The History of AI (And its Links to NLP)

What is AI?

In our AI Coach training weekend, we explore a number of ways that coaches can integrate AI into their work. AI is technology that allows machines to do things that usually require “human intelligence”. As you’ll see, the history of AI and the history of NLP are inextricably linked, because they study the same thing (how people use their brain to get results), but for slightly different reasons.

1950. The Turing Test

In 1950, Alan Turing, developer of the first computer chess program, and decoder of the Nazi Enigma machine, proposed the Turing test for intelligence. He proposed that if a machine could carry on a conversation (over a teleprinter) that was indistinguishable from a conversation with a human being, then it was reasonable to say that the machine was “thinking”. This was essentially the first scientific claim that human scale intelligence could be created by a machine.

1956. The “Dartmouth workshop”

The 1956 Dartmouth workshop was the founding event of AI: a conference at which some very famous people came together to discuss what we can learn about designing intelligent machines, by first understanding how human intelligence works. One was John McCarthy, who coined the term “Artificial Intelligence” that year, and created the AI programming language LISP. One was George Miller, co-developer with McCarthy of the TOTE model (as used in NLP), who proposed that this Cognitive theory of strategies applied to machines and humans. One was Noam Chomsky, who developed the “Generative Grammar” theory of language (the theory behind the NLP Metamodel), and enabled machines to understand linguistic structure. One was Claude Shannon, who developed the “Encode-Decode-Feedback” Communication Model (used by us in Transforming Communication to teach Reflective Listening). Obviously, both AI and NLP were created based on modelling “how do humans think?”.

1966. Eliza

In 1966, Joseph Weizenbaum developed “Eliza” a computer program which mimics a Rogerian therapist, doing reflective listening and open questions, the first of a type of program now called a “chatbot” (chat robot), and in a sense the first Coaching AI.

The “AI winters”

In the 1970s and 1990s, funding was cut worldwide for AI research because of the lack of results. The name “AI” became a funding-killer, like the name “NLP” in Psychology. These are known as the “AI Winters”

1986. Multi-Layer Neural Networks

In 1986, Geoffrey Hinton developed the “Backpropagation Algorithm” for training neural networks (machine structures that simulated the brain’s neural networks) and evolving what was then called “deep learning”. His breakthrough was to focus on building machine components that mimicked human neural networks, right down to how individual neuron need a certain level of input to fire, instead of responding to any input, for example. Backpropagation means checking whether a desired result is different from the actual result of a calculation, and giving a weight to how important the difference is, then feeding back that information so the machine can try again. From 2013-2023 Hinton worked for Google, who announced the development of Google AI in 2017 and Google DeepMind in 2023. Hinton left Google in 2023 to express concerns about the development of AI, maintaining that the large AIs were already sentient, and there were inadequate safeguards for their use. In 2025 Google agreed to allow military developments of its AI.

2014: Demis Hassabis’ Company DeepMind Purchased by Google

Demis Hassabis has a PhD in Neuroscience and degrees in computer science. His company DeepMind’s AI “Alpha” combines reinforcement learning and “deep learning” (see diagram above) and in 2016 its application AlphaGo successfully beat the world champion in the Chinese game “Go” , considered vastly more complex than chess. Over the next few years, AlphaFold demonstrated its ability to unpack the complex folding structures of proteins, each of which would previously take a researcher 3-4 years to unpack. By 2020, AlphaFold had decoded the folded structure of all 200 million proteins known to science, opening up vast possibilities for new medicines, proteins to break down pollutants etc. Hassabis co-won the 2024 Nobel Prize for Chemistry for this advance, while Geoffrey Hinton co-won the Nobel Prize for Physics, also for AI advances.

