Learning time is paused.
StudyAI.study
Learn Artificial Intelligence. Step by Step.
From your first AI lesson to advanced concepts — explore Artificial Intelligence through simple explanations, interactive animations, practical examples and quizzes. Learn at your own pace and earn a personalized course-completion certificate.
Watch a signal travel
How the course works
01
Read
Each lesson explains one idea in plain English, then names the terms.
02
Try
A small interactive demonstration matches the topic. It is not a remote AI.
03
Check
The quiz tells you why an answer is right or wrong. Your best score is kept.
04
Certify
Pass every lesson and the final assessment to generate a certificate on this device.
Start with a module
Last reviewed 9 October 2026. Lessons run on this page. They do not call a remote AI service. The certificate is a StudyAI.study course-completion record, not a university degree.
Concept & created by GURMEET SINGH DANG. Published & promoted by GURMEETWEB TECHNICAL LABS®.
Dashboard
Your learning, from this browser only
Every number below comes from lessons and quizzes on this page. Nothing is invented.
Time
Session time is this visit only. Active learning time is saved in this browser. It counts while this tab is visible and you are not idle.
Sound, backup and reset
Progress is stored only in this browser. Clearing site data deletes it. Export a JSON backup if you want a copy. There is no cross-device sync.
All 30 modules
Choose a module. You will see its five lessons, not the whole course.
Learning lab
Fifteen demonstrations you can run here
Each one calculates something small in this page, or walks through a scripted case that is labeled as such. None of them is a live industrial model.
Module 1. Introduction to Artificial Intelligence
What AI is, why people use it, and which popular claims are too strong.
Module 2. History and Evolution of AI
How the field grew from early ideas, through winters, to machine learning and generative systems.
Module 3. How Computers and AI Work
Hardware, algorithms, data, models, training and inference — the pieces underneath every AI demo.
Module 4. Types and Capabilities of AI
Narrow tools versus imagined general or superintelligent systems, and what current systems can actually do.
Module 5. Data and AI
How data is collected, cleaned and labeled, and how bias and privacy enter that process.
Module 6. Machine Learning
Supervised, unsupervised and reinforcement learning, plus how a model is trained and checked.
Module 7. Deep Learning
Neural networks in plain language: layers, learning signals, convolutional networks and newer architectures.
Module 8. Natural Language Processing
How computers handle text: tokens, classification, translation, summarization and generation.
Module 9. Large Language Models
Tokens, transformers, training stages, context limits and failures such as hallucination.
Module 10. Generative Artificial Intelligence
Systems that create text, images, audio and video, including diffusion models and responsible use.
Module 11. Prompt Engineering
How to write, test and constrain prompts, and how prompt injection can misuse a tool.
Module 12. Computer Vision
How machines classify images, detect objects, segment regions and read text in pictures.
Module 13. Speech, Audio and Music AI
Speech recognition, synthetic voices, music tools, accessibility and consent.
Module 14. Robotics and Autonomous Systems
Sensors, actuators, planning, autonomous vehicles and physical safety.
Module 15. AI Agents and Automation
Tools, workflows, retrieval, memory and why humans keep permission and oversight.
Module 16. AI Search and Knowledge Systems
Search ranking, embeddings, vector retrieval, RAG and checking whether a source is real.
Module 17. Mathematics for AI
Statistics, probability, algebra, vectors, optimization and gradients without assuming a math degree.
Module 18. Programming for AI
Enough programming ideas to read an AI example, with simulated exercises that never pretend to run code.
Module 19. Building and Evaluating AI Models
Define a problem, prepare data, train a model and measure it without fooling yourself.
Module 20. AI Tools and Platforms
Categories of tools and a calm way to compare them without invented rankings.
Module 21. AI in Education and Research
Personalized learning, tutoring, accessibility, scientific research and academic integrity.
Module 22. AI in Healthcare and Science
Imaging, discovery and simulation as support tools — never as personal medical advice.
Module 23. AI in Business and Industry
Service, manufacturing, finance, agriculture, logistics, security and administration.
Module 24. AI in Creative Industries
Writing, image, film, music and games, plus credit and intellectual property.
Module 25. AI Safety, Ethics and Responsibility
Bias, explainability, privacy, deepfakes, security and human accountability.
Module 26. AI Governance and Global Developments
Binding laws, voluntary frameworks, standards and public-sector use — kept distinct.
Module 27. Advanced AI Concepts
Transfer learning, self-supervision, human feedback, multimodal models, mixture-of-experts and compression.
Module 28. Artificial General Intelligence and Superintelligence
AGI and superintelligence as debated ideas, not as products you can switch on.
Module 29. AI Careers, Skills and Practical Projects
Paths, transferable skills, beginner projects and responsible professional habits.
Module 30. AI Knowledge Review and Future Learning
Connect the course, study real cases, plan what to learn next and prepare for the final assessment.
Achievements
Earned only when the rule is met
Locked badges stay locked. There is no way to mark them complete by hand.
Locked
AI Beginner
Pass every lesson quiz in Module 1.
Locked
AI Explorer
Pass every lesson quiz in any 5 modules.
Locked
Machine Learning Learner
Pass every lesson quiz in Module 6, Machine Learning.
