Artificial General Intelligence
Artificial General Intelligence (AGI) refers to a computer system that could solve any intellectual task a human can handle – not just individual specialized tasks. Whether such a thing is possible, and when, remains open; the term is therefore above all a goal and a point of contention, not an existing product.
Today’s computer programs that learn from examples are specialists. One recognizes tumors in X-ray images, another translates texts, a third drives a car. None of them can take over another’s task. Artificial General Intelligence, or AGI for short, means instead a system that could master any intellectual task – as flexibly as a human. Such a system would have to learn new things without experts having to rebuild it specifically for that purpose. AGI does not yet exist. The term describes a goal that some researchers consider close and others believe is decades away.
Why billions are being bet on a single word
AGI is the reason companies like OpenAI, Google DeepMind, or Anthropic were founded in the first place. Their founding documents state the goal explicitly. Investors are pouring billions into these companies even though their current products bring in only a fraction of that. The bet is: whoever achieves AGI first controls a technology that can replace or complement human labor in almost every field.
That’s why the word is also an economic factor. If a CEO says AGI is imminent, stock prices and headlines react. Critics consider this marketing, since no one provides a binding definition. Without a clear benchmark, it’s hard to disprove claims of getting closer to the goal.
Nevertheless, the debate deserves to be taken seriously. Governments are now writing laws that refer to particularly capable systems. And part of the research focuses on how to prevent a highly capable system from pursuing goals that no one intended. This field is called AI safety.
How AGI was supposed to be recognized – and why that fails
The oldest proposal is the Turing test from 1950. A human chats with an unknown counterpart and is supposed to guess whether a machine is answering. Today’s language models often pass such conversations. Nevertheless, hardly anyone considers them AGI. The test measures convincing speech, not understanding.
Other definitions rely on tasks. OpenAI, for instance, describes AGI as a system that performs most economically valuable work better than humans. Researcher François Chollet turns it around: intelligence is the ability to learn something completely new from just a few examples. His test set ARC-AGI contains colorful puzzle grids that humans solve easily and machines long failed at.
Technically, there is no blueprint for AGI. The path currently being pursued is called scaling: more training data, more computing chips, larger models. This has greatly improved capabilities. Whether general intelligence will eventually emerge from this, or whether a fundamentally different idea is needed, nobody knows. This is exactly what the field is arguing about.
AGI in headlines and contracts
In the news, the term mostly appears in quotes from company executives. Predictions range from a few years to several decades. A simple follow-up question is useful: what definition is the person basing this on? Without that information, the figure is hardly assessable.
AGI even appears in contracts. The partnership between OpenAI and Microsoft included a clause stating that certain rights end once AGI is achieved. A fuzzy term thus becomes a question involving billions of dollars – and a case for lawyers.
In everyday life, you don’t yet encounter AGI, but narrow AI: chatbots, translators, image generators, recommendation systems. They appear broadly applicable but remain bound to their training data. When an ad claims a product is “almost AGI already,” that’s sales talk. After all, there is still no reliable way to recognize genuine AGI.