Artificial Intelligence, Explained

The idea of an artificial mind predates computers. For centuries, people imagined beings that could speak, obey or escape control; those stories shaped the hopes and fears through which we now view machines.

In the *Iliad*, Hephaestus is served by golden attendants endowed with voice and intelligence. Greek myth also tells of Talos, the bronze giant who guards Crete; Jewish tradition preserves the Golem, a clay creature animated by a word. These images are far removed from modern technology, yet they already pose a familiar question: what happens when people build something that appears to act on its own?

Mary Shelley’s *Frankenstein* appeared in 1818. The creature is more than a threat: the novel asks what responsibility belongs to a creator who abandons their work. In 1920, Karel Čapek’s play *R.U.R.* staged artificial workers and popularised the word “robot”, derived from the Czech *robota*, or forced labour. In 1942, Isaac Asimov formulated the Three Laws of Robotics; his stories show how difficult it is to translate simple rules into safe behaviour in every situation.

HAL 9000 in *2001: A Space Odyssey* (1968) does not act out of malice: it comes into conflict with the crew while trying to complete its mission. In Kazuo Ishiguro’s *Klara and the Sun* (2021), the question shifts from threat to relationship: what does it mean to care for someone, and can a machine do so? Fiction does not predict the future, but it makes visible the moral questions technology brings with it.

In 1950, Alan Turing published “Computing Machinery and Intelligence” and proposed the imitation game, later known as the Turing test: can a machine appear intelligent in a conversation? In the summer of 1956, a workshop at Dartmouth College gave the emerging field its name: artificial intelligence. Its organisers proposed that aspects of intelligence could, in principle, be described precisely enough for a machine to simulate them.

Two approaches developed. Symbolic AI seeks to represent knowledge and reasoning through explicitly written rules. The connectionist approach uses artificial networks that learn from examples. Neither has disappeared, but the growth of data and computing power has made neural networks central to the latest wave.

Table 1. Selected milestones in AI history; this is not a comprehensive chronology.

The Transformer, described in 2017 by Google researchers, underpins much of today’s large language models. On 30 November 2022, OpenAI made ChatGPT available to the public, turning a conversational interface into an accessible gateway to generative models. Scientific recognition followed in 2024 with Nobel Prizes: Physics to John Hopfield and Geoffrey Hinton for contributions to neural networks, and Chemistry to Demis Hassabis and John Jumper for their work on AlphaFold.

When we say “AI”, we often mean very different systems. Language models generate text and code; multimodal models work across content types such as images, audio and video; agents can use tools and carry out sequences of actions. Other systems specialise in scientific, industrial or physical tasks, such as analysing protein structures, controlling robots or driving vehicles.

Even language models differ in training, access, cost, capability and use. Some are available only through a provider’s service; others make weights or components available for third parties to run and adapt. The labels “open” and “closed” cover different degrees of transparency and freedom: they are not a simple ranking of quality or safety.

Models can produce convincing but incorrect answers, lose details, reproduce biases in their data or respond differently to small changes in a prompt. Reasoning and tool-use features expand what they can do, but they do not make every answer verified. Human oversight remains important when decisions, people or sensitive data are involved.

Behind a chat interface lies physical infrastructure: chips for training and running models, data centres, networks, water and cooling systems. Large technology companies are investing at scale to build this capacity. Financial estimates change quickly and are not always comparable; rather than relying on one headline figure, it is useful to see the direction of travel: computing has become a strategic and costly resource.

The competition is geopolitical too. The United States and China lead much of the ecosystem, while export controls, semiconductor access, public investment and energy availability affect the ability to train models. The release of DeepSeek-R1 in January 2025 renewed debate about model cost and efficiency; on its own, it does not show that infrastructure has become irrelevant.

The question of an “AI bubble” remains open. Investment may produce sustainable products and revenue, but it may also exceed demand or concentrate in companies unable to recover their costs. The history of the internet is a reminder that a technology can transform society while many of the companies financing it disappear. Innovation and financial success are not the same thing.

