AI Certification Glossary

The credentials, and what they are actually called

These words are not interchangeable, and the difference between the first two decides what a credential is worth to an employer. Each definition below is how the providers themselves use the term.

AI certification

A credential showing you completed a structured AI course or passed an AI exam. Two very different things get called this: a completion certificate, awarded for finishing graded coursework, and an exam certification, awarded for passing a supervised test set by a vendor. Nearly everything sold online as an “AI certification” is the first kind.

Certificate

What a provider issues when you finish their course. It records completion, not an examined standard, and it does not expire. Coursera, DataCamp and Udemy all issue certificates — and the name on it, Google or IBM or Stanford, is usually what carries the weight rather than the format.

Certification

Awarded by a body that sets a standard and examines you against it — AWS, Microsoft, Google Cloud, NVIDIA, Databricks, Oracle. It normally means a proctored exam, a fee separate from any course, and an expiry date. Harder and more expensive, and the version a hiring manager is most likely to recognise. See our exams ranked by difficulty.

Professional certificate

A multi-course program on a platform like Coursera, built by a company — Google, IBM, Microsoft — to prepare you for one specific job role. Typically weeks to months of work, ending in a shareable credential.

Specialization

Coursera’s term for a themed series of related courses ending in one shared credential, such as the Machine Learning Specialization from DeepLearning.AI and Stanford Online. Narrower than a professional certificate, and usually built by a university rather than an employer.

Career track / skill track

DataCamp’s two shapes. A career track aims at a job title and runs to ten or more courses; a skill track covers one capability in a handful. Both are included in a subscription rather than bought individually, so the cost is time more than price. Untangled in the DataCamp credentials guide.

DataCamp certification

Separate from DataCamp’s tracks and routinely confused with them: eleven exam-based credentials, eight at Associate level and three at Professional, as checked in September 2026. These are examined; the tracks are coursework. Our DataCamp platform review covers both.

Vendor certification

Issued by the company whose product it covers. It carries weight because that vendor’s own customers hire against it — which is also its limit: an Azure credential says little to a team running AWS. Compared in AWS vs Azure vs Google.

Certification tier

Vendors stack their exams: foundational (broad concepts, no prerequisites), associate (working knowledge, some hands-on), then professional or specialty (design and operate real systems). The tier tells you whether you are ready far better than the vendor’s name does.

Foundational certification

The entry tier — AWS AI Practitioner, Azure AI Fundamentals (now exam AI-901), Google Cloud Generative AI Leader. Tests broad concepts rather than hands-on engineering and normally has no prerequisites. The easiest recognised certifications sit almost entirely at this level.

Recertification

The requirement to renew a certification after a validity period. AWS certifications expire and must be renewed; Microsoft fundamentals certifications do not. Worth checking before you book an exam, because renewal is a cost the advertised price never includes.

Digital badge

A verifiable online credential you can add to LinkedIn, issued for both certificates and exam certifications. Unlike a PDF it carries the issuer, the issue date and any expiry, so anyone can check it without asking you for anything.

Credly

The service most large vendors use to issue those badges. A Credly link is the shortest honest answer to “can you prove it?” — see AI certifications on your résumé.

Micro-credential

An umbrella term for any short, stackable credential below a degree. It is a category rather than a standard: the word tells you nothing about whether an examiner was ever involved.

MOOC

Massive open online course — the 2012-era name for the format Coursera, edX and Udacity introduced. Still common in academic writing about online learning, and rarely used by the platforms themselves now.

PD / CE / CPE credit

Professional-development or continuing-education hours some professions must log: teachers, nurses, accountants, lawyers. Whether an AI course counts is decided by your licensing body or district and never by the course provider, so ask before you pay.

How the courses themselves work

These are the terms that decide whether you finish — which, on a self-paced course nobody is chasing you about, is the part that usually fails.

Audit track

The name Coursera used for taking a course free without the certificate. Its enrolment options no longer use the word: most courses now offer a free Preview of the first module, graded work included, with later modules locked, and select courses offer “Full Course, No Certificate” — every lesson and graded assessment, but no certificate. Some programmes, Google AI Essentials among them, cannot be taken free at all; there the route is financial aid.

Self-paced

No cohort and no deadlines — you start when you enrol and finish when you finish. Almost every certification we cover works this way. The trade is that nothing external makes you finish, which is why completion realism is one of the six factors we score.

