What is artificial intelligence? Do chess computers, auto-completion, or learning systems already count – or does the AI age only begin when a super AI wakes up? Anyone who deals intensively with artificial intelligence will quickly find that every AI advance has two characteristics: Explaining it is pretty complicated. And there is always someone who writes in the comment column: “But it has nothing to do with real artificial intelligence!” Is that true?
The Beginnings Of Artificial Intelligence
The aim of this field: the study and development of intelligent agents. An intelligent agent is any device that senses its surroundings and takes measures that maximize its chance of successfully achieving its goals. They are usually divided into categories ranging from simple to complex agents. In the standard work “Artificial Intelligence: A Modern Approach” .
There are five categories:
- simple reflex agent
- model-based agent
- goal-based agent
- benefit-based agent
- learning agent
The order reflects a hierarchy: How effectively can the agent achieve its goals in a changing environment? A reflex agent reacts statically to environmental influences. A simple thermostat fulfills these conditions and is therefore considered a reflex agent in the hierarchy of intelligent agents. In nature, this includes all living beings whose behavior is exclusively determined by reflexes.
On the other hand, a learning agent operates freely in unfamiliar environments and learns new behaviors and its own goals. So far, there is no artificial variant of such a learning agent. This level would be comparable to human intelligence. In practice, complex systems such as robots can consist of several intelligent agents that share control of motor function, perception, planning, and orientation.
Machine Learning Takes Over Artificial Intelligence
In recent years, however, one has heard of intelligent agents less often, at most in the context of reinforcement learning (explanation). There is a simple reason for this: By the 1980s, at the latest, it was recognized that symbolic AI research, in which static programs performed logical operations with knowledge from databases, no longer made any progress.
Logical thinking is easy to describe in a programming language, but understanding the environment, seeing, or speech is too complex to master with the static solutions. “Good Old-Fashioned Artificial Intelligence” is dying out, and the term “learning” appears in more and more specialist books. It is the beginning of the triumphant advance of machine learning: Neural networks master many tasks that symbolic systems failed to achieve.
From 2010, the breakthrough with deep learning will follow: deep neural networks break one AI record after another. These neural networks can be part of intelligent agents. Still, especially in English-speaking countries, machine learning is used almost synonymously with AI research – also brought about by companies such as Google, Amazon, or Netflix.
The Meaning Of The Term Artificial Intelligence Is Changing
During this period, two worlds merge: AI research meets pop culture. In the public eye, the term artificial intelligence does not stand for a research field, but for an intelligent, artificial entity, influenced by science-fiction depictions such as HAL 9000 from “2001: A Space Odyssey”.
These intelligent entities act like humans, only without humanity. They are often even more intelligent than we are, self-determined, conscious, and have big plans, either the destruction or the salvation of humanity. Driven by the success of machine learning, companies, the media, public relations workers, and some researchers are soon adopting this object-related AI term.
So whoever writes, “But it has nothing to do with real artificial intelligence!” Today’s comment columns do not mean: “This is not a result of the research field Artificial Intelligence.” He probably means: “This system is not intelligent.” Unsurprisingly, the focus of this AI criticism is on the vague concept of intelligence itself. Even if the term “artificial” is just as unclear – but that is a different topic.
AI Is What Hasn’t Been Achieved
So we have to clarify what intelligence means, and the discussion of what artificial intelligence is and what is not would be over once and for all. Unfortunately, it’s not that easy. It is worth taking a look at the history of AI development, which shows that AI critics themselves often do not know exactly what they understand by intelligence.
Time and again, tasks that once required intelligence suddenly seem to be possible without it: character, image and speech recognition, playing chess, or credible writing texts were previously considered meaningful AI tests.
Common AI systems now meet these requirements, which, however, no longer or less count for many. This phenomenon is called the “AI effect” and was summed up by computer scientist Larry Tesler as follows: “AI is whatever has still not been achieved.”
The AI effect shows how difficult the term intelligence is to grasp. The search for a simple definition proves that: Sometimes intelligence is what we measure with intelligence tests. Sometimes it is a collective term for all possible cognitive abilities. Sometimes what behaves intelligently is intelligent. Sometimes, what functions intelligently inside.
Biological Intelligence Has Gradations
A look at cognitive research, psychology, or cognitive-behavioral research clarifies that there is no simple answer to the demanding intelligence question. Intelligence comes in gradations. So far, everyone seems to agree: In addition to people, we also call dolphins, dogs, birds, or bees more or less intelligent.
Conceptual distinctions such as narrow and general artificial intelligence try to grasp this differentiation of different forms of intelligence in language. Close AIs learn and master only one or a few tasks and cannot cope with changing environments. A general AI, on the other hand, should show a generalized learning ability and intelligence.
Depending on the idea, it is human or vastly superior to us, since humans as a product of evolution cannot find a solution for every problem or even recognize it at all. The adjectives “narrow” and “general” help in the AI context to express gradations. But that does not yet prove that such systems can rightly be called intelligent.
Intelligence, Change, And Learning
By the end of this article, I cannot offer a definitive solution to this linguistic problem. So far, I have hopefully understandably explained the reasons for this. But I would like to contribute my assessment: I consider the attribution of intelligence for current AI systems appropriate.
The fact that we describe intelligence in degrees shows that the term is either completely unclear or encompasses a wide range of skills with different levels of development. The use of time in the natural sciences confirms the latter assumption.
However, intelligence is not the same as these skills – on the contrary: intelligence is the capacity to learn new skills efficiently. This capacity comes in varying degrees. Some living things are more efficient at learning new skills than others.
So the more efficiently a system learns a new skill, the brighter it is. There are different ideas about what exactly constitutes such a learning process. A biological or artificial system that can autonomously produce such inferences and thus learn is then considered intelligent.
Intelligence Is A Measured Value
In this sense, intelligence would be a measured value that indicates how efficiently a system autonomously forms abstractions and thus learns. Current machine learning systems still require plenty of training material and human support for supervised learning (explanation).
Nevertheless, they learn and, depending on their size, form different forms of abstraction. In this sense, they are prescribed on the intelligence scale – albeit very low. Calling them “tight AI” makes sense because they have to learn everything from scratch outside their domain.
With the development of massive language models such as GPT-3, we may be experiencing a transition phase in which AI systems emerge that master several domains and are capable of so-called few-shot learning (explanation). These systems only need a few examples to learn new ones based on existing skills.
The term “Transitional AI” for such systems, as they could be a transition from closer to general or even human-like AI. A definitive statement on this will only be possible when OpenAI opens GPT-3 completely to other scientists for tests or a similarly powerful system from another manufacturer becomes open source.
Conclusion: AI Is What You Make Of It
Artificial intelligence can mean a research field, a human-like intelligence, or a system that learns how biological intelligence occurs with different capacities. All of these attributions are legitimate.
With the latter understanding of the term, a general or human-like artificial intelligence would not be the first AI to be a “real” intelligent, artificial entity. About its predecessor models, it would be what humans are to the first unicellulars.