What is AGI? AGI stands for artificial general intelligence: an AI that doesn't just master one task but handles a broad range of mental tasks at least as well as humans. It is often equated with strong AI. Its counterpart is weak AI, also called narrow AI, which solves specific tasks without truly understanding or thinking like a person: playing chess, spotting spam, translating text. There is no single definition of AGI, and which level today's models such as ChatGPT or Claude reach depends on which definition you follow.
I have worked with Claude every day since February 2025, and I now run my business with it. My fourteen AI employees run on Claude Team, some also on open models or through the API. This post is only about the terms strong AI and weak AI and about AGI. What AI is in the first place is covered in What Is AI? Artificial Intelligence Explained, and how software learns from examples in Machine Learning Explained Simply. All AI basics are on my AI Basics page.
Where the terms strong and weak AI come from
The terms come from philosophy. In 1980 the philosopher John Searle published the paper "Minds, Brains and Programs", in which he describes his well-known Chinese Room thought experiment. The Stanford Encyclopedia of Philosophy sums up his distinction like this:
- Strong AI is the view that a suitably programmed computer really understands language and has mental capabilities similar to the humans whose behaviour it mimics.
- Weak AI is the much more modest claim that computers are useful because they can simulate mental abilities. It explicitly makes no claim that computers understand or are intelligent.
Searle wanted to show that a program doesn't produce understanding just by processing symbols. Whether he is right is still debated in philosophy today. What matters for you: in this original sense, strong AI is about understanding and consciousness, not about performance.
What weak AI is
In everyday use today, weak AI is mainly a question of breadth. IBM describes weak AI, also known as narrow AI, as AI that focuses on one specific task, such as answering questions or playing chess. Typical examples:
- the spam filter in your inbox,
- the text recognition that reads figures from a scanned receipt,
- a chess engine like Stockfish,
- AlphaFold, which predicts the three-dimensional structure of proteins.
Narrow doesn't mean weak in the sense of bad. Researchers at Google DeepMind place Stockfish and AlphaFold in their framework as "superhuman narrow AI": at their one task they beat every human. Outside that task they can't do anything.
What is AGI, and what does strong AI mean?
IBM equates strong AI with AGI and describes it as a hypothetical form of AI that, if it could ever be developed, would have intelligence and self-awareness equal to those of humans. That mixes both meanings: the breadth of abilities and the consciousness from Searle's distinction.
Most of the nine AGI definitions collected by the Google DeepMind research group (more on that in the next section) leave consciousness out and only ask what a system can do. One example is OpenAI's charter, which defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". That is a completely different yardstick from Searle's: he asks about what goes on inside, OpenAI about the work a system gets done.
Why there is no single definition of AGI
In a 2023 paper, the research group around Meredith Ringel Morris and Shane Legg at Google DeepMind put nine well-known AGI definitions side by side, including OpenAI's. Their finding: the definitions start from very different places. Some measure against the Turing test, some against the human brain, some against consciousness, others against economically valuable work. So the group proposes levels instead of a fixed line. A system is placed along two axes:
The performance levels run from "emerging", equal to or somewhat better than an unskilled human, through "competent", at least as good as half of skilled adults, up to "superhuman", better than all humans.
The term is disputed within the industry too. Dario Amodei, the CEO of Anthropic, writes in his October 2024 essay "Machines of Loving Grace" that he dislikes the term AGI. He finds it imprecise and loaded with sci-fi baggage and hype, so he prefers to speak of "powerful AI".
Even the law gets by without the term. The EU AI Act doesn't speak of AGI but of a "general-purpose AI model": a model that displays significant generality and is capable of competently performing a wide range of distinct tasks. Obligations for the providers of such models depend on it. I am a developer, not a lawyer, and this is not legal advice, just the operator's perspective.
What today's AI models can and can't do
Under Searle's original distinction, there is no way to measure whether a model like Claude "really understands". Anthropic itself writes about its models that it doesn't understand how they do most of the things they do. How a language model computes its answers is explained in How Do LLMs Work? A Plain-English Guide.
