Few technical terms are currently used as differently as “Artificial Intelligence.” A chess program from the 1990s, a spam filter, and a language model such as Claude all fall under this term, even though they have almost nothing in common technically.
Why the term is used so differently
The media uses “AI” as a catch-all term for anything that makes automated decisions. Marketing departments apply the label to products that sometimes use nothing more than simple statistics. In research, the individual methods are usually distinguished much more precisely. In everyday language, however, “AI” now almost exclusively refers to language models such as ChatGPT or Claude.
These four contexts—media, marketing, research, and everyday use—lead to four different meanings of the same word. As a result, a discussion about “AI” can fail over the definition before it even reaches the actual topic.
Artificial Intelligence is not a product
ChatGPT is not synonymous with Artificial Intelligence. Claude is not synonymous with Artificial Intelligence. Gemini is not synonymous with Artificial Intelligence.
All three are language models—a single subfield within a much larger discipline. Artificial Intelligence does not refer to a single program, but to an umbrella term for various methods that automate tasks previously performed primarily by humans: pattern recognition, language processing, and decision support.
Since when has Artificial Intelligence existed?
The term was coined in 1956 at a conference in Dartmouth. The following decades produced expert systems—programs with hard-coded rule sets for individual fields such as medicine and chemistry. Periods of intensive research were followed by so-called AI winters: periods during which funding declined because the results failed to meet expectations.
Machine learning fundamentally changed the field: systems learn patterns from data instead of having them programmed explicitly. Today’s language models are the latest result of this development.
Artificial Intelligence did not begin in 2022. Language models are the newest and most publicly visible part of a development that has been underway for seven decades.
Why there is no single definition
Different contexts define AI differently, and each of these definitions is correct within its own context:
- Research: Distinguishes methods with technical precision—rule-based, statistical, neural
- Legislation: Defines AI for specific purposes, such as making regulation practical (the EU AI Act uses its own broad definition)
- Business: Often uses the term as a marketing label, regardless of the actual technology
- Everyday use: Usually reduces AI to its most visible current form—today, language models
There is no universally accepted definition because each context serves a different purpose. In this fundamentals series, we use the term in a clearly defined sense on which the following articles build.
Regardless of the definition used, there is broad agreement about which technical methods are considered part of the field of Artificial Intelligence today.
Which technologies fall under the term
- Rule-based systems: Fixed if-then logic, programmed entirely by humans
- Machine learning: Patterns are learned from data instead of being programmed
- Neural networks: A machine learning method modeled after the way biological nerve cells process information
- Language models: Neural networks specialized for language processing
- Computer vision: Neural networks for image recognition and image generation
- Robotics: AI methods used to control physical systems
These methods are not competitors. Modern systems frequently combine several of them—for example, a language model with image understanding uses both technologies at the same time.
Why almost everyone talks about language models today
Language models are the first AI application that can be used without technical background knowledge—a text box is enough. Earlier AI technologies usually operated invisibly in the background: spam filters, credit scoring, and product recommendations. Language models are visible, directly accessible, and produce results that appear human.
This explains why hardly anyone talks about expert systems anymore, even though they were the dominant AI technology for decades: they operated invisibly. Language models operate visibly—and therefore shape the public perception of the entire field.
What distinguishes language models from other AI
Language models such as Claude process and generate text based on patterns learned from large collections of text. They differ from many traditional approaches in that they are designed for general-purpose language processing, and the same model architecture can be used for a wide variety of tasks: translation, summarization, and programming all rely on the same underlying system without requiring separate training for each task.
The technical details—how training works, what parameters are, and why models hallucinate—are covered in the next article in this series.
Terms that are often confused
These terms are often used interchangeably in everyday conversation, but they refer to different concepts:
- AI / Artificial Intelligence: The umbrella term discussed in this article
- Machine Learning: A subfield of AI—systems learn from data instead of programmed rules
- Deep Learning: A subfield of machine learning that uses multi-layer neural networks
- Neural Network: The technical structure on which deep learning is based
- Language Model / LLM (Large Language Model): A neural network specialized for language processing
- Generative AI: AI systems that create new content—text, images, or audio—instead of only classifying or predicting
Each term refers to a different layer. A language model is generative AI, which is based on deep learning, which in turn is a subfield of machine learning, which itself is a subfield of AI.
How this series defines AI
In this fundamentals series, we use the term Artificial Intelligence as an umbrella term for the technologies described in this article. The primary focus of the series is on modern language models, since they represent the most relevant current application. However, the underlying concepts—such as neural networks, training, and inference—apply to large parts of the field as a whole.
The following articles explore these topics step by step in greater depth: machine learning, neural networks, training and inference, tokens and context, and hallucinations.
What is intentionally not covered
This fundamentals series focuses on the technical aspects. Social, ethical, and philosophical questions are covered elsewhere.
With this foundation in place, the terminology on which the remaining articles in the series build has now been established.
Further reading:
Comments
Comments will be enabled soon.