AI news
AI model reviews
AI tools
How-to guides
Our staffAboutContact

What are foundation models in generative AI? A plain‑English guide for 2026

Foundation models are huge AI models trained once on broad data, then adapted to many jobs. Here's how they work, real examples, and their limits.

Rack of servers in a server room
Photo: Kevin Ache on Unsplash

Foundation models in generative AI are very large models trained once on huge, broad datasets (text, images, code, audio) so they can be adapted to many different jobs afterwards. ChatGPT, Claude and Gemini all run on them. If you've used any AI writing or image tool lately, you've almost certainly used a foundation model, even if nobody told you.

The name is a bit grand, I'll admit. But it points at something real: one expensive model sits underneath, and lots of cheaper products get built on top of it.

Short answer: foundation models in generative AI, defined

A foundation model is a general-purpose AI model, usually a neural network with billions of parameters, that learns patterns from a massive pile of mostly unlabelled data. Once it's trained, you can prompt it, fine-tune it or wrap it in an app to write emails, summarise contracts, generate images or answer support tickets. The model isn't built for any single task. That's the point.

The term itself is young. Researchers at Stanford's Center for Research on Foundation Models coined it in August 2021 in a long report called On the Opportunities and Risks of Foundation Models, written by more than 100 authors. They picked "foundation" because these models are the base that other systems stand on, and (this is the part people forget) because any crack in the base shows up in everything built above it.

Inside a foundation model: pre-training and adaptation

Most foundation models are built on the transformer, an architecture Google researchers described in 2017. A transformer reads a sequence of tokens (small chunks of words, pixels or sound) and learns which parts of the sequence matter to each other. If you want the basics of tokens first, our explainer on what a token is in AI covers it.

Training happens in two broad stages. First comes pre-training, where the model sees an enormous amount of data and does one simple exercise over and over: predict the next token, or fill in a missing piece. Nobody labels the data by hand. The model works out grammar, facts, coding conventions and visual styles on its own, because getting good at prediction needs all of them. This stage can take months on thousands of chips, and it's where most of the cost sits.

Then comes adaptation. A company takes the pre-trained model and narrows it: instruction tuning so it follows requests, human feedback so it's more helpful and less harmful, or fine-tuning on legal or medical text for a specialist product. Lots of apps skip this step and just write clever prompts.

Here's an analogy I find useful. A foundation model is a bit like a medical school graduate. Years of broad training come first; the specialism (cardiology, paediatrics) comes later and takes far less time. You wouldn't send a fresh graduate into heart surgery, and you probably shouldn't send a raw foundation model into a regulated workflow either.

AWS's explainer on foundation models makes a similar point in plainer terms: instead of building AI from scratch, teams "use a foundation model as a starting point to develop ML models that power new applications more quickly and cost-effectively." That sentence probably explains the business appeal better than any diagram.

Foundation model vs large language model

People mix these up constantly. A large language model (LLM) is one kind of foundation model, trained mainly on text. Foundation models also include image models, audio models and multimodal ones that handle several types of input at once. So every LLM is a foundation model, but not every foundation model is an LLM. Our pages on LLM AI and LLMs vs generative AI go further into the difference.

Real examples of AI foundation models

You'll recognise most of these names. What's less obvious is that each family is used in dozens of products their makers don't own.

Model familyMakerMain data typesWhere you meet it
GPTOpenAIText, images, audioChatGPT, Microsoft Copilot, many third-party apps
ClaudeAnthropicText, images, codeClaude apps and the Claude API
GeminiGoogle DeepMindText, images, audio, videoGemini, Google Search's AI features
BERTGoogleTextSearch ranking and classification tools (an early, smaller example)
Stable DiffusionStability AIImagesImage generators and design plugins

Google is now on its fourth generation of Gemini; the Gemini page lists Gemini 4 Argon as its flagship for Google AI Pro and Ultra subscribers. Google's AI search features sit on the same family, which is why our Google AI Mode review is really a review of a foundation model in a search box.

Open-weight models are a separate branch worth knowing about.

