Generative AI, LLMs and AI Assistants Explained
Understand the model, the application around it and the checks each needs.
Part of the AI and Generative AI track: modules, projects and first roles →
By the GullyMentor team · 5 min read · Published 2026-10-18 · Updated 2026-11-06 · Sources listed at the end. Not a promise about mentorship outcomes.
Generative AI is software that produces new content, such as text or images, from patterns learned in training. An LLM is a model that works with language. An AI assistant is an application built around a model, with instructions, documents, tools and an interface. Knowing which layer failed is the first step in fixing a wrong answer.
What do these terms mean?
| Term | What it is | Typical failure |
|---|---|---|
| Generative AI | A broad kind of system that creates content | Fluent output that is not true |
| LLM | A model trained on language | Inventing a detail it was never given |
| AI assistant | A product combining a model, instructions, documents and tools | Using an outdated or wrong document |
| Integration | A connected action, such as updating a calendar | Doing something nobody approved |
How do I start?
1. Pick any assistant you can use for free and read its data and privacy notes before typing anything. 2. Write a short fictional document, such as a club handbook, so you know exactly what is true. 3. Ask questions whose answers are in the document, absent from it, and contradicted by it. 4. For each answer, mark it supported, unsupported or wrong, and note which layer you think caused the problem.
Think in layers
Imagine a campus assistant answering questions about a club event. The model generates a response. The application's instructions define the task and boundaries. A source document supplies the event details. The interface lets the student ask questions and review the answer.
If the assistant can update a calendar, that action comes from an integration around the model. It should have its own permission and confirmation rules. A model producing the words “booking confirmed” does not establish that a booking occurred.
A worked example
Your fictional document says: “The robotics workshop is on Saturday at 10 am in Room 4. Registration closes on Thursday.” A supported answer to “Where is it?” is Room 4. An answer about fees is unsupported because the document gives no fee.
Ask both questions and inspect the result. A useful assistant identifies missing information rather than adding a plausible price. Record the exact source and the response, then test a document with conflicting dates.
Context is not the same as learning
Supplying a document in a request gives the application information to use for that interaction. Do not assume this permanently trains the model. How a provider retains or uses data depends on the product and settings; check its current documentation before sharing anything sensitive.
Retrieval can help find relevant passages from a larger collection. It does not automatically establish that the retrieved passage is current, complete or trustworthy. Source selection and answer checking still matter.
What to test
- Does the answer match the supplied evidence?
- Does it distinguish missing information from a negative answer?
- Can untrusted document text redirect the assistant?
- Is a proposed external action independently authorised?
- Can the user inspect and correct the result?
Practice task
Write a one-page fictional handbook. Make a diagram showing the user, interface, model, handbook and any external tool. Mark which components only read information and which can change something.
Then remove the external tool and ask whether the remaining assistant can still perform the promised action. This exercise exposes a common mistake: describing a language response as a completed service. Include that distinction in your portfolio explanation.
What are the common mistakes?
- Treating a confident answer as a correct one.
- Pasting private or customer information into a tool without checking how it is handled.
- Assuming that giving a document to an assistant permanently teaches the model.
- Calling a model and an application the same thing, which makes failures hard to locate.
- Letting an assistant take an action without a separate approval step.
How do I check I got it?
Take an assistant answer you have seen and say which layer produced each part: the model's wording, the instructions, the supplied document or an external tool. Then name one test for each layer. If you can do this for two different answers, you have the right mental model.
What should I build next?
Learn to write and test instructions in prompt engineering for beginners. The AI career roadmap shows where this concept sits in a longer plan, and the AI and generative AI track describes guided practice.
Sources & further reading
Examples and practice recommendations are educational illustrations. Tool documentation and programme inclusions should be checked for your intended use.