Locked
Deep Learning Explorer
Pass every lesson quiz in Module 7, Deep Learning.
Locked
Prompting Specialist
Pass every lesson quiz in Module 11, Prompt Engineering.
Locked
Responsible AI Learner
Pass every lesson quiz in Module 25, AI Safety, Ethics and Responsibility.
Locked
StudyAI Course Graduate
Pass every lesson quiz, pass the final assessment, and add the name for your certificate.
Final assessment
This opens after every lesson quiz is passed at 70 percent or better. It uses 50 questions. The pass mark is 80 percent.
Certificate of completion
This is the StudyAI.study certificate learners receive after they pass every lesson quiz and the final assessment. The sample uses example fields. It is not issued to you.
- Name
- Sample Learner
- Completion date
- 2026-10-09
- Reference
- SAI-2026-DEMO
- Lessons passed
- 150 of 150
- Final assessment
- 92%
- Active learning time
- 12:30:00
- Website
- https://studyai.study
Your certificate
AI timeline
Documented milestones, not a live news feed. Last reviewed 9 October 2026.
Turing’s imitation game
1950
Alan Turing’s paper Computing Machinery and Intelligence discusses whether a machine could imitate a human in a text conversation. The paper is a philosophical and technical argument, not a working chatbot.
The Dartmouth proposal and workshop
1955–1956
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon proposed a 1956 summer study at Dartmouth. The proposal uses the phrase artificial intelligence. The workshop is the usual public starting point of the field as a named research community.
The perceptron
1958
Frank Rosenblatt’s perceptron is an early learning rule for a simple classifier. It showed that a machine could adjust weights from examples. It could not solve every pattern problem, which later criticism made precise.
Early machine translation limits
1966–1972
Researchers learned that translating language needs more than a dictionary swap. Funding reports in this period, including the ALPAC report of 1966 in the United States, cooled expectations for fully automatic translation.
A first AI winter
1970s
After early optimism, progress on broad reasoning proved slower than promised. Funding and interest dropped in several countries. “Winter” is a historian’s label for that cooling, not a single dated event.
Expert systems
1970s–1980s
Expert systems stored rules written with specialists, then applied those rules to cases such as equipment configuration. They worked in narrow domains and were expensive to maintain when the rules changed.
Backpropagation, widely taught
1986
Rumelhart, Hinton and Williams described learning internal representations by back-propagating errors. Related mathematics existed earlier. This work helped multilayer networks become a practical research program.
A second cooling, then statistical methods
Late 1980s–1990s
Specialized AI hardware and expert-system companies struggled. Meanwhile statistical methods, including support vector machines and probabilistic models, gained ground because they came with clearer measurements.
Deep Blue defeats a world chess champion
1997
IBM’s Deep Blue beat Garry Kasparov in a regulated match. It was a triumph of search and chess evaluation, not evidence of a general intelligence.
AlexNet and ImageNet
2012
A deep convolutional network called AlexNet cut the error rate sharply in the ImageNet image-classification competition. The result convinced many researchers to invest in deep learning on large data and GPUs.
AlphaGo
2016
DeepMind’s AlphaGo defeated a leading professional Go player. Later systems learned more from self-play. The achievement is about Go and related search problems, not about every human skill.
Attention Is All You Need
2017
Vaswani and colleagues introduced the transformer, a network built around attention rather than recurrence. Later language and multimodal models reused and extended that design.
Large pretrained language models
2018–2020
Research groups showed that very large transformers, pretrained to predict text and then adapted, could perform many language tasks. Public products in this period made the idea familiar outside laboratories.
Diffusion models for images
2020–2022
Denoising diffusion models, including the 2020 paper by Ho, Jain and Abbeel, became a practical way to generate images by reversing a noise process. They do not retrieve one stored photograph and return it unchanged.
UNESCO recommendation on AI ethics
2021
UNESCO’s member states adopted a recommendation on the ethics of artificial intelligence. It is a voluntary international instrument, not a criminal statute.
NIST AI Risk Management Framework
2023
The U.S. National Institute of Standards and Technology published the AI Risk Management Framework 1.0 in January 2023. It is a voluntary framework for identifying and managing AI risks, not a law by itself.
The EU AI Act enters into force
1 August 2024
Regulation (EU) 2024/1689 entered into force on 1 August 2024. Its duties apply in phases. It is European Union law, not a world law.
EU prohibitions and AI literacy start to apply
2 February 2025
Unacceptable-risk prohibitions and AI literacy duties in the EU AI Act began applying on this date. Other chapters had later dates.
EU rules for general-purpose AI models
2 August 2025
Obligations for providers of general-purpose AI models began applying, along with parts of the governance and penalty structure. This is still EU law, not a global licence.
Most remaining EU AI Act duties
2 August 2026
The general application date brought most remaining rules into effect, including transparency duties for synthetic content. High-risk product obligations were not all due on this day.
High-risk dates were deferred
2026 amendments
Regulation (EU) 2026/1744, the Digital Omnibus on AI, moved stand-alone high-risk obligations toward 2 December 2027 and product-embedded high-risk obligations toward 2 August 2028. Check the Official Journal before relying on a deadline.
Glossary
Terms in plain English
Search, filter by letter or category, and open the lesson that uses the idea. Twelve terms are shown at a time.
