In the 1950s, there were widely-held beliefs that "Artificial Intelligence" will quickly become both self-conscious and smart-enough to win chess with humans. Various people suggested time frames of e.g. 10 years (see Olazaran's "Official History of the Perceptron Controversy", or let say 2001: Space Odyssey).

When did it become clear that devising programs that master games like chess resulted in software designs that only applied to games like the ones for which they were programmed? Who was the first person to recognize the distinction between human-like general intelligence and domain-specific intelligence?


3 Answers 3


Many publications from the middle of the twentieth century prove the questioner's statement that it was a widely held belief during that period that AI would quickly become conscious, self-aware, and smart.

Great Success

Many tasks and forms of expertise once the exclusive domain of human intelligence, after the development of the Von Neumann general purpose computing architecture became, by the end of that century, more or less the exclusive domain of computers. These are only a few examples.

  • Scientific and statistical computation
  • Drafting and manufacturing process automation (CAD and CAM)
  • Publishing and typesetting
  • Certain forms of algebraic and calculus reductions (Maxima and its derivatives)
  • Circuit analysis
  • Masterful board game playing
  • Profitable stock speculation
  • Pattern recognition (OCR, fingerprint, voice recognition, sorting, terrain)
  • Programming in predicate logic and recursive predicates
  • Strategy evaluation

Disappointments (thus far)

In contrast to this impressive array of successes, there is an equally long list of failed expectations.

  • Consumer available bipedal robots
  • Automated vacuum cleaning (major disappointment for this answer's author)
  • Autonomous mechanical factory workers
  • Automated mathematicians (creative hypothesis generation and proof/disproof to extend theory)
  • Natural language comprehension
  • Obedience to arbitrary commands
  • Human-like expression in conversation
  • Automated technical innovation
  • Computer morality
  • Human (or at least mammalian) emotional states
  • Asimov's three laws operating system
  • Adaptive strategy development in arbitrary and shifting set of domains

Domain and Domain-free Distinction

When did it become clear that devising programs that master games like chess resulted in software designs that only applied to games like the ones for which they were programmed?

Although the general public may have thought that a cybernetic chess master would also be smarter than people in other ways, those creating those programs were well aware of the distinction between developing software that exhibited excellence in chess play hard coded and developing software that exhibits the ability to learn chess play and develop excellence iteratively from novice.

The end goal had always been high powered general intelligence. More short term achievable objectives were created to facilitate the demonstration of progress to investors. It was the only way to maintain a continuous stream of research funding from the military.

The first milestone was to master a single game without machine learning. Then research turned to the building of domain knowledge so that a class of solutions, adaptations, and forms of planning could be realized in real time during warfare. As economic domination became more preferable to military domination during the third quarter of the twentieth century, the vision for AI scaled to embrace the domains of economics and natural resource management.

Consider this spectrum of automation maturity.

  • A program that enumerates current move sequence possibilities at each turn in the play of a chess game, eliminating probable bad moves at each projected move point, and selects the next move most likely to lead to a win
  • A program that does the above but also skews probability based on pattern recognition of known winning chess strategies
  • A program that is designed to be a run time optimized rules engine that centralizes and abstracts the redundant operations of the play of an arbitrary game and isolates and aggregates the representation of chess rules, chess strategies, and chess patterns and anti-patterns
  • A program that, given a set of rules of a game, can generate a next move based on any game state, remembers success and failure results and the sequences that led to those outcomes, and has the ability to assess the probable loss or gain of individual moves and the game patterns in space and time around them based on history, and then leverages these abilities to learn an arbitrary game, reaching the masterful level of play of chess through learning
  • A program that learns how to learn games such that, after learning several games, it can learn chess faster than an intellectually gifted human can

The first is easy. The last is extremely challenging.

When the distinctions between these phases of automation maturity became apparent and how clear people became of those distinctions in which research groups is a complex probabilistic function.

Key Contributors

Who was the first person to recognize the distinction between human-like general intelligence and domain specific intelligence?

Norbert Wiener was likely the first to deeply comprehend the distinction between electronic control of relays (investigated theoretically by Claude Shannon) and closed loop control. In his book, Cybernetics, a primarily mathematical work, he precisely established the foundation for self-correcting and adaptive systems. John von Neumann had a comprehension of the distinction between programming good game play and the human ability to learn good game play and published much on the topic.

It was Arthur Lee Samuel who actually wrote the first impressive demonstration of the distinction between game playing software and machine learning. It was he who bridged Wiener's work with the contemporary digital computer and first coined the term Machine Learning.

Distorted Restatements of Authentic Research and Innovation

The categories artificial narrow intelligence (ANI), artificial general intelligence (AGI), and artificial super intelligence (ASI), proposed in The AI Revolution: The Road to Superintelligence by blogger Tim Urban (Huffington Post, THE BLOG, posted 2/10/2015, updated 4/12/2015), is referenced in AI Stack Exchange in multiple places, but the distinctions between these categories are not precisely defined and the ideas contained therein are neither peer review nor validated by other research or statistics.

