Talk:Artificial general intelligence

Rewrite definition

We should rewrite the definition of this article by including the different performance levels of general AI which are already mentioned in the article body, as the current definition given in the lead is quite vague:

"[...] match or surpass human capabilities across virtually all cognitive tasks" – which humans? It can't mean all humans as that would be the definition of ASI. It also can't mean any human since this would imply that we have already achieved AGI, as ChatGPT is a general-purpose AI that is equal to (or somewhat better) than a completely untrained and unskilled human across virtually all cognitive tasks.

Most researchers and academics probably refer to the cognitive equivalent of an average trained human when they mention AGI, which would arguably be classified as a "Competent AGI" on Google DeepMind's performance scale (≥50 percentile of a skilled adult); while anything above that would be approaching ASI. Maxeto0910 (talk) 19:32, 20 July 2025 (UTC)

ChatGPT is a general-purpose AI that is equal to (or somewhat better) than a completely untrained and unskilled human across virtually all cognitive tasks. Do you have a reliable source for this claim? Elestrophe (talk) 21:21, 20 July 2025 (UTC)

ChatGPT is classified as an "Emerging AGI" system by Google DeepMind on their performance scale, which they define as a general AI "equal to or somewhat better than an unskilled human": https://aibusiness.com/ml/what-exactly-is-artificial-general-intelligence-ask-deepmind- Maxeto0910 (talk) 22:27, 20 July 2025 (UTC)

Okay, but that only shows that DeepMind's researchers believe this, which is not the same as showing that this (quite extraordinary) claim is true. Elestrophe (talk) 04:19, 21 July 2025 (UTC)

Sure, but that can be said about the conclusions of every source you could possibly imagine, including scientific studies and professional examinations. By that logic, no non-objective claim could ever be fully verified. Maxeto0910 (talk) 08:25, 21 July 2025 (UTC)

Anyway, regardless of whether or not this statement made by DeepMind is valid, it's certainly important to be precise when defining the capabilities of AGI, as we are arguably crossing a line now where the borders are fluid between the different interpretations of what constitutes AGI; although there is one broad and widely agreed-upon definition of what is AGI, there are several interpretations of and associated approaches to measuring when it is reached, which is why it's crucial to quantify the vague term "human capabilities" to ensure we have a solid basis and are at least to some extent talking about the same thing. Maxeto0910 (talk) 05:42, 30 July 2025 (UTC)

it refers to average human it not hard, ie 100 IQ human is the average. O1Proando3ASI (talk) 17:47, 30 July 2025 (UTC)

I think so too, at least that's the intelligence level most people refer to when they talk about AGI: an averagely intelligent, trained human. It's important to clarify this aspect in the definition. Maxeto0910 (talk) 21:16, 30 July 2025 (UTC)

perhaps rather then modifying the introduction it may be useful to add short part about something like AGI and Human Intellgence as it own section somewhere on the page O1Proando3ASI (talk) 15:09, 2 August 2025 (UTC)

Most people seem to define AGI based on what the average human can do, similarly to Google DeepMind's "Competent AGI". But sometimes discussions on whether an AI is AGI focus on tasks that only a small fraction of humans can do, like coding or research. Since there is vagueness in how people use the term, it may not be bad for the Wikipedia definition to also be vague in way that encompasses multiple possible definitions. So I'm ok with the current definition, but not necessarily against changing it. Feel free to propose a new definition here, or to make a bold edit. Alenoach (talk) 02:39, 31 July 2025 (UTC)

But sometimes discussions on whether an AI is AGI focus on tasks that only a small fraction of humans can do, like coding or research.
Yes, but I think it has been long settled in academic discussions that AGI encompasses being able to perform (virtually) all cognitive tasks a human could do.
it may not be bad for the Wikipedia definition to also be vague in way that encompasses multiple possible definitions
I'm fine with being vague and using various definitions as well, but in that case we should explain the vagueness and state that the term AGI, in most contexts, refers to an averagely intelligent, trained human, making this the primary definition. Maxeto0910 (talk) 13:42, 31 July 2025 (UTC)

Archiving the last 2 years of discussions totally unwarranted and counterproductive

Replying on Talk:Artificial general intelligence - Wikipedia curprev 22:27, 20 July 2025 Alenoach talk contribs 1,889 bytes −72,288 Archived the content from 2023 and 2024 undothank Tag: Replaced This modification is totally unwarranted as talk contained active discussion this decision should be reversed, and it undo. O1Proando3ASI (talk) 17:45, 30 July 2025 (UTC)

Archiving talk pages is standard procedure for Wikipedia, and discussions that haven't had a reply in more than 8 months are not active. MrOllie (talk) 18:14, 30 July 2025 (UTC)

it was largely counter productive; the thread were useful reference to current editors now buried for no good reason O1Proando3ASI (talk) 02:49, 24 March 2026 (UTC)

Adding other proposed definitions

There was a recent publication: https://www.agidefinition.ai/ which sought to provide a useful working definition. It's reported on here: https://newsletter.safe.ai/p/ai-safety-newsletter-63-new-agi-definition. That could be referenced alongside the 2023 DeepMind working definition.

I'm also sure I've come across others as well, for example the concept of 'Transformative AI' or 'Economically transformative AI' which is sometimes operationalised as being able to perform some fraction of 'economically useful work' according to some reference basket of work. BemusedObserver (talk) 15:37, 11 November 2025 (UTC)

WP:BE

AGI traditionally refers to a machine with human-level intelligence, nothing more and nothing less. Its relationship to economic output is not intrinsic to the definition. The definitions being cited appear revisionist and risk diminishing, rather than clarifying, the concept of AG O1Proando3ASI (talk) 02:51, 24 March 2026 (UTC)

However, that said, you could add it as a test for AGI rather than as a definition of AGI. O1Proando3ASI (talk) 03:05, 24 March 2026 (UTC)

Stick to reliable sources. The Center for AI Safety is a pro-AI advocacy group, so this website would need context from a reliable, independent source, otherwise this is just WP:OR and "criti-hype" from within the industry bubble. Grayfell (talk) 21:13, 24 March 2026 (UTC)

Somebody keeps deleting tests. I noticed the Employment Test was removed, and the Robot College Test (high school exam test) was also deleted.

WP:BE

Whoever is removing these tests should stop. They are relevant to AGI, and their removal unnecessarily diminishes the scope of the article. O1Proando3ASI (talk) 02:37, 24 March 2026 (UTC)

I have added a short list of major AGI tests so that their inclusion is explicit and any future removal is easier to identify and discuss:

Turing Test: conversational imitation of a human.

Robot College Student Test: passing university study at degree level (or high school level).

IKEA Test: robotic furniture assembly from parts and instructions.

Coffee Test: autonomous coffee-making in an ordinary home.

Employment Test: performing useful work at human level.

Suleyman’s Test: converting a fixed starting sum into much greater value.

Video Game Test: learning and succeeding across unfamiliar games. O1Proando3ASI (talk) 02:59, 24 March 2026 (UTC)

@O1Proando3ASI: Hello. I am the one who removed this. There are multiple serious issues with this content that would each need to be addressed.

The scope of the article is set by sources, not by individual editors.

All of this information needs to be directly support by reliable sources as it relates to 'artificial general intelligence'. Many of these sources merely documented the existence of this test without connecting it to this topic. This is what's known in Wikipedia parlance as Synthesis of published material.

Additionally, sources need to be reliable, and in most cases, this will also mean they are independent. When opinions are included, which should be rarely, they are directly attributed as opinions. This means, among other things, that promotional claims by Sam Altman that AGI has been achieved cannot be passed along as plain facts. Many of the sources you added are not reliable at all and should not have been cited. Many of the rest are primary sources and would again need attribution, and this most likely should come from independent sources.

Another issue is WP:NPOV. As an encyclopedia, Wikipedia has a 'mainstream' bias and should not be used to argue for a or against a novel position. The claim that the debate has 'shifted' would need much, much better sources than a single tech-gossip column in Windows Central. Likewise, list articles from WP:BI which do not mention AGI or even the Turing test are not useful for a topic as sweeping as this one.

Instead of compiling examples of things which might indicate AGI, start from reliable, independent sources about AGI and expand with sources which directly and unambiguously connect to that topic.

Grayfell (talk) 21:05, 24 March 2026 (UTC)

Journal of Artificial General Intelligence

I have removed this source from the lead:

• Goertzel, Ben (2014). "Artificial General Intelligence: Concept, State of the Art, and Future Prospects". Journal of Artificial General Intelligence. 5 (1): 1–48. Bibcode:2014JAGI....5....1G. doi:10.2478/jagi-2014-0001.

I don't see any indication that this journal is reliable. The Journal of Artificial General Intelligence is published by the Artificial General Intelligence Society via Paradigm Publishing Services (AKA reference-global.com). Paradigm's listed 'partners' include a few legit schools along with some seemingly predatory publishers, so I don't think it flatly confers legitimacy. Grayfell (talk) 23:01, 24 March 2026 (UTC)

The paper in question has 1100+ citations, and seems to be a good overview of the many various conceptual definitions and formalisms for AGI in 2014. The author of the paper is Ben Goertzel, who is credited with helping popularize AGI as a term in (Goertzel and Pennachin, 2007), and it looks like the 2014 paper is still highly regarded: it is cited by recent influential academic works such as (McLean et al. 2021) and (Hadi et al. 2024), as well as the book Artificial General Intelligence (Togelius, 2024) for popular audiences.

I agree that it isn't a good source for the definition given in the lead, though; the Goertzel paper does not support that AGI must "[match] or [surpass] human capabilities across virtually all cognitive tasks" as is currently stated. The Goertzel paper -- and virtually every other academic publication I could find -- suggest that a looser definition is needed, and while the current definition may be consistent with how some popular media uses the term, it does not appear to reflect how the term is used by AI researchers.

Note that the second source for the definition seems incompatible, as it does not contain the term "artificial general intelligence" or "AGI" anywhere in the paper body. So I've marked that as having failed verification, and I will try rewriting the lead sentence to match the 2014 Goertzel paper if the consensus is that its accceptable as a reliable source. I agree that the journal its published in doesn't have much information about it -- is that a valid reason on its own for considering the 2014 paper unreliable, and if so, is there a better source that can be used instead? Ideally not just a source that gives a single definition, but one that actually analyzes multiple definitions, like the 2014 paper. CeruleanStyle (talk) 16:33, 13 April 2026 (UTC)

I understand that the source may be convenient, but as I said, I do not think it is an ideal source. Being widely cited is often misleading, for a few reasons. (As one extreme example, Hans Eysenck had one of the highest cite-counts in psychology, but this was because he was extremely prolific and aggressively cited himself in his own work. A lot of papers existed specifically to refute his findings and methods, which further inflated his cite count. Being widely cited was not a good thing in that case).

So, if more recent reliable sources cite this paper to provide a definition, it would be better to use of those newer sources. If they don't provide a definition, than being cited by these sources doesn't matter here.

Grayfell (talk) 22:47, 13 April 2026 (UTC)

What Virtuality

That "virtually" in the cited definition for AGI in the intro sentence leaves loopholes for anything (stupid) to be called AGI.
Has anyone a reputable source for an AGI definition with less .. virtuality? --Bernd.Brincken (talk) 16:56, 7 June 2026 (UTC)

In case there is a confusion, "Virtually" isn't really about virtuality, it's a synonym of "almost". No problem with changing "virtually" to "almost" or "nearly" if it's clearer.

Or maybe you really meant that there could be an AI that does almost any intellectual task at least as well as humans but that is stupid? That would sound implausible to me, since even with "almost" it's a high bar. Alenoach (talk) 20:55, 7 June 2026 (UTC)