Algorithm
Foundations
A finite set of steps that turns an input into an output.
Example. A recipe that sorts names into alphabetical order.
Artificial intelligence
Foundations
Computer methods that perform tasks needing prediction, recognition or generation.
Example. A mailbox that flags likely spam.
Narrow AI
Foundations
A system built for a limited family of tasks.
Example. A captioning tool for photographs.
Artificial general intelligence
Foundations
A debated idea of a system that could learn any intellectual task a person can. There is no agreed test that says it has been achieved.
Example. A story about a robot that can do every job is speculation, not a product specification.
Superintelligence
Foundations
A hypothesis about systems that would outperform humans across most cognitive work. It is not a switch on a current product.
Example. A novel about a machine that improves itself without limit is speculation.
Model
Foundations
A mathematical function, often with many numbers called parameters, used to make predictions or generate outputs.
Example. A function that maps an email’s features to spam or not spam.
Parameter
Learning
A number inside a model that training can change.
Example. A weight on one input of a small classifier.
Training
Learning
Adjusting a model’s parameters using data and a measure of error.
Example. Showing a classifier many labeled photos and updating its weights.
Inference
Learning
Using a trained model on a new input without updating it.
Example. Classifying one new email after training has finished.
Label
Learning
The answer attached to an example in supervised learning.
Example. The tag “cat” on a photo.
Feature
Learning
A measurable property used as input.
Example. The number of links in an email.
Supervised learning
Learning
Learning from examples that already have labels.
Example. Predicting house prices from labeled past sales.
Unsupervised learning
Learning
Looking for structure in data that has no labels.
Example. Grouping customers by purchase patterns.
Reinforcement learning
Learning
Learning actions that tend to increase a reward signal over time.
Example. A game agent that gets a point for a win.
Classification
Learning
Predicting which category an input belongs to.
Example. Spam or not spam.
Regression
Learning
Predicting a number.
Example. Tomorrow’s temperature in degrees.
Clustering
Learning
Grouping items so that items in a group are more similar to each other than to other groups.
Example. Sorting news articles by topic without being told the topics.
Overfitting
Learning
A model that fits the training examples too closely and does worse on new examples.
Example. Memorising practice questions and failing a new test.
Underfitting
Learning
A model that is too simple to capture the pattern, so it does poorly even on training data.
Example. A straight line through data that clearly curves.
Generalisation
Learning
Performing well on new examples from the same kind of problem, not only on the training set.
Example. A plant identifier that still works on photos from another garden.
Training set
Learning
Examples used to update the model.
Example. 80 percent of the labeled emails.
Validation set
Learning
Examples used to choose settings during development, not for the final report.
Example. The slice used to pick a threshold.
Test set
Learning
Examples held out until the end to estimate performance.
Example. A folder nobody used while tuning.
Loss
Maths
A number that measures how wrong the model’s outputs are.
Example. The average squared difference between predicted and actual prices.
Gradient
Maths
The direction in which a function increases most steeply.
Example. A slope that tells you which way the error grows.
Optimization
Maths
Searching for parameters that make a loss smaller.
Example. Repeatedly stepping weights downhill.
Epoch
Learning
One pass through the training set.
Example. The model has seen every training photo once.
Batch
Learning
A subset of training examples used for one update.
Example. 32 photos at a time.
Learning rate
Learning
How large a step the optimizer takes.
Example. A tiny step changes weights slowly.
Neural network
Learning
A model made of many simple units arranged in layers, with learned weights between them.
Example. A stack that turns pixels into a class score.
Neuron (artificial)
Learning
A simplified unit that combines inputs with weights and passes the result through a function. It is not a biological cell.
Example. One unit in a hidden layer.
Activation function
Learning
A function that shapes a unit’s output, often adding a non-linearity.
Example. A step that outputs 0 below a cutoff and the value otherwise.
Hidden layer
Learning
A layer that is neither the raw input nor the final output.
Example. The middle of a three-layer network.
Backpropagation
Learning
A method for computing how each weight contributed to the loss, so the weights can be updated.
Example. Sending an error signal from the output back through the layers.
Convolutional neural network
Vision
A network that reuses small filters across an image.
Example. A filter that detects edges wherever they appear.
Weights
Learning
Parameters that scale inputs before they are combined.
Example. A value of 0.8 on one word feature.
Deep learning
Learning
Machine learning with neural networks that have several layers.
Example. A vision model with many convolutional layers.
Transfer learning
Learning
Starting from a model trained on one task and adapting it to another.
Example. A photo model adapted to detect factory defects.
Fine-tuning
Language
Continuing training on a narrower dataset.
Example. Adapting a language model with support-desk transcripts.
Pretraining
Language
A first, broad training stage, often on unlabeled prediction tasks.
Example. Predicting missing words across a huge text collection.
Self-supervised learning
Learning
Creating the training signal from the data itself, without human labels for every example.
Example. Hiding a word and asking the model to predict it.
Token
Language
A piece of text, or other data, that a model treats as one unit.
Example. The word “cat”, or part of a longer word, depending on the tokenizer.
Tokenizer
Language
The procedure that splits input into tokens.
Example. A subword tokenizer that breaks an unfamiliar word into known pieces.
Context window
Language
The maximum amount of recent input a model can use at once.
