A foundation model is an AI model trained on broad data so it can serve as a starting point for many tasks. It is not the complete chatbot, search service or assistant you use. Those applications add instructions, interfaces, information sources and sometimes tools around a model.
That distinction helps explain a common puzzle: two applications can use the same underlying model yet behave differently. One might summarize a document; another might search a collection before answering. To understand the result, inspect the whole system, not just the model name.
What makes a model a foundation model
“Foundation” describes a reusable starting point. Instead of building every task-specific model from nothing, developers can adapt capabilities learned during earlier training. The term emphasizes breadth of training and reuse across tasks, not a particular product or a universal minimum number of parameters. Stanford researchers introduced this framing in their foundation models report.
A parameter is an adjustable numerical value inside a model; weights are one important kind. Training changes these values. A pretrained model has already completed an earlier training stage.
Not every pretrained model is necessarily a foundation model. A narrowly trained model can also be reused. The useful distinction here is the broad starting point and range of possible adaptations, rather than a badge that guarantees quality.
From pretrained model to useful application
For a foundation model, pretraining typically involves learning statistical patterns from a broad collection of examples by adjusting the model’s parameters. Later training can adapt those learned capabilities for particular tasks or behavior. When the trained model processes a new input, that is inference. Google’s generative AI glossary explains training, inference and adaptation.
An application also decides what information reaches the model and what happens to its output. It might supply a question alongside document passages, display citations, restrict available actions or route an uncertain answer to a person.
For a beginner, four things are worth separating:
- Model: the trained component that processes inputs.
- Context: information supplied for the current response.
- Application: software arranging the interaction and connecting other components.
- Evaluation: checks showing whether the complete system works for its intended task.
These distinctions offer different places to diagnose a problem. If a library assistant gives an outdated opening time, first ask whether it received the current schedule. Changing the model is not the only possible remedy.
Follow a fictional library assistant
Imagine a library creates an assistant to explain its services. The following situations are invented teaching examples, not results from a deployed system. Keep the underlying model unchanged unless a step explicitly says otherwise.
Change the instruction
The application first asks for a short explanation of borrowing rules. It then adds: “Use plain language for someone visiting a library for the first time.”
A different answer may result because the instructions changed. That does not establish a permanent change to the model’s learned weights.
This also does not tell us whether the service logs the interaction. Immediate model behavior and a provider’s separate storage or training policies are different questions.
Supply a document
Now the application includes a fictional current notice: “The reading room closes at 6 pm on Fridays.”
The user asks when the reading room closes on Friday. The assistant has relevant evidence in its working context. Supplying the notice does not itself train those hours into the model.
Suppose the answer says 8 pm. The notice was available, but the generated answer failed to follow it. Having evidence and using it correctly are separate checks.
Add retrieval
Instead of inserting the notice by hand, the application searches the library’s document collection and supplies a relevant passage.
This adds retrieval to the system. It does not mean the generator has been retrained. If the search returns an obsolete notice, a fluent answer may still be wrong. Our RAG guide follows this evidence trail in more detail.
Train for a particular task
Separately, developers might fine-tune a model using examples of library questions and appropriate responses. That is a training operation, unlike simply adding the current notice to a prompt.
It needs a defined objective and evaluation. More training is not automatically the right solution to frequently changing opening hours. The transfer learning guide explains how pretrained parameters can be adapted.
Connect a borrowing tool
Finally, imagine an authorized tool can look up the signed-in visitor’s loan record. The application checks access before returning that information.
Reading a loan record is not the same as renewing a book. Renewal would be a separate action requiring the appropriate permission and checks. The model’s ability to produce a convincing sentence does not grant it either capability.
Try the distinction: if the assistant gives the correct closing time after receiving a new notice, what changed in the setup? The supplied evidence. If it can now renew a book, what changed? The application’s connected actions and permissions. Neither observation alone demonstrates retraining.
How the names fit together
AI terminology often labels different aspects of the same system rather than competing products.
On narrow screens, scroll this table sideways. Keyboard users can focus the comparison and use the arrow keys.
| Term | What it describes |
|---|---|
| Foundation model | A broad, reusable model that can support many downstream tasks |
| Large language model | A large model concerned with language |
| GPT | A family of generative pretrained transformer models |
| Transformer | A model architecture, not an application |
| Chatbot | A conversational application or interface |
| AI agent | A system that can pursue tasks through actions, within its design and permissions |
A chatbot can use a language foundation model built with a transformer architecture. These descriptions can all apply without meaning the same thing. Some chatbots use rules rather than a foundation model.