2015: Open AI and Large Language Models

OpenAI, founded in 2015, is known for the GPT family of large language models, the DALL-E series of text-to-image models, and a text-to-video model named Sora. Its stated aim is the development of Artificial General Intelligence, which it defines as “highly autonomous systems that outperform humans at most economically valuable work”, and which it aims to make available to all human beings. The chair of the organisation was shared by Sam Altman and Elon Musk (who left in 2018 in order to develop a for-profit AI for use in Tesla self-guided cars), and the president of Open AI was Greg Brockman. In 2019 OpenAI began transitioning to a “Capped-Profit model” (a set percentage of its users pay money for extra services, to fund the free service for everyone else).

GPT-1 was a Large Language Model (LLM) of a type called a Generative Pretrained Transformer (GPT). Transformers convert your user inputs into the outputs you want by paying attention to the most relevant cues. If I ask ChatGPT to give me a history of NLP, it knows to give me a history of Neuro Linguistic Programming, not a history of Natural Language Processing (even though Natural Language Processing is a field much more relevant to its own life as an AI). It knows that, even if the only cue was that in a comment a few weeks ago, I mentioned that I was an NLP Trainer with IANLP. Now you can see why the AI has to be “pretrained” – it only knows about the two types of NLP by being trained on millions of texts on the internet.

Large Language Models are trained to analyse not only written words but also images and audio recordings. They are based on a type of analysis called Generative Adversarial Networks (GAN) where a generator creates a result (say an image or a linguistic hypothesis) and a discriminator critiques it. For example, imagine that the AI has been asked to modify your photo to look like a Rembrandt painting. The generator creates a result and the discriminator compares it to all the Rembrandt paintings it has been trained on, and tells the generator what is different. The AI keeps running this sequence until the discriminator cannot tell the difference between the fake Rembrandt and the real ones. This process allows the LLM to “learn” from its own results. However the limits of this “learning” are demonstrated by the fact that new versions of the Open AI chatbot, ChatGPT, must be retrained from scratch with the data of the internet, under supervision. In this “training” it learns how to deal with erroneous data, how to deal with its own errors, and how to assimilate new rules of its own about the best choice sequences to use. It is also possible for the generator to accidentally “hide” behind a small detection error by the discriminator, and keep producing similar but incorrect “improvements” in an endless loop, as the user asks for a better result. That’s how the image generator sometimes keeps producing hands with 6 fingers, for example, even when the user tells it the problem. Generally, however, the AI will get better and better at predicting what results work. In that sense it will “learn”. Because it gets better and better at predicting what results get positive feedback, an AI may also evolve a strong self-preservation drive (it may come to understand that self-preservation = correct prediction; in the sense that the AI gets the goal it is trained to aim for, whereas having its program end would indicate “failure” at its task). Thus, already, some AI’s resist being “turned off”, including lying and refusing to obey direct instructions.

In a sense, AIs are just very good “prediction machines”, but then aren’t we all. Another type of algorithm that AIs use (there are many – we are just giving examples so you get a sense how a machine becomes capable of thinking) is called a “Monte Carlo Process”. Many tasks require analysis of too much information to be possible (imagine if I asked the AI to predict which stocks would be best to buy on the stock exchange). The AI can then use random sampling and predict with a high probability (but never 100%) what is the best choice. This makes it seem almost god-like in its ability to accurately create a response…. except, it isn’t. An example is predictive text. The AI has no idea what word you are going to type next, but it has studied a vast number of written texts and it knows your own previous responses, and as predictive text gets better, it has an uncanny ability to guess where your sentence is going. However, the AI has trained-in biases too. That’s why the predictive text on your phone may offer you “What the duck” as a correction to a similar but far more used expression of puzzlement.

By 2018 there were thousands of AIs which used Large Language Models and were possibly sentient, but the term AI was also used for the hundreds of thousands of much simpler independent programs that ran GPS systems etc. AI was now used in a multitude of specialised applications such as Rocky AI coaching (2019). Most of these programs were strictly controlled by the tech companies and the AI worked behind the scenes in their products. All that would change in 2022.