Their makers publish the trained weights, so anyone with enough hardware can run and fine-tune them. That's how a small startup can ship a specialist chatbot without spending millions on pre-training. If that interests you, our list of open source AI tools is a good next stop.

Benefits and limitations of foundation models

The big benefit is reuse. Pre-training is brutally expensive, but it only has to happen once per model; after that, thousands of companies can adapt the same model for a tiny fraction of the cost. Before 2018 or so, most teams trained a separate model for each task, one for sentiment, one for translation, one for spam. That world's mostly gone.

Foundation models also handle tasks nobody trained them for. Ask one to turn a recipe into a shopping list sorted by supermarket aisle and it'll usually manage, despite nobody ever setting that as a goal. I think this flexibility, more than raw quality, is why they spread so fast.

The limitations are just as real, and the Stanford authors flagged most of them in 2021:

  • They make things up. Prediction isn't the same as knowing, so confident wrong answers (hallucinations) still happen.
  • Flaws get inherited. A bias or blind spot in the base model can turn up in every app built on it.
  • Training data is murky. Most makers don't say exactly what went in, which has led to copyright lawsuits from authors, artists and publishers.
  • Only a few companies can afford to build them, so a handful of labs set the defaults for everyone else.
  • Running them uses a lot of electricity and water, though exact figures per query are hard to pin down.

That concentration point bothers me more than hallucinations, honestly. Hallucinations are improving year by year. A market where four or five firms own the base layer isn't obviously fixing itself.

Practical takeaways for you

For most people, the practical takeaway is simple: the app you're using matters less than the model under it. Two writing tools that both run on the same foundation model will probably give you similar quality, so compare price, privacy terms and features instead of the marketing.

A few habits help. Check which model a tool uses; good vendors say so on their pricing or docs page. Treat facts, figures and citations from any foundation model as drafts to verify, especially on health, money or legal questions. And read the data terms before pasting in client work (some free tiers may use your inputs for training, others don't).

For your career, the picture's mixed. Building foundation models is a small, very well-paid field. Building on top of them is a much bigger one, and there's a growing market for people who check their output; our pieces on remote AI software engineer jobs and AI review jobs look at both ends. My guess is that knowing how to evaluate a model's output will stay useful longer than knowing any one tool's buttons.

The next few years for foundation models

Some trends are already visible. Models are getting more multimodal, reading and producing text, images, audio and video in one system. They're also being turned into agents that plan and use software tools, which Google now pitches openly in its Gemini announcements. Smaller models that run on a phone or laptop are getting better too.

What's less certain? Whether scaling up keeps paying off at the same rate, whether courts force changes to how training data is collected, and whether regulation like the EU AI Act (which has specific duties for general-purpose AI models) changes who can afford to build them. I'd be cautious about anyone who claims to know how those three play out. My own bet is that the base layer stays concentrated while the layer on top gets more crowded, more specialised and cheaper for you.

FAQ

What are foundation models in generative AI?

They're large AI models pre-trained on broad, mostly unlabelled data such as text, images and code, which can then be adapted to many tasks. Generative AI tools like ChatGPT, Claude and Gemini are built on them. Instead of training a new model for each job, developers prompt or fine-tune one shared foundation model.

What is a foundation model in generative AI, in simple terms?

It's a general-purpose AI that learned from a huge amount of data. A company trains it once at great cost, then many apps reuse it for writing, chat, coding or images. Other products are built on top of it.

What is an AI foundation model, and who coined the term?

An AI foundation model is any large model trained at scale on broad data that can be adapted to a wide range of downstream tasks. Stanford's Center for Research on Foundation Models introduced the term in an August 2021 report. It covers language models, image models and multimodal models, and goes well beyond chatbots.

Is ChatGPT a foundation model?

Not exactly. ChatGPT is an app; the GPT models underneath it are the foundation models. OpenAI adapts those models with instruction tuning and human feedback so they work well as a chatbot, then adds features like file uploads and web search. The same GPT models also power other companies' products through OpenAI's API.

Keep reading