Prerequisite

What you are expected to know already before the first lesson, which is regularly not what the provider’s level badge implies. A track badged Beginner that opens on scikit-learn assumes you can write Python. See what “beginner” actually means.

Capstone project

A final, larger piece of work that pulls a whole programme together. It is the part of a certificate an interviewer can actually ask you about, and the part most often skipped once the certificate is already within reach.

Guided project

A short hands-on exercise with the steps supplied. Good for learning a tool, weak as portfolio evidence — the decisions were made for you before you started.

Hands-on lab

A sandboxed environment — a cloud console, a notebook, a browser IDE — where you run the real tool instead of watching a video of someone else running it. Often the biggest difference between two courses that look identical on paper.

Proctored exam

A supervised test, online or at a test centre, where your identity is verified and your screen and room are monitored. Cloud certifications use them; most Coursera certificates do not.

Practice exam

A timed mock paper. Its value is diagnostic rather than motivational: the score matters far less than which domains you lost marks in. How to use practice exams covers the method.

Exam voucher

A prepaid code covering an exam fee, often bundled with training or issued by an employer. See getting your employer to pay.

Completion rate

The share of enrolled learners who actually finish. Providers rarely publish it. One of our six scoring factors, completion realism, exists because of it — and it is why we record course length in hours rather than months.

Paying for it

Three ways of paying sit behind almost every course on this site, and they fail differently. A subscription punishes taking your time; a one-off purchase does not, and buys a certificate nobody screens for.

Coursera financial aid

A per-course application for a discount off a paid Coursera course, certificate included, whose size Coursera says may depend on your application and where you live. Coursera asks you to allow up to 16 days, and approval starts a 180-day completion window — so a five-course certificate means five applications and five clocks. Method in the financial aid guide.

Coursera Plus

A subscription covering most Coursera specializations and professional certificates, billed monthly or annually. It beats paying per certificate only if you finish more than one — the arithmetic is in is Coursera Plus worth it?

Free trial

A short preview of a paid tier that converts to a subscription unless you cancel. Useful for confirming a course is what you think it is, and not a way to finish one. See the Coursera free trial explained.

Subscription

You pay for a period of access and lose it when you stop paying. Coursera bills its programme subscriptions monthly, so the real price there is the monthly figure multiplied by how long you take; DataCamp sells monthly and annual plans, priced by country, and its pricing page shows both. On a monthly plan, finishing early is genuinely cheaper and drifting an escalating cost.

One-off purchase

You pay once rather than by the month, and nothing renews. Udemy sells courses this way, with what Udemy calls lifetime access, which its terms say does not cover courses taken through its Personal Plan subscription. The trade is the credential — there is no examined standard behind it and employers do not screen for it. Coursera vs Udemy weighs the two.

Employer-sponsored training

A learning budget, a platform licence your company already holds, or a reimbursement policy nobody advertises internally. The most under-claimed money in this market — how to ask for it.

The technology the courses teach

Enough of each term to read a syllabus and judge whether it covers what you need. These are working definitions rather than textbook ones.

Machine learning (ML)

Building systems that learn patterns from data instead of following hand-written rules. The core technical skill behind most AI engineering roles.

Supervised learning

Training a model on examples already labelled with the right answer — spam or not spam, price paid, disease present. Most of an introductory ML course is supervised learning, because it is where the method is clearest.

Unsupervised learning

Finding structure in data nobody has labelled: clustering customers, reducing hundreds of features to a handful. Usually a week or two of an introductory course rather than a course of its own.

Neural network

A model built from layers of simple units whose connection strengths are adjusted during training. “Deep” simply means many layers.

Deep learning

The branch of machine learning built on multi-layered neural networks. Powers image recognition, speech and modern language models, and is taught in depth by the Deep Learning Specialization.

Transformer

The neural-network architecture behind current language models. It reads a whole sequence at once and learns which parts of it matter to which other parts. Introduced in 2017, and the reason the field moved as fast as it did afterwards.

Token

The unit a language model actually reads and writes — roughly a short word or a fragment of one. Context limits, API pricing and rate limits are quoted in tokens rather than words, so the conversion matters when you are costing a project.

