Artificial general intelligence (AGI) is the idea of a system that can learn and handle almost any intellectual task a person can, and move from one to another without being trained again. Today there is no single definition and no accepted test to show it has been reached.
A question from the app. Pick an answer.
What helps a text AI sound human?
Fluent language does not show that it thinks like a person. It learns patterns during training.
It can generate new sentences and also repeat text. It is not limited to copying and pasting.
It learns language patterns. It can sound human and still give a false answer.
Many AI systems are "narrow": good at one task and nothing else. A chess program beats any champion, but it cannot recommend you a song. Your email spam filter cannot translate a sentence. Each one was trained for its own task.
Language models blur the line: the same model writes, translates, summarizes and codes. So some researchers say they are more general than earlier AI, and others answer that they still fail at things that are easy for a person and do not learn from experience the way we do.
Each group defines it differently. For some it means matching a person at most intellectual tasks. For others, being able to do most of the jobs people do today. For others, learning something new from very few examples, like a child who sees a giraffe once and recognizes it from then on. Depending on the definition, the same system passes or fails.
Tests have the same problem. Chatting like a person, passing an exam or solving puzzles measure parts of intelligence, not all of it. And every time a system beats a test, people debate whether the test measured what they thought it did.
Whether it can be reached by improving current methods or needs something new. Whether it will happen and when: experts disagree, and estimates vary a lot. And what benefits and risks it would bring, and how to prepare. These are open questions, not facts.
If you read "an AI has reached general intelligence" or "AGI is just around the corner", ask yourself three things. Which definition does it use? What test was run and who checked it? Who benefits from the announcement? A system scoring well on a test shows it is good at that test, not at everything.
The reverse holds too: something not being AGI does not make it useless. Some AI already translates, recommends and catches fraud without being general. To see how it learns, start with machine learning.