users experience the model and where speed, cost and reliability become operational concerns.
Connected concepts
Why it matters
The endpoint is where model capability meets authentication, scaling, latency, monitoring and cost control.
ALSO: Model API
Connected concepts
Why it matters
It helps turn a base model that predicts text into a useful assistant that responds to intent.
Connected concepts
J
Why it matters
Jailbreak resilience matters, but application security must not rely on the model refusing every harmful request.
Connected concepts
L
Why it matters
In supervised learning, inconsistent or biased labels directly teach inconsistent or biased behaviour.
Connected concepts
Why it matters
LLMs power conversational assistants, coding tools, search interfaces and many agent systems.
ALSO: LLM
Connected concepts
Why it matters
A capable system can still fail as a product if it is too slow for the workflow it serves.
Connected concepts
M
Why it matters
Most modern AI capability comes from learned statistical behaviour rather than hand-written logic.
ALSO: ML
Connected concepts
Why it matters
It can increase total model capacity without using every parameter for every request.
ALSO: MoE
Connected concepts
Why it matters
Production reliability depends on repeatable pipelines and ownership after the prototype works.
ALSO: MLOps · LLMOps
Connected concepts
Why it matters
The model is the reusable capability produced by training and later called during inference.
Connected concepts
Why it matters
Good documentation helps users judge fitness for purpose rather than treating a model name as assurance.
Connected concepts
Why it matters
A shared connection layer can reduce one-off integrations and make capabilities portable across compatible AI clients.
ALSO: MCP
Connected concepts
Why it matters
Distillation can move capability into faster, cheaper or private deployment environments.
ALSO: Knowledge distillation
Connected concepts
Why it matters
Production AI needs ongoing measurement because accuracy at launch does not guarantee accuracy later.
Connected concepts
Why it matters
Specialisation can help complex workflows, but coordination overhead may outweigh the benefit for simpler tasks.
Connected concepts
Why it matters
Real work is rarely text-only; multimodality lets AI operate across documents, screens, speech and the physical world.
Connected concepts
N
Why it matters
Neural networks underpin language models, computer vision, speech systems and most generative AI.
Connected concepts
O
Why it matters
Open weights can improve control, customisation and deployment choice, but do not necessarily mean the training data or code is open.
ALSO: Open weights
Connected concepts
Why it matters
Reliable agent systems depend as much on orchestration and recovery paths as on model capability.
Connected concepts
P
Why it matters
Parameter count can indicate model scale, but it does not by itself measure quality, intelligence or usefulness.
ALSO: Model weight · Weight
Connected concepts
Why it matters
Planning helps agents handle longer tasks, but each extra step creates more opportunities for error or drift.
Connected concepts
Why it matters
The right balance depends on the cost of false alarms versus missed cases.
Connected concepts
Why it matters
Clear prompts improve consistency, but cannot compensate for missing capability, evidence or governance.
Connected concepts
Why it matters
Agents that read external content may encounter hostile instructions hidden in pages, files, messages or tool output.
Connected concepts
Q
Why it matters
Quantization can lower memory, cost and latency, sometimes with a modest trade-off in quality.
Connected concepts
R
Why it matters
Extra reasoning can improve difficult analytical and coding tasks, but usually increases latency and cost.
Connected concepts
Why it matters
Adversarial testing exposes failure modes that routine quality checks often miss.
Connected concepts
Why it matters
It is used for sequential decisions, control and aligning model behaviour with preferences.
ALSO: RL
Connected concepts
Why it matters
It frames AI success as more than technical performance: impacts on people and institutions count too.
Connected concepts
Why it matters
RAG can make outputs more current, specific and verifiable without retraining the base model.
ALSO: RAG
Connected concepts
S
Why it matters
It signals genuine demand, but can expose sensitive data and create unmanaged legal, security and quality risks.
Connected concepts
Why it matters
Smaller models can be cheaper, faster, easier to run privately and strong enough for focused tasks.
ALSO: SLM
Connected concepts
Why it matters
Speech turns AI into a lower-friction interface and extends access beyond keyboards and screens.
Connected concepts
Why it matters
It is effective when the target is clear and enough trustworthy labelled data exists.
Connected concepts
Why it matters
It can address scarcity and privacy constraints, but may reproduce or amplify the assumptions of the generator.
Connected concepts
Why it matters
It is an important control layer, but it is not a security boundary on its own.
Connected concepts
T
Why it matters
Lower values often improve repeatability; higher values can increase variety but do not create deeper reasoning.
Connected concepts
Why it matters
Token counts affect context limits, response length, latency and usage cost.
Connected concepts
Why it matters
Tools let models work with live information and real systems instead of relying only on learned text patterns.
ALSO: Function calling
Connected concepts
Why it matters
Training creates the capability; inference applies that capability after training is complete.
Connected concepts
Why it matters
Keeping the test set genuinely separate reduces the risk of mistaking memorisation for generalisation.
ALSO: Data split · Train-test split
Connected concepts
Why it matters
Transformers are the dominant architecture behind modern language models and many multimodal systems.
Connected concepts
U
Why it matters
It helps discover hidden structure and build representations when labelled data is scarce.
Connected concepts
V
Why it matters
It is a common retrieval layer for grounding model responses in private or current knowledge.
ALSO: Vector store
Connected concepts
Why it matters
Vision models enable inspection, document understanding, medical imaging and visual computer use.
Connected concepts
W
Why it matters
World models are important for robotics, simulation and agents that must plan beyond the next immediate step.
Connected concepts
A note on language
Working definitions · Context over certainty
AI language moves quickly. Meaning still needs judgement.
These are concise working definitions, not claims that every researcher, regulator or vendor uses each term identically. The aim is a clearer conversation—and better questions.
Move from terms to practice