What is artificial general intelligence?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Artificial general intelligence, usually shortened to AGI, describes the idea of an AI system with broad, adaptable intellectual abilities. Rather than being designed for one narrow task, it would handle many kinds of problems and transfer what it has learned to unfamiliar situations. There is no single definition that all researchers accept.
The contrast is with narrow AI: systems built for a limited purpose, such as identifying unwanted email or playing a particular game. A system can be outstanding at that purpose without being able to do unrelated work. Generality concerns the range of abilities, not simply exceptional performance on one task.
Imagine a system that can learn an unfamiliar game, interpret a new set of instructions, plan a project and adapt when circumstances change. Such examples help describe the ambition, but they do not establish a universal test. Researchers disagree about which abilities matter, how strong performance must be and how much help from people or other software should be allowed.
It is useful to separate breadth, competence and independence. Breadth asks how many different tasks a system can handle. Competence asks how well it performs them. Independence asks how much human direction it needs. A system could be broadly useful while requiring frequent supervision, or highly independent within a very narrow task.
A benchmark is a collection of tasks used to compare systems. Benchmarks provide evidence about particular abilities, but no single score settles every question about general intelligence. Results also depend on whether the tasks are genuinely new to the system and whether success transfers beyond the test itself.
AGI should not be treated as a synonym for consciousness, meaning subjective experience or awareness. Nor does it necessarily mean superintelligence, a proposed capability substantially beyond human performance across many intellectual activities. Those are different questions that are often mixed together in public discussion.
Consider two systems that receive the same score on a set of puzzles. One may have encountered nearly identical examples during training, the process through which an AI system learns from data. The other may solve unfamiliar puzzles from a small number of instructions. The scores alone do not tell us whether the systems have the same ability to learn something new. Tests of generality need to examine that distinction, not just count correct answers.
Comparisons with people also need a clear reference point. Does “human-level” mean the average person, someone trained for the task or a specialist? Are the system and the person allowed the same tools and time? These choices change what the comparison establishes. Researchers have proposed frameworks to make such comparisons more precise, but a proposed framework is not a universally accepted standard.
Neither usefulness nor risk arrives only when a system crosses an agreed AGI threshold. A narrowly capable system can already be useful or harmful. Broad capability also does not establish good judgment, trustworthy behavior or a legitimate right to make decisions on someone’s behalf. Those questions need their own evidence and human decisions.
In healthcare, answering difficult licensing examination questions, summarizing a longitudinal chart and coordinating a patient’s care are different demonstrations of ability. Success at one does not establish all the others, or confer authority to act as a clinician. A health system need not settle the definition of AGI before evaluating a bounded use. It can ask what the system does reliably, where it fails and which decisions remain human responsibilities. Those answers are more useful for operating a service than the label alone.
Sources: Morris and colleagues — Levels of AGI; Chollet — On the Measure of Intelligence.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy