AI has already had a long and spasmodic history of development so far.
Inspired by Alan Turing’s speculative essay of 1950, “Can machines think?”, a group of researchers in 1951 devised the first working AI program which they ran on the Ferranti Mark 1 at the University of Manchester, England. Then in 1956 a group at Carnegie Mellon in Pittsburg wrote the first program, called “Logical Theorists”, engineered to perform automated reasoning.
Then in 1964 development of the first expert system commenced, resulting in 1969 in DENDRAL to assist organic chemists in planning complex synthesis – however it relied on input of human-originated heuristics derived from bench experience. This was followed in 1971 by MYCIN, a medical diagnosis and treatment recommending program written in a new specialized language LISP. It was very successful when operated by an experienced doctor able to order extra lab tests to clarify possible ambiguities.
Building on this valuable success many other expert programs were developed in the 1970s and 1980s – all based around a module known as a knowledge base/inference engine. However the first wave of expert systems came to an end by the end of the mid-1980s. These had depended on human expert knowledge and skills which were often difficult to transfer into a computer program in a form that was easy to access and able to be expressed as useable heuristics. They could also rapidly become very complex as the made use of both forward chaining – where a rule could activate an appropriate algorithm to move forward – and backward chaining – where an emerging hypothesis is checked for compatibility against the original evidence.
Thus expert systems came to be seen as expensive and slow to design – yet they live on in applications where: (a) transparency and the possibility of review are important, e.g. credit scoring, job application assessment; or (b) where the for speed of operation is essential, e.g. within self-driving car applications.
Parallel to the development of expert systems were two other lines of experimentation often initiated by persons from outside engineering. As far back as 1957 a researcher called Frank Rosenblatt, who had a psychology background, made an early attempt to program a computer to learn by trial and error, called the “Perceptron”. It was a first attempt at using a neural network called “connectionism”. With some image capability it was able to distinguish between dogs and cats – but it failed to impress its financial backers. The second line of development was an early attempt at machine translation, but this failed badly and was discontinued in 1966.
These three disappointments, a lack of well defined data, limitations in the computer architecture at the time and inadequate processing power led to a period of disillusion and a lack of funding which was known as the “First AI Winter” (1974 – 1980).
There was a tentative restart with “explanation based learning” which relied heavily on input from a human expert who “explains” how one example can be validly used to produce a general rule. This approach produced a faster program, especially where data was imprecise or poor. However it remained only a good as the input from the human expert. Nevertheless it could be very focused and efficient, with an audit trail back to a human. This made for an expert system that was fast and reliable enough to be deployed in the field of legal argumentation.
In 1983 an effective type of recurrent neural network appeared where recurrent activity of one “neuron” causing the “firing” of the second one results in an increase in the sensitivity of the process leading to patterns of activity being established analogous to learning.
This improvement to the expert system concept resulted in earlier systems becoming obsolete and 1987 saw the final collapse of the earlier LISP-based expert systems and the start of the “Second AI Winter” (1987 – 2000).
Development did of course move along during this so-called “winter”. Of note were Support Vector Machines designed to improve data classification and regression analysis, of which the dog/cat sorting is a trivial example. In 1995 these were enhanced by the introduction of “Random Forests” – ensembles of decision trees based on random subsets of the presented data from which a corresponding number of predictions were made, which were then averaged or a majority vote taken on the classification to get the final result. Shortly after that, in 1996, the oddly named Long Short Term Memory (LSTM) units were incorporated into recurrent neural networks which, if unmodified, had a tendency to lose data that might be helpful later on in the learning process. The LSTMs took decisions on retaining or forgetting earlier assessment of data and if retained, doing so over hundreds of time intervals. So some data was indeed discarded but other was retained over a longer timescale that was represented in conventional “short term memory”. The practical effect was greater sensitivity to processing results which the LSTMs judged likely to be useful later on, at the expense of other results judged expendable. However this was expressed in terms of improved speed and accuracy.
During the next 13 years development concentrated on market testing robotic products like Roomba, a domestic robot vacuum cleaner and robot grass cutters while image processing was improved to sort and standardize machine readable images from large databases and for use in self-driving vehicles.
2009 saw the first LSTM recurrent neural network with pattern recognition software and this enabled cursive hand writing to be read and Google built an “autonomous car”. At the same time the ImageNet visual dataset containing 14 million hand annotated images was produced using a team of 49,000 individuals from 167 countries working from a base of 162 million candidate images!
The next decade was one of intense innovation and progress. Google greatly improved natural language processing in 2013 paving the way for the “chatbot”. Generative AI moved machine capabilities beyond algorithms and predictions to “prompts” and new content generation, in which natural language capability produces human-like text and responses. By now the rather unattractive and obviously AI images (which no doubt reflect the artistic preferences of some Californian geeks) generated by OpenAI’s DALL-E 3 are all too familiar!
In 2017 OpenAI published its research on generative models and Google’s DeepMind team based in London, UK simplified and accelerated its demonstration game playing system, “AlphaGo” to the extent that AlphaGo Zero was able to teach itself how to play Go – by playing against itself – did the same with chess and excelled in subsequent contests with other systems and with humans.
Shortly after this, in 2018, Google Duplex enabled an AI Assistant program to book appointments over the phone in Los Angeles. The following year Open AI’s GPT-4 greatly acclaimed language model was launched, but the “hallucination problems” that originated with GPT-3 were said to have persisted.
In November 2022 Open AI launched ChatGPT (also based on GPT-3) but was criticized because of the persistence of the “hallucination problem”, prompting political and public realm discussions. At the same time a tsunami of court cases erupted against several late entrants into AI with a number of “me too” products where the actions centred on copyright infringement of proprietary databases or unauthorized use of private and personal data or images for training their AI engines.
Finally Joe Biden issued a Presidential Order late in 2024 setting out 8 objectives for the ethical development of AI systems in the US. These included protection of US national interests, respect for copyright issues, protection of personal data of the public and general requirements for the products and applications to be factual, unbiased and non-discriminatory.
However, one of the first acts of Donald Trump on his first day of office was to sign his own Presidential Order rescinding that of Joe Biden and thus allowing US companies to develop AI free from any responsibilities or restrictions. This is likely increase the already established pattern of internecine litigation between the AI giants but stifle any complaints from the public.
On February 10 and 11, 2025, France hosted the Artificial Intelligence Action Summit. 61 countries, including China, India, Japan, France and Canada, signed a declaration on “inclusive and sustainable” AI. Sadly, both the UK and US refused to sign. Anglo-Saxon exceptionalism?