The work is no less conjecture than mediocre science fiction — entertaining enough to gain some popularity but not rational conclusions drawn from either repeatable experiments or randomized studies. The trend graphs provided in the article are of invented shape, not graphical representations of actual data.

Some of the material may later be found to have some truth in it, as in the case for many lay interpretations of scientific research or the futuristic thoughts of science fiction authors. However, much of the material leads to misconception and false assertions.


I expect a very precise answer to this question may be lost to the sands of time, although I hope somebody can given such an answer. In the meantime, here's one clue on the trail... This anthology of papers from 2007 starts with the following blurb:

Our goal in creating this edited volume has been to fill an apparent gap in the scientific literature, by providing a coherent presentation of a body of contemporary research that, in spite of its integral importance, has hitherto kept a very low profile within the scientific and intellectual community. This body of work has not been given a name before; in this book we christen it “Artificial General Intelligence” (AGI). What distinguishes AGI work from run-of-the-mill “artificial intelligence” research is that it is explicitly focused on engineering general intelligence in the short term.

But even if this is the origin of the specific phrase "Artificial General Intelligence", I am pretty sure people were making the distinction between "general intelligence" and "task specific" techniques much earlier.

The Wikipedia article on AGI also has a clue, where it states:

However, in the early 1970s, it became obvious that researchers had grossly underestimated the difficulty of the project. The agencies that funded AI became skeptical of strong AI and put researchers under increasing pressure to produce useful technology, or "applied AI".

That section cites this this book as support for that statement. And indeed, it contains the following verbiage:

Although most founders of the AI field continued to pursue basic questions of human and machine intelligence, some of their students and other second-generation researchers began to seek ways to use AI methods and approaches to tackle real-world problems. Their initiatives were important, not only in their own right, but also because they were indicative of a gradual but significant change in the funding environment toward more applied realms of research. The development of expert systems, such as DENDRAL at SAIL, provides but one example of this trend.

Given that DENDRAL began around 1965, it appears that some significant body of researchers (or at least funders) became strongly aware of the distinction between research into "general intelligence" and "applied AI" somewhere around the end of the 1960's. If you keep reading, other passages support the notion that DARPA in particular started pushing a more "applied" approach to AI research throughout the 1970's.

So, not a definite answer, but it looks like we can say that the distinction was known and taken into account at least by 1970, although use of the exact term "artificial general intelligence" appears to be of more recent coinage.


In 1973, the British government hired Sir James Lighthill to commission a "general survey" on the state of artificial intelligence. His report was a condemnation of current AI research, leading a wave of pessimism among AI scientists and the First AI Winter. You may view Lighthill's report (and contemporary criticism of his report) here, but I will summarize Lighthill's key points.

Sir James Lighthill divided AI into three categories:

  1. Advanced Automation - task-specific work
  2. Computer-based CNS research - research into the the "central nervous system" of humans
  3. The Bridge between Advanced Automation and Computer-based CNS research. This bridge would generally be seen as "general-purpose" robotics, so Lighthill would also use the term Building Robots.

Advanced Automation (or "applied AI") is obviously useful. Computer-based CNS research is useful because we want to know more about human intelligence. Both fields of AI had some successes, but its practitioners were overly optimistic, leading to disappointment in those fields. Sir James Lighthill was still very supportive of research in these two fields though.

Building Robots, on the other hand? Sir James Lighthill was very hostile to the very idea, probably because it was more overly hyped up than the other two categories and produced the least amount of valuable output.

He mentioned chess in particular as an example where "robotic" research has failed. At the time the report was published, the chess-playing engines were at the level of "experienced amateur standard characteristic of county club players in England". However, these chess-playing engines relied on heuristics that were made by human beings. The engines weren't intelligent at all...they merely were following the heuristics that were created by intelligent humans. The only advantage the robots bring to the table is "speed, reliability and biddability", and even that wasn't enough to beat the chess grandmasters.

Now, today, we would probably not treat chess as an example of general-purpose problem solving. We would more accurately classify it as "advanced automation", a "narrow AI" problem divorced from broader real-world implications of general problem-solving. But Sir James Lighthill probably would agree with us. He never used the term "narrow AI" and "AGI" (neither of those terms existed yet) but he would write:

To sum up, this evidence and all the rest studied by the present author on AI work within category B during the past twenty-five years is to some extent encouraging about programs written to perform in highly specialised problem domains, when the programming takes very full account of the results of human experience and human intelligence within the relevant domain, but is wholly discouraging about general-purpose programs seeking to mimic the problem-solving aspects of human CNS activity over a rather wide field. Such a general- purpose program, the coveted long-term goal of AI activity, seems as remote as ever.

Sir James Lighthill believed that the only thing that connects Advanced Automation and Computer-based CNS research is the existence of the Building Roobts "bridge" category. But he's very pessimistic about this category actually producing anything worthwhile. So instead, the AI field should instead breakup into its own its constituent parts (automation and research). Any robots that are built could then be specialized within their subfield...either industrial automation or CNS research. Trying to build the holy grail of "general-purpose program" would be worthless...for the time being, at least.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .