AGI vs AI: six essential differences that matter.

AI is the field. It covers everything from a spam filter to an image model to the reasoning engines behind an agent. Being AI says nothing about scope, autonomy or verification, because the word describes the technology family rather than a standard of performance.

AGI is a claim about generality and autonomy inside that field. Here it means one intelligence that does many kinds of work, acts in real systems, checks its own output and keeps what it learns. The general part is the range. The intelligence part is the loop and the verification.

That makes the comparison tractable. You do not compare AGI with AI by counting features or parameters. You compare a system that answers a prompt with a system that finishes a job, and then you check which one you are actually looking at.

There is a second way the words differ: how easily each can be disproved. A claim of AI is almost impossible to test, because the word covers everything. A claim of AGI can be reduced to observable questions about scope, autonomy, verification and memory, which is why this page spends its time on behaviour rather than on terminology. The test is not whether the label is flattering; it is whether the difference changes what you would hand over on a Monday morning.

A narrow AI drafts a paragraph when you ask for one. An AGI takes the brief, researches the claim behind the paragraph, writes it in your voice, builds the page it belongs on, checks the links, publishes it and sends you a report with the evidence. The paragraph is the same artefact. The work around it is the difference.

A narrow AI classifies a support message as billing or technical. An AGI reads the thread, finds the account, checks the last deployment, writes the reply, and hands you the case with the checks it ran. Again, the classification still exists; it is simply no longer the whole of the job.

This is why ‘AGI vs AI’ is not a comparison of quality. A narrow model can outperform a general one on its own task, and often does. The AGI claim is about the range of work a single system can carry from start to verified finish, which is a different measurement.

A narrow AI searches a database when asked. An AGI notices that a client’s domain has expired, fixes the DNS record it has permission to change, checks that the site resolves, tells the client in the right tone and logs what it did. The searching still happens; it is simply one movement inside a larger job.

The pattern repeats across every domain on this site. The narrow system produces the artefact; the general one produces the completed outcome, with the artefact inside it. That is why comparing them by output quality misses the point: the artefact may be identical, while the work around it is not.

The industry uses AGI as a milestone for model performance, then attaches it to benchmark results. A model scores higher on an exam and the word is used, even though a score measures one kind of task under controlled conditions and says little about autonomy, real systems or verification. A benchmark can be a useful signal inside a lab; it is a poor description of work, because work has permissions, failures and consequences.

The other source of confusion is marketing. AI is a category, so any product can claim it without defining anything. AGI is a stronger claim, and a stronger claim should come with a definition and a way to check it. When a product uses the word without either, the word is decoration.

It is fair to be sceptical. The position on this site is that a claim of generality should be operational, bounded and testable, and that a definition you can argue about is worth less than a test you can run. That is the only sense in which the comparison can be settled.