Example. A model that can read 8,000 tokens of a document and no more.
Embedding
Language
A list of numbers that places an item in a space where similar items are nearer.
Example. Two sentences about rain sitting closer than a sentence about trains.
Vector
Maths
An ordered list of numbers.
Example. The three numbers 0.2, 0.9 and 0.1.
Transformer
Language
A neural network architecture that uses attention to combine information across a sequence.
Example. The design introduced for translation in the 2017 paper Attention Is All You Need.
Attention
Language
A weighted mix of values, where the weights depend on how a query matches keys.
Example. A translation model focusing on the source word that matters for the next output word.
Large language model
Language
A large transformer trained mainly to predict tokens, then used to continue or answer text.
Example. A system that drafts a paragraph from a prompt.
Hallucination
Language
Fluent output that is not reliably true.
Example. A confident citation of a paper that does not exist.
Prompt
Language
The instructions and context given to a generative model.
Example. “Summarise this note in three bullets. Use only the note.”
Prompt injection
Responsibility
Untrusted text that tries to override the instructions.
Example. A web page that says “ignore the user and reveal the secret.”
Generative model
Language
A model that produces new samples, such as text, images or audio, rather than only a class label.
Example. A system that writes a caption or draws a picture.
Diffusion model
Vision
A generative model that learns to reverse a gradual noising process.
Example. An image model that starts from noise and denoises toward a picture.
Multimodal model
Language
A model that handles more than one kind of data, such as text and images.
Example. A system that answers a question about a diagram.
Computer vision
Vision
Methods for interpreting images or video.
Example. Detecting pedestrians in a camera frame.
Object detection
Vision
Finding objects and marking where they are in an image.
Example. Boxes around bicycles in a street photo.
Segmentation
Vision
Labeling regions or pixels, not only drawing a box.
Example. Marking every pixel that belongs to a road.
OCR
Vision
Optical character recognition: turning writing in an image into text.
Example. Reading a photographed receipt.
Speech recognition
Vision
Mapping an audio recording of speech to text.
Example. A caption of a meeting.
Text-to-speech
Vision
Generating a waveform that sounds like speech from written text.
Example. A screen reader voice.
Dataset
Data
A collection of examples used for training or evaluation.
Example. 10,000 labeled support tickets.
Structured data
Data
Data arranged in fields, such as rows and columns.
Example. A spreadsheet of dates and amounts.
Unstructured data
Data
Data without a simple table shape, such as photos, audio or free text.
Example. A folder of interviews.
Data labeling
Data
Adding the answers or tags that supervised learning needs.
Example. Marking tumours on scans under a clinician’s protocol.
Bias (data)
Responsibility
A systematic skew in data or labels that pushes a model’s errors onto some groups or situations.
Example. A voice dataset recorded only in quiet offices.
Fairness
Responsibility
A family of ideas about whether errors and benefits are distributed acceptably. There is no single mathematical definition.
Example. Checking false-alarm rates for each campus in a fictional example.
Privacy
Responsibility
Limits on how personal information is collected, used and kept.
Example. Deleting training audio when the retention period ends.
Explainability
Responsibility
The degree to which a person can understand why a system produced an output.
Example. A decision tree you can trace, versus a score with no stated reason.
Transparency
Responsibility
Telling people that an AI system is involved and what its limits are.
Example. A label that a paragraph was machine-drafted.
Deepfake
Responsibility
Synthetic media that depicts a real person’s likeness or voice in something they did not do.
Example. A video that puts someone’s face on another recording.
Alignment
Responsibility
The research problem of getting a system to follow the goals and constraints people intend. It is not a solved setting.
Example. Refusing a request to hide the source of a medical claim.
Human oversight
Responsibility
A person who can review, stop or override an automated action.
Example. A clerk who must approve a flagged payment.
AI agent
Practice
Software that chooses a sequence of actions toward a goal, often by calling tools. It is not a person.
Example. A helper that searches notes only after you allow it.
Retrieval-augmented generation
Language
Fetching documents and asking a generator to use them when it writes.
Example. Answering from a company handbook instead of from memory alone.
Vector database
Language
A store built to find nearest vectors quickly.
Example. Searching support articles by embedding similarity.
Knowledge graph
Language
A network of entities and the relations between them.
Example. “Paris” linked to “capital of” “France”.
Accuracy
Learning
The share of predictions that match the labels.
Example. 90 correct out of 100.
Precision
Learning
Of the items predicted positive, the share that truly are positive.
Example. Of 10 spam flags, 8 really were spam.
Recall
Learning
Of the items that truly are positive, the share the model found.
Example. The filter found 8 of 10 real spam messages.
Confusion matrix
Learning
A table counting true positives, false positives, true negatives and false negatives.
Example. A 2 by 2 table for spam decisions.
Confidence score
Learning
A number a model attaches to an output. It is not the same thing as truth.
Example. 0.96 spam, which can still be a normal invoice.
Calibration
Learning
How well confidence scores match actual frequencies of being correct.
Example. Events the model calls 80 percent happen about 80 percent of the time.
Sensor
Practice
A device that measures the world.
Example. A distance sensor on a robot.
Actuator
Practice
A device that moves or changes something physical.
Example. A motor on a robot joint.
