The earliest attempts to develop artificial intelligence were based on what we often call rule-based expert systems. In these models, responses were governed by detailed instructions on how data should be processed, based on expert knowledge.

For example, during the Cold War, the U.S. government attempted to replace its human translators with computers. The goal was faster translation of important documents from Russian to English. Using expert systems, sentences were translated word by word, but these systems generally failed to convey coherent meaning and context. Eventually, the US government returned to human translators.

As hardware improved, the 1980s saw the development of machine learning, which involves training models to find correct answers within large datasets. Such machine learning can be either supervised—meaning the dataset is labeled with correct answers by humans—or unsupervised, where the models must identify patterns in the data on their own.¹

This type of technology eventually enabled computers to defeat even the best human chess players by the 1990s. Throughout the 2000s, this development further enabled facial recognition, voice recognition, and —most notably— better translation programs.

The image below illustrates the relationship between artificial intelligence, machine learning, deep learning, and the development over time—from simple chess programs to deep learning in neural networks.

Illustrasjon: Relasjonen mellom kunstig intelligens, maskinlæring og dyplæring og utviklingen
over tid - fra enkle sjakkprogram til dyplæring i nevrale nettverk.
Fra Kunstig intelligens og personvernLinks to an external site.. (s.5), av Datatilsynet, 2018.

Illustration: Datatilsynet. (2018). Kunstig intelligens og personvern, s. 5. datatilsynet.no

With the recent introduction of large neural networks, the machines' ability to analyze patterns could be enhanced through what is known as deep learning. In deep learning, models—such as those for language—are formed by encoding such patterns into vectors, lists of numbers representing layers of nodes in the network. Because today’s hardware allows for the creation of highly complex networks, these vector patterns can model and recognize richer contexts than earlier language models.

Unlike expert systems, where the machine’s processing of input was fairly understandable, network models are so complex that we humans struggle to understand what happens before the answers appear. The internal workings are therefore often referred to as a “black box.” This applies, for example, to the large language models developed by OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), and Meta (Llama).

Reutter, L., Grolid, S. Å. og Birkeland, R. M. (2020). Kunstig Intelligens, NDLA.