Generative AI Fundamentals
Generative AI learns patterns in existing data and uses those patterns to produce new content. It generates plausible outputs—not guaranteed facts.
Learning objectives
What you will be able to do
- Explain generative ai fundamentals accurately in your own words.
- Recognise how the concept appears in real AI products and professional work.
- Identify an important limitation, risk or evaluation requirement.
- Apply the concept through a practical activity and knowledge check.
Core explanation
Build the right mental model
Generative systems model a data distribution: the patterns that make text, images, audio or code resemble their training examples. At use time, sampling and instructions guide the creation of a new output.
Discriminative models learn a predictive relationship from inputs to targets; generative models represent a data distribution from which outputs can be produced. The same application may combine both—for example, classifying a request, retrieving documents and generating a grounded response.
Because generation is probabilistic, the same prompt can produce different answers. Quality control must address factual accuracy, copyright, privacy, bias and unsafe content.
Distinguish producing content from establishing truth
A generative system creates an output such as a paragraph, image, audio segment or code fragment. The usefulness of that output depends on the task. An imaginative story can be successful without being factual; a policy summary must preserve the policy’s conditions. Before opening a tool, decide whether you want variation, faithful transformation or a verifiable answer. The same fluent paragraph can be suitable for one purpose and unacceptable for another.
Generation is different from looking up an exact record. A chatbot can compose a sentence that was never stored as a complete sentence. It can also reproduce familiar wording or memorised material, so “newly generated” does not guarantee originality or permission to reuse. Treat attribution, factual accuracy and allowed data use as separate checks rather than assuming one solves the others.
Build a small generation contract
Give the assistant a purpose, approved input, audience, output format and boundaries. For a summary, specify what must be retained: dates, exceptions, responsibilities and numerical limits. Ask it to mark missing information rather than invent it. For creative brainstorming, specify how many alternatives you want and what makes an idea acceptable. A clear contract makes review easier even when the model fails to follow it.
The source material matters more than a request to “be accurate”. A grounded response uses information supplied or retrieved for the task, but it still needs checking. Retrieval can select the wrong passage, a summary can drop a condition, and a citation can point to a real page that does not support the sentence. Inspect the relationship between each important claim and its evidence.
Review content at the level of the decision it supports
Begin with important claims rather than spelling. In a course notice, a wrong submission date matters more than an awkward adjective. In generated code, an unsafe file operation matters more than elegant variable names. Separate factual checks from style checks and finish the factual checks before polishing. Preserve an original copy of the source so you can compare the output line by line.
Use public or invented material while practising. A task does not become harmless merely because the output is only a draft: private information can be disclosed at input time. Tool settings and institutional approval matter before submission. The fictional policy below contains no actual student information and is designed to make omissions visible.
Generation
Producing a new sample consistent with learned patterns.
Prompt
Instructions and context supplied at use time.
Sampling
Selecting among possible outputs according to model probabilities.
Grounding
Connecting generation to trusted source material or tools.
Worked example: summarising a fictional booking policy
Source card: “Practice rooms may be booked by registered course participants for up to 90 minutes. Bookings must be cancelled at least 12 hours before the start time. Access is available from 08:00 to 18:00 on weekdays. Staff may close rooms for maintenance.” Request: write a plain-language summary under 60 words without adding rules.
Illustrative weak draft: “Anyone can book a practice room for 90 minutes. Rooms are open daily, and you can cancel beforehand.” It broadens eligibility, changes weekdays to every day, loses the 12-hour deadline and omits the maintenance exception. These are content errors, even though the language is readable.
A better summary is: “Registered course participants may book practice rooms for up to 90 minutes on weekdays between 08:00 and 18:00. Cancel at least 12 hours before the booking starts. Rooms may close for maintenance.” Compare each clause with the source card: all four conditions are present and no new entitlement is introduced.
The result is an illustrative revision, not a report of a particular model’s performance. A useful review record would list the four omitted or changed conditions and show the corrected clause for each.
Apply your learning
Practical activity
- Spend 4 minutes writing a fictional three-rule policy for equipment loans. Include eligibility, a duration and an exception.
- Spend 5 minutes requesting or manually drafting a 50-word summary. Keep the source visible but review the summary as though it came from an untrusted assistant.
- Spend 6 minutes creating a source-to-claim table. Mark every condition as preserved, changed, omitted or invented.
- Spend 3 minutes revising the summary and writing one sentence about what an automated accuracy score would not establish.
Compare with an explained example answer
Example source: staff may borrow training microphones for two days; collection requires a booking number; damaged units must be reported before use. A valid summary preserves all three requirements without adding a deposit, a fee or permission for students.
The answer is acceptable when every material statement traces to the source and every necessary condition is represented. Different wording can be correct. A model’s confidence or a second fluent approval is not independent evidence of fidelity; the source comparison provides the evidence.
Write your answers in a learning journal. The value comes from connecting the concept to your own profession.
Knowledge check
Test your understanding
Q1Does a generated summary become reliable because it contains a citation?
Answer: No. The cited source must support the specific claim, and important conditions must survive the summary.
Q2What is the difference between a creative and a faithful transformation task?
Answer: Creative work may welcome invented alternatives within constraints. Faithful transformation must preserve the supplied meaning and should not introduce unsupported details.
Q3Why review inputs before generating a harmless-looking draft?
Answer: Privacy exposure can occur when source data enters a service. A low-risk output does not make every input appropriate.
If you cannot explain the answer without reading it, revisit the core explanation and example before continuing.
Continue with authoritative sources
Important industry resources
These links lead to official organisations, industry laboratories or established open-source learning projects.