AI agents are systems that can do more than generate responses: they may use tools, consult documents or take successive steps towards a goal. They are useful when a task is well defined and actions can be checked. Once they have access to data, accounts or external tools, permissions, limits and the ability to stop them become just as important.

The phrase *artificial general intelligence* (AGI) has no shared definition. It can refer to a system able to handle many cognitive tasks, but people use different criteria to measure it: performance, autonomy, adaptability or comparison with human work. Superintelligence is a further hypothesis, describing capabilities that would greatly exceed human ones across many fields. There is no reliable date for either milestone.

One line of research measures specific capabilities rather than assigning a date to AGI. METR, for example, estimated the duration of software tasks that some agents can complete autonomously at a defined success rate. Its 2025 report observed an approximate doubling every seven months over the period studied, while noting that results depend on task selection and methodology. This is a bounded experimental trend, not a guaranteed timetable for general autonomy.

One visible change is a move from software designed in advance for a particular use towards tools that can be described and adapted in natural language. *Vibe coding* — a term popularised by Andrej Karpathy in 2025 — describes creating software by asking a model to write or change code, then assessing the result. It can make prototypes and small tools accessible to people who do not code; maintenance, security and reliability still require expertise and checks.

Interfaces are changing too. On 7 October 2026, OpenAI announced Intelligent UI in ChatGPT: responses that can include visual and interactive elements. It is a concrete example of an interface generated in the moment to suit a request. The promise is fewer steps between a question and an action; the challenge is preserving transparency, control and accessibility.

Research on work suggests real opportunities, but results depend on task and context. Two often-cited studies illustrate different scales and measures:

Table 2. Results are not directly comparable: methods, samples and definitions of productivity differ.

These studies measure different effects. Time saved on a task does not, by itself, tell us whether the result is better, whether the benefit is fairly distributed or whether the whole economy will grow. Organisations need to track quality, errors, workload and learning as well as speed.

The most meaningful uses often address practical problems. Be My Eyes combines its volunteer network with a tool that describes images and helps blind and low-vision people interpret their surroundings. NotebookLM can turn uploaded materials into summaries and audio conversations. Voice cloning, when authorised and carefully managed, may help some people who have lost their voice; the same technology can enable fraud and impersonation. AlphaFold has made large-scale protein-structure predictions available to biological research. These are different cases, each with benefits and limits that need to be assessed in context.

The questions are not limited to distant futures. Generative systems can invent information, reproduce bias, facilitate deceptive content or expose information entered by users. Agents add the risk of unauthorised actions, especially when they have broad permissions or do not ask for confirmation before doing something difficult to reverse.

There are also questions about work and power: which tasks will change, who will receive the benefits and who will bear the cost of transition? The most capable models depend on capital, data, energy and expertise concentrated in a limited number of organisations. In *Atlas of AI*, Kate Crawford reminds us that AI is not immaterial: it relies on natural resources and the labour of many people. In their paper on “stochastic parrots”, Emily Bender, Timnit Gebru and co-authors warn against confusing fluent language with human understanding.

Regulation sets some responsibilities, but it does not settle every ethical or organisational question. In the European Union, the AI Act follows a phased timetable: some prohibitions and the duty to promote AI literacy have applied since February 2025; obligations for general-purpose AI models since August 2025; and the general rules since August 2026, with later deadlines for specific categories of high-risk systems. Each organisation needs to assess its role, the actual use and the provisions that apply.

Artificial intelligence has a long history, but the speed at which it has entered everyday activity is new. Its most tangible promise is not that every problem will disappear: it is that some capabilities — writing, searching, translating, prototyping and analysing — may become more accessible. The same ease can amplify errors, dependence and concentrated power.

Mary Shelley had already raised the question of a creator’s responsibility. Today the answer does not belong only to those who build models: it concerns institutions, businesses, professionals and the people who use them. To be an apocaloptimist, as Daniel Roher’s documentary title suggests, is to hold together the possibility of substantial benefits and the need to govern their risks. The choice is not between enthusiasm and fear; it is how willing we are to understand, verify and decide.