Performance can be measured. The DeepMind group placed the large language models as of September 2023, such as ChatGPT, Bard and Llama 2, in its framework like this. The revised version from September 2025 keeps that assessment:
- On some tasks, such as short essays or simple coding, they reached the "competent" level.
- On most tasks, such as maths or tasks where factual accuracy matters, they were still at the "emerging" level.
- Overall that made them "emerging AGI", the lowest level of general AI. The "competent AGI" level, which matches most earlier ideas of AGI, had not been reached by any public system at the time of writing, according to the authors.
That explains a contradiction you'll run into in daily work: the same model writes a usable email in seconds and invents a source that doesn't exist in the next sentence. Why that happens and what helps is covered in AI Hallucinations: Why AI Makes Things Up.
So when will AGI arrive? There are plenty of estimates and no facts. Amodei wrote in October 2024 that in his view powerful AI could come as early as 2026, though it could also take much longer. In the same essay he notes that many people doubt it will be built soon, and some doubt it will ever be built at all. Treat such statements as forecasts, even when they come from AI vendors.
Why this matters for your work with AI
For a small business, the question of strong AI isn't philosophy but planning. If you buy a tool because it is supposedly "almost AGI", you expect more breadth and reliability than the measurements support.
In my business it looks like this: each of my fourteen AI employees has their own name, their own job and their own personnel file. For those who prepare emails or quotes, approval and sending stay with me. How that works in detail is described in My AI Workforce: Fourteen Employees, One Boss. The model behind them is general purpose. The task I give each of them is narrow.
A made-up example: an electrical contractor wants AI to pre-sort enquiries from the inbox and draft replies. That is a clearly defined task with a check by the owner at the end. "The AI takes care of the office", on the other hand, isn't a task but a hope for AGI.
What to watch out for
- Ask for the definition. When a vendor or an article talks about AGI, ask which one they mean: consciousness, breadth or economic performance.
- General doesn't mean good at everything. A model that can do many tasks can't do them all equally well. Check figures, names and sources.
- Forecasts aren't facts. Plan with what a tool can demonstrably do today, not with what has been announced for next year.
- Keep tasks narrow. The clearer the task and the check at the end, the easier it is to judge an AI result.
Frequently asked questions
Is ChatGPT strong or weak AI?
That depends on the definition. In Searle's sense it is open whether ChatGPT understands anything; it can't be proven. IBM counts AI that answers questions based on your input as weak AI. The Google DeepMind research group, on the other hand, classed ChatGPT (as of 2023) as "emerging AGI", that is, general AI at the lowest performance level.
What does AGI stand for?
AGI stands for artificial general intelligence. It means an AI that doesn't just master one task but handles a broad range of mental tasks at least as well as humans. How broad and how well exactly is set differently by every definition.
Does strong AI exist yet?
An AI with consciousness has not been demonstrated. IBM describes strong AI as hypothetical. According to the DeepMind group's framework, no public system had reached the "competent AGI" level at the time of their assessment (September 2023).
What is the difference between narrow AI and weak AI?
In today's usage, none: both terms mean an AI geared to one specific task. Originally, Searle's weak AI had a philosophical meaning, namely that a computer only simulates mental abilities without understanding.
What is superintelligence?
That's the name for an AI that outperforms humans at practically every task. The technical term is ASI, artificial superintelligence. The DeepMind group lists it as the top level of general AI in its framework and notes there that it has not yet been achieved.
Where to go from here
If you want to go over the basic terms at your own pace, read Generative AI Explained, with Examples and look up anything unfamiliar in AI Terminology: A Plain-English Glossary. What AI systems already do on their own today and where that ends is explained in Agentic AI Explained: What It Is and Isn't. How other self-employed people use AI day to day without waiting for AGI is what I show in my community AI Employee Framework.