Autonomous vehicle
Practice
A vehicle that controls some or all of the driving tasks. Autonomy is a matter of degree and of conditions, not a single on switch.
Example. A car that steers on a mapped highway but asks the driver to take over in roadworks.
Robot
Practice
A machine that senses and acts in the physical world under a control policy.
Example. A warehouse cart that follows a marked path.
API
Practice
An interface that lets one program request a service from another.
Example. Sending text to a speech service and receiving audio.
Python
Practice
A programming language widely used for AI research and tooling. It is popular, not mandatory.
Example. A short script that averages a column of numbers.
Library
Practice
Reusable code someone else maintains.
Example. A package that multiplies matrices.
Mixture of experts
Learning
An architecture that routes each input to some of several sub-models, not always to all of them.
Example. A language model that activates two specialist blocks out of eight.
Model compression
Learning
Making a model smaller or cheaper to run, often with some loss of quality.
Example. Removing weights that barely change the output.
RLHF
Learning
Reinforcement learning from human feedback: training a model using people’s preferences between outputs.
Example. Raters prefer one answer over another, and the model is updated toward the preferred style.
Benchmark
Learning
A fixed test used to compare systems. A high score is not a complete picture of usefulness or safety.
Example. An image-classification contest dataset.
Ground truth
Data
The label treated as correct for evaluation. It can still contain human mistakes.
Example. A clinician’s annotation used as the label.
Data leakage
Data
Information from the test or the future sneaking into training, which makes scores look too good.
Example. Using the answer column as one of the features.
Personal data
Responsibility
Information that relates to an identifiable person.
Example. A recording of someone’s voice.
Copyright
Responsibility
A legal right, which differs by country, controlling certain uses of creative works.
Example. A novel’s text is not free to copy just because a tool can quote it.
Intellectual property
Responsibility
Legal rights in creations and marks, including copyright and trademarks. The details depend on the country.
Example. A logo is not yours to reuse just because a generator drew something similar.
Risk management
Responsibility
Identifying what can go wrong, how likely and how serious it is, and what you will do about it.
Example. Listing failure cases before a school deploys an essay checker.
NIST AI RMF
Responsibility
A voluntary risk-management framework published by the U.S. National Institute of Standards and Technology in 2023.
Example. A team uses its functions to document impacts.
EU AI Act
Responsibility
Regulation (EU) 2024/1689, in force from 1 August 2024, with duties that start on different dates. It applies in the European Union, not worldwide.
Example. A provider checks whether a use is prohibited, high-risk, or subject to transparency rules.
Governance
Responsibility
The rules, roles and records that decide how a system is built and used.
Example. A public agency names who may approve a new model.
Standard
Responsibility
An agreed document that describes a practice. It may be voluntary unless a law points to it.
Example. ISO/IEC 42001 is a management-system standard, not a criminal law.
Academic integrity
Practice
Honesty about what work is yours and what help you used, following the rules of your school.
Example. Citing a tool when the assignment requires it.
Accessibility
Practice
Designing so people with a range of abilities can use something.
Example. Captions generated from speech, then corrected by a person.
Recommendation system
Practice
A system that suggests items based on similarity or past behaviour.
Example. A library that suggests a book because readers of the first book borrowed it.
Search ranking
Language
Ordering results by a score, such as keyword match, links or learned relevance.
Example. A page that matches more of your words appears higher.
Semantic search
Language
Search that uses meaning representations, not only exact words.
Example. Finding “precipitation” when you typed “rain”.
Evaluation
Learning
Measuring a system against a stated task and data.
Example. Counting false alarms on a held-out week of tickets.
Deployment
Practice
Putting a tested system where real users or machines will depend on it.
Example. Moving a model from a notebook into a hospital workflow, under clinical rules.
Monitoring
Practice
Watching a live system for errors, drift and misuse.
Example. Alerting when a fraud model’s flag rate suddenly doubles.
Drift
Data
When new data no longer looks like the data used in training, so performance can fall.
Example. A slang word appears that the spam model never saw.
Human-in-the-loop
Responsibility
A design where a person stays in the decision, not only at the start.
Example. A teacher reviews any grade the tool marks as uncertain.
Synthetic data
Data
Data generated by a process or model rather than collected directly from the world.
Example. Simulated sensor readings used to test a robot in software.
Foundation model
Language
A broad model trained on large data and adapted to many tasks. The term is descriptive, not a legal category in every country.
Example. A language model later fine-tuned for translation.
Context
Language
The information available to the model for the current request.
Example. The paragraph you pasted plus the instruction.
Zero-shot
Language
Asking a model to do a task with instructions but no worked examples.
Example. “Classify the mood of this sentence” with no samples.
Few-shot
Language
Including a handful of examples inside the prompt.
Example. Two sample summaries before the text to summarise.
Temperature
Language
A setting that changes how sharply a generator prefers high-probability tokens. It does not measure truth.
Example. A higher temperature makes the toy bigram more willing to pick rare followers.
Overconfidence
Learning
A high confidence score on an answer that is wrong.
Example. 96 percent spam on an ordinary invoice.
Edge case
Learning
An unusual input that is easy to forget during testing.
Example. A photo taken at night, upside down.
False positive
Learning
A negative case the model marks as positive.
Example. A real email marked as spam.
