What is artificial intelligence (AI)?

What is artificial intelligence (AI)? | Databricks Blog Skip to main content

Summary

AI’s impact comes from its breadth: the same pattern-recognition capabilities now power everything from fraud detection and medical imaging to personalization and content generation.

AI capability is advancing faster than AI reliability, making evaluation, human oversight and governance essential for production use.

The next phase of AI adoption will be defined less by model access and more by how effectively organizations connect models to trusted data, real workflows and measurable outcomes.

Artificial intelligence (AI) is a branch of computer science that lets machines perform tasks that normally require human intelligence, like learning, reasoning, problem-solving, recognizing patterns and making decisions. Put more simply, AI is software that learns from data and uses what it learns to make predictions, decisions or new content without being explicitly programmed for each task. Today’s AI runs everything from spam filters and recommendation engines to chatbots like ChatGPT and image generators. It draws on a range of techniques, most notably machine learning and generative AI , and it has moved from research labs into products people use every day. Stanford computer scientist Fei-Fei Li, writing in the Stanford Emerging Technology Review , places AI in the same category as the most transformative technologies in modern history: “AI is a foundational technology that is advancing other scientific fields and, like electricity and the internet, has the potential to transform how society operates.” Adoption is now scaling across every sector, from healthcare and financial services to retail and manufacturing, and the pace is accelerating. This page covers how AI works, the main types of AI, real-world examples, the limitations to watch for and a brief history of the field. What is AI in simple words? Think of AI as teaching a computer by example instead of writing step-by-step instructions. Show a system thousands of photos of cats and it learns to recognize cats on its own, not because someone told it that cats have whiskers and pointed ears, but because it has seen enough examples to figure out the pattern. AI is not “thinking” the way you or I do. It is finding patterns in data and using those patterns to make a best guess. That distinction matters: AI can get remarkably good results in narrow domains, but it does not understand anything in the human sense. The same pattern-matching approach that lets a model recognize cats also lets it spot cancer cells in a biopsy or flag fraudulent transactions among millions of legitimate ones. The underlying mechanism, finding patterns in data, is the same even when the application looks dramatically different. It is already part of everyday tools: search engines, voice assistants, navigation apps, spam filters and the recommendations you see on streaming services. How does AI work? Most modern AI works by learning patterns from large amounts of data, then applying those patterns to new situations. Instead of a developer writing rules (“if email contains ’free money,’ mark as spam”), the system is shown many examples and figures out the rules itself. The basic process looks like this: Collect data. The system is fed large amounts of relevant text, images, numbers, audio or video, and gaps in that data become gaps in the model. Train a model. An algorithm studies the data and tunes its internal weights and parameters until it reliably produces correct outputs, the computationally expensive step that can run for hours, days or weeks across many GPUs. Test and refine. The model is evaluated on a “held-out” test set it wasn’t trained on, where catching mistakes is far cheaper than catching them in production. Make predictions. Once trained, the model answers questions, classifies inputs, generates content or triggers actions on data it has never seen, the “inference” step end users actually interact with. Learn and improve. Many AI systems keep improving as they are exposed to more data and feedback, including signals from how people respond to their outputs.

Modern AI training is also a question of scale: frontier models train on trillions of tokens of text, run on tens of thousands of GPUs and cost hundreds of millions of dollars to build. Most organizations don’t train models from scratch. Instead, they fine-tune existing foundation models on their own data, which is dramatically faster and cheaper while still producing models tailored to a specific task or domain. The quality of an AI system depends heavily on the data it learns from: when training data is incomplete, biased or low-quality, AI outputs will be too. You can read more about the building blocks in our overviews of machine learning models and neural networks . What are the 4 types of AI? Researchers commonly group AI into four categories based on capability, a taxonomy usually attributed to Michigan State University researcher Arend Hintze, who proposed it in 2016 as a way to think about how AI might evolve. Only the first two categories exist in the real world today, while the other two remain open questions in research and philosophy. The taxonomy is useful because it draws a clean line between what AI can actually do now and what it can only do in theory or fiction. Type What it does Status today Example Reactive machines Responds to a specific input with a fixed output. Has no memory of past events, no ability to learn from experience and no model of the world beyond the immediate input. Among the earliest AI architectures; still in use for narrow tasks today. IBM’s Deep Blue, which defeated chess world champion Garry Kasparov in 1997, evaluated the board from scratch every turn. Simple spam filters that match keywords against a fixed list belong to the same category. Limited memory Learns from historical data to make predictions or decisions. Can use recent inputs to refine its outputs but does not retain a persistent long-term memory the way humans do. Powers nearly all modern AI in production, including the most capable systems. Self-driving cars that pull from short-term sensor history to anticipate the road ahead. ChatGPT, which holds the context of the current conversation but starts fresh in a new session. Netflix’s recommendation engine, which learns from viewing patterns over time. Theory of mind Would understand the emotions, intentions and beliefs of other people, the...