Artificial intelligence began as a philosophical question posed by Alan Turing in 1950 and became a formal research field at the 1956 Dartmouth workshop. Early optimism about symbolic reasoning gave way to cycles of hype and disappointment, known as “AI winters”, before expert systems, statistical machine learning, and deep learning each pushed the field forward.
The 2017 Transformer architecture enabled large language models, and the public release of ChatGPT in late 2022 brought generative AI to hundreds of millions of people. Today’s focus on AI agents that can plan and act raises fresh questions about reliability, jobs, and governance that echo debates as old as the field itself.
AI can feel like a sudden arrival, as if chatbots appeared out of nowhere a few years ago. In reality, today’s tools rest on more than seventy years of ideas, failures, rediscoveries, and steady engineering progress.
I have previously traced the evolution of computing and of the internet. This article completes the trilogy by following artificial intelligence from a thought experiment to the agents now booking meetings and writing code.
Turing’s Question (1950)
In his 1950 paper “Computing Machinery and Intelligence”, British mathematician Alan Turing asked, “Can machines think?” Rather than define thinking, he proposed the “imitation game”, now called the Turing Test: if a machine could converse indistinguishably from a human, should we not call it intelligent?
The question shaped decades of debate. It shifted attention from what intelligence is to what intelligent behaviour looks like.
Dartmouth and the Birth of a Field (1956)
The term “artificial intelligence” is credited to John McCarthy, who organised a summer workshop at Dartmouth College in 1956 with Marvin Minsky, Claude Shannon, and Nathaniel Rochester. The proposal optimistically suggested that significant progress could be made over a single summer.
The following years produced impressive early programs:
- Logic Theorist, which proved mathematical theorems
- Programs that played checkers and learned from experience
- ELIZA (mid-1960s), a simple chatbot that surprised users with how human it seemed
Symbolic AI and the First Winter
Early AI relied on symbolic reasoning: humans encoded knowledge as rules, and computers manipulated those symbols. It worked well for puzzles and narrow problems but struggled with the messiness of the real world, such as language, vision, and common sense.
By the 1970s, promises had outrun results. Critical government reports and reduced funding in the US and UK led to the first “AI winter”, a period of scepticism and shrinking budgets.
Expert Systems and the Second Winter
The 1980s brought a revival through expert systems, programs that captured specialist knowledge in thousands of if-then rules. Businesses used them for tasks such as configuring computer systems and assisting diagnosis.
But expert systems were costly to build and maintain, and brittle when faced with situations outside their rules. As specialised AI hardware markets collapsed in the late 1980s, a second winter followed.
Machine Learning Takes Over
From the 1990s, researchers increasingly let computers learn patterns from data instead of hand-coding rules. Statistical methods powered spam filters, recommendation engines, and speech recognition.
Milestones from this era include:
- 1997: IBM’s Deep Blue defeated world chess champion Garry Kasparov
- 2011: IBM Watson won the quiz show Jeopardy!
- Growing use of machine learning in search engines and online advertising
Cheap computing power and the explosion of internet data made this approach increasingly practical.
The Deep Learning Revolution
Neural networks, inspired loosely by the brain, had existed since the 1950s but were long considered impractical. Around 2012, a deep neural network known as AlexNet dramatically improved image recognition results, helped by graphics processors (GPUs) and large labelled datasets.
Deep learning quickly spread to translation, voice assistants, and medical imaging. In 2016, DeepMind’s AlphaGo defeated top professional Lee Sedol at Go, a game long thought too intuitive for machines.
Transformers, Large Language Models, and ChatGPT
In 2017, Google researchers published “Attention Is All You Need”, introducing the Transformer architecture. It allowed models to process language efficiently at enormous scale, leading to large language models (LLMs) trained on vast amounts of text.
When OpenAI released ChatGPT in November 2022, generative AI went mainstream almost overnight. Image generators, coding assistants, and open-weight models that run on ordinary laptops followed in rapid succession.
The Age of AI Agents
The current frontier is agents: AI systems that do not just answer questions but plan multi-step tasks, use tools, browse the web, and take actions on a user’s behalf. Workflow tools like n8n, coding agents, and personal assistants all reflect this shift.
Agents also surface old concerns in new forms:
- Reliability: models can still produce confident errors
- Accountability: who is responsible when an agent acts wrongly?
- Work: which tasks will be automated, and which will be augmented?
- Regulation: governments, including the EU with its AI Act, are setting rules
Conclusion: Seventy Years of Overnight Success
The history of AI is a story of bold predictions, humbling winters, and quiet breakthroughs that eventually compounded. Turing’s question has not been fully answered, but machines now converse, create, and act in ways that would have astonished the Dartmouth pioneers.
Knowing this history is a useful antidote to both hype and panic. Progress has rarely been as fast as optimists promised or as stalled as sceptics claimed, and the next chapter will likely follow the same pattern.