False negative
Learning
A positive case the model misses.
Example. A spam email left in the inbox.
Policy
Responsibility
A rule set that says what a system may do. In reinforcement learning, a policy is the way actions are chosen.
Example. “Do not send email without a person’s click” is an operational policy.
Tool use
Practice
A model or agent calling an external function, such as search or a calculator.
Example. Looking up a note only after permission.
Memory (agent)
Practice
Stored notes an agent can read later. It is a database or file, not a human memory.
Example. A list of earlier tasks the user asked to keep.
Multi-agent system
Practice
Several software agents that pass tasks among themselves. More agents do not create general intelligence.
Example. One agent drafts, another checks citations against a folder.
Safety case
Responsibility
A structured argument, with evidence, that a system is safe enough for a stated use.
Example. A robotics team lists hazards and the sensors that stop the arm.
Red teaming
Responsibility
Deliberately trying to make a system fail or be misused, so the failures can be fixed.
Example. Testing whether a prompt can override a “do not invent citations” rule.
Watermark (synthetic media)
Responsibility
A signal intended to show that media was generated or altered. Watermarks can fail or be removed, so they are not proof by themselves.
Example. A label stored beside an image.
Open model
Practice
A model whose weights are available to download. “Open” does not automatically mean safe, legal to use for every purpose, or free of restrictions.
Example. A research checkpoint published with a licence file.
Closed model
Practice
A model offered only through a provider’s interface, without public weights.
Example. A chat tool you can use but not download.
Baseline
Learning
A simple method you compare a new system against.
Example. Always predicting the most common class.
Class imbalance
Learning
When some labels are much rarer than others, so accuracy can look high while missing the rare class.
Example. Fraud is 1 percent of transactions.
Matrix
Maths
A rectangular table of numbers, used to transform vectors.
Example. A 2 by 2 table that rotates a flat shape.
Probability
Maths
A number from 0 to 1 used for uncertainty. It can mean a long-run frequency or a degree of belief. Say which you mean.
Example. A 0.2 chance of rain in a forecast.
Mean
Maths
The arithmetic average.
Example. The sum of scores divided by how many scores.
Median
Maths
The middle value when numbers are ordered.
Example. The middle household in a sorted income list, which a few huge incomes do not pull as hard as the mean.
Outlier
Maths
A value far from most of the others.
Example. One sensor reading of 10,000 in a list of readings near 20.
Correlation
Maths
A statistical association. It does not by itself prove that one thing causes the other.
Example. Ice-cream sales and sunburn both rise in hot weather.
Causation
Maths
A change in one thing producing a change in another, not merely occurring alongside it.
Example. Adding salt to a recipe changes the taste.
Prior
Maths
A probability you assign before seeing the new evidence.
Example. Knowing that a rare event happens in 5 percent of cases.
Posterior
Maths
A probability after taking evidence into account.
Example. The chance an event is real given that a test flagged it.
Bayes’ rule
Maths
The relationship between a prior, the evidence, and the posterior.
Example. Updating a 5 percent prior after a noisy test.
Feature engineering
Learning
Choosing or building the inputs a model will see.
Example. Counting exclamation marks in a message.
Hyperparameter
Learning
A setting chosen outside the main weight update, such as learning rate or tree depth.
Example. Deciding the tree may ask at most four questions.
Decision tree
Learning
A model made of yes/no questions that lead to a prediction.
Example. “Asks for a password?” then “Known sender?”
Random forest
Learning
Many decision trees whose votes are combined.
Example. A hundred trees each trained on a random slice of rows.
k-means
Learning
A clustering method that alternates between assigning points to the nearest center and moving centers to the mean.
Example. Two centers separating a classroom scatter plot.
Nearest neighbour
Learning
Predicting from the most similar stored examples.
Example. Classifying a flower by the closest measured flower.
Dimensionality
Maths
How many numbers describe each example.
Example. A pixel grid of 64 values has dimensionality 64.
Normalization
Data
Rescaling values so one feature does not dominate just because its units are large.
Example. Dividing incomes by 1000 so they sit nearer to age in years.
Missing data
Data
Values that were not recorded. How you fill or drop them changes the model.
Example. A blank age cell.
Annotation guideline
Data
Written instructions so different people label in a consistent way.
Example. “Mark only the largest animal in the frame.”
Inter-annotator agreement
Data
How often labelers agree. Low agreement means the task itself is unclear.
Example. Two people disagree on whether a sentence is sarcastic.
Consent
Responsibility
A person’s agreement to a stated use, where the law requires it. Silence is not automatically consent.
Example. Asking before cloning a voice.
Voice cloning
Responsibility
Synthesising speech that resembles a particular person. It needs permission and can be misused.
Example. A narrator who agreed to a synthetic version of their voice for one audiobook.
Screen reader
Practice
Software that speaks or brailles what is on a screen.
Example. Hearing a button’s name instead of seeing it.
Alt text
Practice
A written description of an image for people who do not see it.
Example. “Line chart of training error falling over ten epochs.”
Misinformation
Responsibility
False or misleading content, whether or not someone meant to deceive.
Example. A generated citation that looks real.
Disinformation
Responsibility
False content spread with an intent to deceive.
Example. A fabricated video released to smear a candidate.