Large language model (LLM)

A neural network trained on very large amounts of text to understand and generate language. The technology behind AI chat assistants and the subject of most generative-AI curricula.

Generative AI

AI that produces new content — text, images, code, audio — rather than only classifying data. Most 2026-era beginner certifications are generative-AI certifications; the generative AI roundup ranks them.

Hallucination

When a model states something fluent and false. It follows from how these models work rather than being a bug awaiting a patch, which is why retrieval, grounding and evaluation appear on every serious AI engineering syllabus.

Prompt engineering

Writing instructions that get reliable, repeatable output from an AI system. A no-code skill with its own courses — and a contested job title, which is prompt engineering dead? takes seriously.

Embedding

A list of numbers representing a piece of text, an image or audio so that similar things sit near each other. What makes semantic search and retrieval possible at all.

Vector database

A store built to find the nearest embeddings quickly. The retrieval half of RAG, and the component most AI engineering courses have you stand up first.

RAG (retrieval-augmented generation)

Having a model look up relevant documents before it answers, so it works on private or current data instead of only what it memorised in training. The commonest thing companies actually build — see RAG courses and certifications.

Fine-tuning

Continuing to train an existing model on your own examples so it adopts a format, a tone or a narrow task. Usually the third thing to try, after prompting and retrieval, because it costs more and locks in whatever the data taught it.

AI agent

A system that uses a model to decide which actions to take — calling tools, reading the results, choosing the next step — rather than only producing text. Explained at length in what is agentic AI?

Natural language processing (NLP)

The field concerned with getting machines to work with human language. The older name for much of what LLMs now do, and still the word a syllabus uses when it means classification, extraction and sentiment rather than generation.

Computer vision

Machine learning applied to images and video — detection, segmentation, recognition. A distinct specialism with a distinct career path: how to become a computer vision engineer.

MLOps

Deploying, monitoring and maintaining machine-learning systems in production — the part that decides whether a model survives contact with real traffic. Ranked in MLOps certifications.

LLMOps

The same discipline aimed at language-model applications: prompt versioning, evaluation, cost and latency control, guardrails. Newer, less standardised, and increasingly what “AI engineer” means in practice — LLMOps courses.

No-code AI

Using AI through interfaces and prompts rather than by programming. Most beginner certifications sit here; the no-coding roundup lists the ones that stay there the whole way through.

AI literacy

Knowing what these systems can and cannot do, where they fail, and when not to use one. The stated goal of most employer-facing AI courses, and the thing the EU AI Act now expects staff to have.

Responsible AI

Fairness, transparency, privacy and safety treated as engineering requirements rather than a closing slide. Increasingly examined — see responsible AI certifications.

AI governance

The organisational half of the same problem: who signs off a model, what gets documented, how risk is classified and audited. A distinct career track with its own credentials — AI governance certifications.

EU AI Act

The European Union’s risk-tiered law on AI systems, with obligations that reach any organisation placing them on the EU market. It is the reason AI literacy training appeared on corporate budgets — the training guide covers who needs what.

Words that mislead

Five words that appear constantly on course pages and mean far less than they sound like. We have published one of them wrongly ourselves, which is why the list exists.

“Accredited”

In education, accreditation is something a recognised body grants to an institution, and it is what makes a degree transfer. Almost no AI certificate is accredited in that sense — where a course page uses the word, it usually means the provider approved it themselves. Treat it as marketing unless an accrediting body is named.

“Official”

Means only that the vendor whose product is covered made or endorsed the course. A third-party course can be excellent and unofficial; an official one can be a slide deck. We had to correct our own pages for describing a third-party exam-prep course as though AWS had written it, and we now record claims like that and check every page against them on each build.

“Lifetime access”

Udemy’s phrase for the access that comes with a course bought on its own. Its Terms of Use say “We generally grant you a lifetime access license, except when we must disable the content because of legal or policy reasons”, and that it does not apply to enrolments through a subscription such as Personal Plan. It is a statement about access, not about updates — which is why we record the month a provider last revised a course and score curriculum currency on it.

“Beginner”

A level set by the provider’s own marketing rather than a statement about your background. We publish our own level for every course, and on several it deliberately disagrees with the provider’s. The long version.

“Industry-recognised”

Almost never checkable as written. The narrower question is answerable: does this credential appear by name in adverts for the job you want? Are AI certifications worth it? works through the evidence.