Artificial general intelligence, or AGI, is a proposed level of AI capability that works well across many different kinds of tasks rather than excelling only at one defined job. The term is useful, but it has no single universally accepted pass/fail test. Researchers and companies place different weight on breadth of ability, performance, and how independently a system can work. A claim that a model is “AGI” therefore needs a definition and evidence, not just an impressive demonstration. Google DeepMind's Levels of AGI research makes these dimensions explicit.
How is AGI different from narrow AI?
Narrow AI is built or deployed for a bounded purpose. A spam filter sorts messages; a system trained to play Go chooses moves in that game. Such systems can be excellent at their tasks without being able to take on unrelated work. Our introduction to AI explains the broader field.
General-purpose AI assistants sit between those simple examples and the proposed AGI goal. As of September 2026, products such as ChatGPT and Claude can help with writing, coding, data analysis, and documents or images. That breadth is real, but capabilities vary by model, tool access, task, and evaluation conditions. Broad usefulness alone does not establish the robust performance or autonomy required by a particular AGI definition.
The distinction concerns what a system can reliably do, not whether it literally thinks like a person. Operational definitions such as DeepMind's and OpenAI's assess capabilities; they do not test consciousness.
Why do definitions of AGI differ?
One approach emphasizes work. OpenAI's charter defines AGI for its mission as highly autonomous systems that outperform humans at most economically valuable work. That is OpenAI's definition, not a universal standard. Google DeepMind's research instead proposes levels based on how broad a system's skills are and how well it performs, while also discussing autonomy and the difficulty of measuring progress.
These approaches can lead to different conclusions about the same system. A model might score well on exams and write useful code yet still fail at a new, long-running task. Another system might complete a tightly defined workflow with little supervision but lack broad ability. Performance, generality, and autonomy should be examined separately.
Are today's AI assistants examples of AGI?
They are examples of increasingly general-purpose AI, not automatic proof of any particular AGI threshold. To assess an AGI claim, ask what tasks were tested, whether the system encountered new problems, how often it succeeded, how much human help it received, and whether the results were independently reproduced. A benchmark score or polished demo answers only part of that question.
Two examples in earlier discussions of AGI illustrate the problem. AlphaGo defeated Lee Sedol in March 2016, an important achievement in Go, but winning at Go did not show competence across unrelated tasks. GPT-4 was introduced in March 2023 with broader text and image capabilities; OpenAI also documented factual and reasoning errors. It is a historical milestone, not a current model benchmark or a certified AGI example.
What would a useful AGI test look like?
Imagine asking a system to investigate an unfamiliar business problem, learn the relevant constraints, analyze a spreadsheet, compare evidence, write a recommendation, and adjust when new facts arrive. A useful evaluation would record its accuracy across the steps, where it needed tools or expert help, whether it noticed its own errors, and whether the result held up in several unfamiliar settings. A single successful run would be insufficient.
For legal work, the standard is especially demanding. A system might draft a case summary or suggest research paths, but a lawyer still has to verify authorities, quotations, jurisdiction, and current law. The presence of an AI agent that carries out several steps does not establish AGI either: a multi-step workflow can be built around systems with narrow or uneven abilities. The NIST Generative AI Profile documents that generative systems can present false information with confidence, including fabricated citations.
AGI describes a research goal and a disputed measurement question. For a lawyer or business owner, the practical question is narrower: Which tasks has this particular system been shown to perform accurately, under what conditions, and what human checks remain necessary? Start with that evidence and the site's responsible AI guide, whatever label a vendor uses.
Sources and date
Product examples were checked September 25, 2026; features and model availability can change. The main sources are Google DeepMind's AGI framework, OpenAI's charter, OpenAI's ChatGPT capability overview, Anthropic's Claude overview, the AlphaGo match record and GPT-4 announcement, and NIST's generative AI risk profile.