For the underlying mechanisms, see GPT explained and transformers in NLP. For the surrounding software, read chatbots and AI agents.
Foundation models are not only chat generators
The category extends beyond conversational text. Stanford’s clarification of the term discusses language, images and other kinds of data.
A useful historical example is BERT. Google’s original BERT explanation describes pretrained language representations adapted to tasks such as sentiment analysis and question answering. This is not the same interaction as asking a chat generator to write an essay.
Likewise, “multimodal” describes support for more than one type of information, such as text and images. It does not tell you whether a particular application can access your photos, browse the web or perform actions. Check the actual system.
Compare the changes
Use this compact checklist when someone says an AI system has “learned” something new.
On narrow screens, scroll this table sideways. Keyboard users can focus the comparison and use the arrow keys.
| Change | What to inspect |
|---|---|
| New prompt or example | The instructions and examples supplied to the model |
| New document or retrieval result | The evidence available in context |
| Fine-tuning | The training process and which parameters were updated |
| Connected tool | The application’s integration, access controls and permitted actions |
For example, storing a visitor preference and supplying it in a later conversation can make an assistant appear to remember. That observation alone cannot distinguish stored context from a training update. Ask how the information is being reused.
Six questions to ask before relying on a system
Use the fictional library assistant as a practice case:
- Which model is involved? Identify the provider and version when disclosed, rather than relying only on the app’s name.
- What adaptation was done? Ask whether the system uses prompting, fine-tuning or another approach.
- Which evidence reaches it? For opening hours, inspect the current notice rather than accepting a confident answer.
- What can it access or change? Separate public information, private loan records and renewal actions.
- What happens to submitted information? Check storage, retention and possible training use before providing private material.
- How was it tested? Ask for task-relevant evidence, including questions it should decline to answer.
For this library example, a sensible test set would include a current notice, an obsolete notice, a missing answer and an unauthorized account request. These are proposed tests, not measured performance results. Model evaluation metrics provides further background on matching checks to the task.
Limits that remain
Broad capabilities do not establish reliability for every use. Generative applications can produce unsupported claims, expose sensitive information or reproduce harmful biases. The NIST Generative AI Profile describes these risks and emphasizes evaluation across the system’s lifecycle.
In the library example, inspect the claim, its evidence and the access boundary separately. A correct public opening time would not demonstrate that private borrowing records are protected.
A foundation model is therefore a starting point for building and testing, not a substitute for either.
Frequently asked questions
What is a foundation model in simple terms
It is a broadly trained AI model that can provide a starting point for many tasks. An application adds the surrounding software and information needed for a particular use.
Is a foundation model the same as an LLM
No. Many large language models serve as foundation models, but foundation models can involve images, audio or other data. The labels emphasize different properties.
Are foundation models always generative
No. Foundation models can learn reusable representations for tasks such as classification or retrieval, rather than being designed primarily to generate new content. BERT is a historical example.
Is a chatbot itself a foundation model
Usually the chatbot is the application you interact with. It may use a foundation model alongside instructions, retrieval and tools, or use rules without one.
Does typing a prompt permanently train the model
Ordinary prompting supplies input for inference; it does not itself update the model’s weights. A service may separately retain or use interactions under its data policies.
Does supplying documents change model weights
Providing documents as context does not itself change the generator’s weights. A separate training process could use documents, but that is a different operation.
Can a foundation model work with images or audio
Yes, depending on its design and training. Do not assume every model or application accepts every input type simply because it is called a foundation model.
Does a bigger model automatically mean a better application
No. Judge the complete application on the intended task, including evidence quality, response time, cost and appropriate safeguards, rather than size alone.
What should I check before submitting private information
Check who operates the service, who can access the data, how it is stored and retained, and whether it may be used for training. Do not assume a chat interface makes information private.
Key takeaways
- A foundation model is a reusable component, not the whole application.
- Prompting, supplying evidence, training and connecting tools change different things.
- Evaluate the resulting system on its actual task and permissions.
Where to Learn Next
Recommended next lesson: Continue with Transfer Learning Explained to learn how capabilities from an existing model can be reused and adapted for a new task.
Continue to Transfer Learning Explained
- Explore how applications supply evidence: Retrieval Augmented Generation Explained.
- Plan checks for a particular task: Model Evaluation Metrics Explained.