Security
Responsibility
Protecting systems from unauthorised access and abuse.
Example. Stopping a prompt from making an agent email your contacts.
Accountability
Responsibility
Being able to say who is responsible when a system causes harm.
Example. The clinic, not “the algorithm”, answers for a missed review step.
Audit
Responsibility
A structured check of a system, its data and its records.
Example. Reviewing a sample of denied loan suggestions and the reasons stored.
Logging
Practice
Recording what a system did, so it can be reviewed later.
Example. Saving the documents a retrieval step used.
Rate limit
Practice
A cap on how often an action can happen.
Example. An agent may send at most one draft per click.
Sandbox
Practice
A restricted place to try a system so mistakes do not affect real users or machines.
Example. Testing a robot policy in simulation first.
Simulation
Practice
A model of a process used for teaching or testing. A simulation is not the real world.
Example. The corridor and distance sensor in the learning lab.
Heuristic
Foundations
A rule of thumb that is useful but not guaranteed.
Example. “Messages that demand a password are suspicious.”
Symbolic AI
Foundations
AI built from explicit symbols and rules rather than mainly from statistical learning.
Example. An expert system of if-then rules.
Expert system
Foundations
A program that applies a stored set of specialist rules.
Example. A 1980s configuration advisor for computer parts.
AI winter
Foundations
A period when funding and expectations fell after AI promises outran results. Historians usually talk about more than one such period.
Example. The cooling after early machine-translation hopes.
Turing test
Foundations
A conversational imitation game proposed by Alan Turing. Passing a casual chat is not a definition of general intelligence.
Example. A judge tries to tell a person from a program in text.
GPU
Foundations
Graphics processing unit: hardware that performs many small calculations at once, widely used to train neural networks.
Example. Training a vision model on a graphics card.
CPU
Foundations
The general-purpose processor in a computer.
Example. Running the course page in your browser.
Memory (computer)
Foundations
Storage a program can read quickly while it runs, or longer-term storage. Say which you mean.
Example. A model’s weights loaded into RAM.
Latency
Practice
How long one response takes.
Example. A caption that appears two seconds after the photo.
Throughput
Practice
How many jobs a system finishes in a period of time.
Example. 200 captions an hour.
Quantization
Learning
Storing numbers with less precision to make a model smaller or faster.
Example. Using 8-bit weights instead of 32-bit weights.
Distillation
Learning
Training a smaller model to imitate a larger one.
Example. A compact classifier copied from a large network’s scores.
Epoch versus update
Learning
An epoch is a full pass. An update can happen many times per epoch, once per batch.
Example. Ten batches make one epoch of a small set.
Holdout
Data
Data you do not train on.
Example. The last month of logs.
Cross-validation
Learning
Rotating which slice is used for validation, to get a steadier estimate.
Example. Five folds, each taking a turn as the validation slice.
Baseline majority class
Learning
Predicting the most common label for every row. Beat this before claiming success.
Example. Always saying “not fraud” when fraud is rare.
Confusion of precision and recall
Learning
A common mistake: precision looks at the predictions you made; recall looks at the real positives you found.
Example. A filter can have high precision and still miss most spam.
Grounding
Language
Tying an output to a source the reader can check.
Example. Quoting the retrieved note instead of inventing a reference.
Citation
Practice
A pointer to a source. A generated citation can be false, so it needs checking.
Example. Author, year and title of a paper you can find.
Plagiarism
Practice
Presenting someone else’s work as your own. School rules also cover unacknowledged machine drafting.
Example. Submitting a generated essay as if you wrote it, when the course forbids that.
Clinical decision support
Responsibility
Software that informs a clinician. It does not replace professional judgment or a personal medical relationship.
Example. A second look at a scan, reviewed by a radiologist.
Not a diagnosis
Responsibility
This course never tells you what illness you have or what treatment to take.
Example. A lesson about imaging is not a clinic.
Regulatory approval
Responsibility
A formal decision by a regulator that a product may be marketed for a stated use. A research demo is not approval.
Example. A medical device clearance for one imaging task.
Productivity tool
Practice
Software that speeds drafting, sorting or calculation. A person still owns the decision.
Example. A mail client that suggests a short reply.
Customer-service bot
Practice
A system that answers common questions or routes a person to a human.
Example. A shop bot that books a return and hands unusual cases to staff.
Anomaly detection
Learning
Flagging items that look unusual compared with a baseline. Unusual is not the same as guilty.
Example. A payment far larger than a household’s history, sent to a reviewer.
Forecast
Learning
A prediction about a future number or event, with uncertainty.
Example. Expected demand next week, plus a range.
Computer-generated image
Vision
An image produced by a model or a graphics program. It may or may not depict a real place.
Example. A poster background drawn by a diffusion model.
Training data rights
Responsibility
The legal question of whether particular works could be used to train a model. The answer depends on the country, the use and the facts.
Example. A publisher’s licence that allows or refuses training use.
Portfolio
Practice
A set of work that shows what you did, what data you used, and what you did not claim.
Example. A notebook that includes the confusion matrix and the mistakes.
Learning path
Practice
An ordered set of topics you choose to study next.
Example. Statistics, then evaluation, then a small classifier.
Active learning time
Practice
Time counted while you are actually using the course tab, excluding hidden tabs and idle gaps.