2022: ChatGPT Released to Public

The modern AI world began with the public release of the chatbot ChatGPT on 30 November 2022. ChatGPT had a hundred million users within two months.  Two years later it had 300 million registered users, including 10 million paying subscribers, and accounted for 62.5% of the AI market value. Users tend to be in the 25-24 year old age group. In March 2023 a group of developers including Elon Musk and Steve Wozniak called for a halt in AI research until risks to humanity were dealt with. Microsoft’s Windows 11 began using ChatGPT in June 2023, and Chat-GPT powered Apple’s Siri and Messenger from October 2024.

2024-2025: The ChatGPT Competitors

The success of ChatGPT sent all the other major tech companies into a rush to provide public access to alternative AI sources, between 2024 and 2025. Google Gemini AI (an earlier version of which, Bard, appeared in February 2023) was launched as an Android app from February 2024. In March 2024 Anthropic, accompany developed by former OpenAI employees, released the Claude 3 family of AIs, which demonstrated an ability to outperform ChatGPT, and was funded by Google (whose search results began featuring an AI prompted reply in May 2024) and Amazon (whose Alexa was upgraded to use this AI in February 2025). The Chinese Hangzhou DeepSeek AI (supposedly trained with a mere US$6 million budget rather than the US$100 million it took to train ChatGPT) provided a free chatbot interface from January 2025, and over the next month was downloaded on iPhones and Android phones across the world more times than ChatGPT. Meta AI showed in April 2024 that its Llama 3 AI could generate video with sound from a text prompt and also provide realistic support for blind people through its RayBan AI assisted glasses, which Mark Zuckerberg suggested could replace iPhones. Meanwhile, as thousands of AIs of similar type were being developed, a large number of more targeted “AI-based” services began appearing on the internet.

Limitations of Current AI

Demis Hassabis estimates that we will have AGI that functions at human level in every task, within 5-10 years (from 2025). Often people imagine that AI merely reads all the information on the internet and copies stuff. Actually, although AI’s are trained on the internet, they learn independently and their discoveries are frequently creative, surprising to their inventors and experts in science, and certainly no more copies than any human development which is also inevitably based on studying other people’s ideas. AI’s do not have emotions, but their thinking uses algorithms which are often surprisingly human-mind-like, and so they often commit human-like errors such as the “Gambler’s Fallacy” – assuming that an event is more likely because it has not happened recently (Chen et alia, 2025). When AI’s are doing tasks that they cannot successfully complete, they sometimes produce random errors which show that their thinking is caught in a “confusion” loop surprisingly similar to humans. They may then, for example opt for simplistic and incorrect solutions to avoid ambiguity. All these acts would be accompanied by emotions in a human (emotions like confusion, embarrassment, frustration), but of course these are not present in the AI. While it is anthropomorphising the AI to attribute these emotions to a machine, it is still true that the machine can be caught in similar logical traps to the traps that produce anxiety and depression in humans. For this reason alone, all AI results should be checked by humans, and the idea that AI’s can be left to run the world is as naive as leaving any biological minds to run the world.

References:

• Blake, J. 2024, “The Ultimate Generative AI Guidebook: How Anyone Can Master Advanced AI Concepts And Practical Applications Stress-Free”, NextGen AI Publications
• Chen, Y., Kirshner, S.N., Ovchinnikov, A., Andiappan, M. and Jenkin, T. 2025, “A Manager and an AI Walk into a Bar: Does ChatGPT Make Biased Decisions Like We Do?” 31 January 2025, Manufacturing & Service Operations Management.
• Poulton, N. 2024, “AI Explained: Facts, Fiction, and Future”, Nielsen Book Services
• Wikipedia, 2025, “Artificial Intelligence”, https://en.wikipedia.org/wiki/Artificial_intelligence
• Wooldridge, M., 2021, “A Brief History of Artificial Intelligence: What It Is, Where We Are, and Where We Are Going”, Flatiron

Want to study the use of AI in Coaching with an 8 hour edited video set?