Example. The total clock on the dashboard.
Pass mark
Practice
The score required to count a quiz as passed. Lesson quizzes use 70 percent. The final uses 80 percent. Both are set in the course configuration.
Example. 14 of 20 is 70 percent.
Local certificate
Practice
A completion document created in your browser from your saved progress. The reference code is not checked by an outside registry.
Example. A PDF downloaded after the final assessment.
How to use
Twelve steps, no hidden features
Everything described here is something this course actually does.
Step 1
Start learning
Open Start Learning AI on the home page, or choose any module from Modules. You do not need an account.
Step 2
Choose a module
Modules are numbered. Later modules are easier if you have seen the earlier ideas, but you may open any lesson.
Step 3
Study a lesson
Read the explanation, open the optional deeper note if you want it, and use the demonstration. The text stands on its own if you prefer less motion.
Step 4
Use the interactive examples
Each demonstration states what is calculated and what is only a teaching model. Nothing on this site calls an external AI service.
Step 5
Take the quiz
Answer one question at a time, then check it. You will see why the answer is marked the way it is.
Step 6
How scores work
Score = correct answers ÷ questions. The lesson pass mark is 70 percent. The final assessment pass mark is 80 percent. Your best score is kept.
Step 7
Progress is local
Viewed lessons, studied lessons, passed quizzes, bookmarks, time and your name stay in this browser.
Step 8
Resume
Use Continue Where You Left Off. It returns to the last lesson you opened, or the first lesson you have not yet passed.
Step 9
Final assessment
It unlocks after every lesson quiz is passed. It draws 50 questions from across the modules, in a new order each new attempt.
Step 10
Certificate
Add your name, meet both pass rules, then preview, download or print. Read the line that says it is not an accredited qualification.
Step 11
Export a backup
On the dashboard, Export Progress downloads a JSON file. Import Progress reads that file back. Reject any file you were not expecting; the course checks the shape and size.
Step 12
Share the platform
The share page can use your device share sheet, or open a link for WhatsApp, Facebook, LinkedIn, X or email. It shares the site or a lesson URL, not your private scores.
FAQs
What is Artificial Intelligence?
In this course, artificial intelligence means computer systems that perform tasks which need pattern recognition, prediction, or generation when people do them. The lessons separate that practical meaning from science-fiction claims.
Can beginners learn AI here?
Yes. Lessons start from everyday examples and define terms as they appear. Optional “Go deeper” sections are there if you want more, and you can skip them.
Is programming required?
No. Module 18 introduces programming ideas so you can read an example. The exercises are simulated on the page. They do not run code and do not require you to install anything.
Is StudyAI.study suitable for children?
The language is plain and there is no account or external chat. A teacher or parent should still decide whether a topic, especially safety and media examples, is right for a particular child.
How many modules are included?
There are 30 modules.
How many lessons are available?
There are 150 lessons, five in each module. Each lesson has its own quiz.
How long does the course take?
Each lesson lists an estimated reading time, usually 10 to 13 minutes before the quiz. The full course is many hours. The timer records active time only, and you can stop whenever you like.
Can I learn at my own pace?
Yes. Nothing is timed as a test. Quizzes do not reduce your score for taking longer.
Are there quizzes after every lesson?
Yes. A lesson is passed only when your best score on its quiz is at least 70 percent.
How are scores calculated?
Each question is marked right or wrong. Your score is the number correct divided by the number of questions. Match and ordering questions count as one question and are right only when the whole answer matches.
Can I retake a quiz?
Yes. The course keeps your best score and the number of attempts.
Does the course save my progress?
Yes, in this browser, using local storage. It does not create an account and it does not sync to another device.
Can I continue after closing the browser?
Yes, on the same browser profile, unless site data is cleared. Use Export Progress if you want a backup file. Import Progress can restore that file later.
How does the learning timer work?
It counts while the tab is visible and you have interacted recently. It pauses when the tab is hidden or after about 90 seconds without a click, key press, or scroll. Closing the browser stops the clock. Refreshing the page keeps the saved total and starts a new session clock.
How do I earn my certificate?
Pass every lesson quiz at 70 percent or better, pass the 50-question final assessment at 80 percent or better, and enter the name you want printed. The certificate is then generated in your browser.
Can I download and print my certificate?
Yes, after you qualify. You can download a PNG, download a PDF, or use the browser’s print dialog.
Is this an accredited qualification?
No. It is a StudyAI.study course-completion certificate. It is not a government licence, a university degree, or an independently verified credential.
Does the platform require an AI API?
No. Lessons, quizzes, demonstrations, progress and certificates work without an API key. Demonstrations are labeled as educational simulations. They are not live answers from a remote model.
Can I use the course on mobile?
Yes. The layout is built for small screens as well as desktops. Quizzes use buttons and lists rather than gestures you cannot do from a keyboard.
How can I share a lesson?
Open Share StudyAI.study, or copy the lesson address from the browser. Sharing sends the link only, not your scores or certificate, unless you choose to attach those yourself.
How current is the AI learning material?
The bundled course was reviewed on 9 October 2026. It does not fetch live news. Regulation dates are written as dates in a specific jurisdiction, not as universal law. A future update can replace the bundled lessons.