Some of the biggest names in AI research recently conspired to define “artificial general intelligence” once-and-for-all in a landmark research paper entitled “A Definition of AGI.”
As one of the first technology reporters assigned specifically to “the AI beat,” I’ve spent my entire career trying to define AGI. And, in my experience, everyone has a definition for AGI, and they’re mostly the same. But nobody can give me a threshold for their definition.
Ghosts?
You can define anything that doesn’t exist using any words you want, as long as you don’t provide a threshold. For example, I can define a ghost as “the lingering spirit of a deceased person.” I can also define a ghost as “an imaginary entity often described as the lingering spirit of a deceased person in fiction.”
As far as science is concerned, there’s no difference. Either you believe in ghosts or you don’t, but the scientific method requires a threshold—it demands measurements, falsification, and the astute application of Occam’s Razor.
In lieu of this, we can’t do science. Without a hyperspecific metric by which “ghostness” can be measured, a threshold that defines the difference between ghostness and all things that cannot be ghosts, any definition for the word “ghost” is functionally useless. And it’s the same with AGI.
In reality, defining AGI is easy. Here’s IBM’s definition:
Artificial general intelligence (AGI) is a hypothetical stage in the development of machine learning (ML) in which an artificial intelligence (AI) system can match or exceed the cognitive abilities of human beings across any task.
Here’s my definition from a 2022 article in Undark Magazine:
“An AGI would in theory be capable of learning anything that a human can, if given the same access to information. Basically, if you put an AGI on a chip and then put that chip into a robot, the robot could learn to play tennis the same way you or I do: by swinging a racket around and getting a feel for the game. That doesn’t necessarily mean the robot would be sentient or capable of cognition. It wouldn’t have thoughts or emotions, it’d just be really good at learning to do new tasks without human aid.”
Anyone can define AGI and, upon doing so, report existing work as evidence for the efficacy of their definition. That doesn’t, however, add credence to the notion that a given developmental paradigm is any more likely than another to achieve AGI.
Much like ghosts, AGI remains theoretical and there’s no scientific evidence that it will be realized within any specific timeframe, or at all.
Defining AGI, circa 2025
In the quest to define, codify, measure, benchmark, and establish a paradigm for AGI, countless people have spent innumerable hours and immeasurable sums of money to move the needle from really good AI to something indistinguishable from human-level intelligence.
Thousands of papers are published on the subject of advanced AI, most referencing AGI technologies, every year. The sheer amount of research is overwhelming and the marketplace of ideas, when it comes to AGI, has become saturated with definitions, new benchmarks, and multi-discipline research.
It’s been my experience that the majority of papers on the subject of AGI are, essentially, attempts to define AGI and describe a potential development paradigm toward achieving it.
With that in mind, I read The Center for AI Safety’s paper, “A Definition of AGI.” The paper starts with the following definition:
“AGI is an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult.”
It bears mention, at this point, the paper was written by the following people:
Dan Hendrycks, Dawn Song, Christian Szegedy, Honglak Lee, Yarin Gal, Erik Brynjolfsson, Sharon Li, Andy Zou, Lionel Levine, Bo Han, Jie Fu, Ziwei Liu, Jinwoo Shin, Kimin Lee, Mantas Mazeika, Long Phan, George Ingebretsen, Adam Khoja, Cihang Xie, Olawale Salaudeen, Matthias Hein, Kevin Zhao, Alexander Pan, David Duvenaud, Bo Li, Steve Omohundro, Gabriel Alfour, Max Tegmark, Kevin McGrew, Gary Marcus, Jaan Tallinn, Eric Schmidt, and Yoshua Bengio.
Pundits such as myself have long called for a scientific quorum to discuss AGI, and, for those who don’t recognize the above names, I’ll just say that if there were an AGI researcher Walk of Fame, most of those people would have stars on it.
Together, they came up with a framework that “dissects general intelligence into ten core cognitive domains—including reasoning, memory, and perception—and adapts established human psychometric batteries to evaluate AI systems.”
In other words, they created a new benchmark to measure the capabilities of existing systems against the proposed capabilities of a system that does not exist.
With all due respect to the researchers, this isn’t good enough. It’s ghost hunting.
I don’t write that as a token of criticism. In fact, I believe their paper represents the best, most current, and most useful definition and benchmark there is. I don’t think any other team of researchers could have done a better job, given the task at hand.
The reality is that AGI doesn’t exist yet. Testing AI to see how close it is to becoming a hypothetical technology is as difficult as measuring how successful a ghost summoning went—on a scale of 1-10.
Reframing the frameworks with more frameworks
As world-renowned technoethicist Wendell Wallach pointed out in his article from last month’s issue of the AGI Ethics News newsletter, any AI model sufficiently advanced enough to be considered an “AGI” will require a modicum of “moral intelligence.”
“Without such faculties,” he wrote, “computational agents will not only fail to fully embody moral intelligence but also fail at other skills necessary for artificial general intelligence.”
Our publisher, KB Miller, recently developed a proposed AGI Trust Table denoting more than a dozen vectors of trustworthiness, complete with examples of corresponding model behavior.
While our table isn’t designed to certify or define an AGI system in the wild, it is, we think, utilitarian. As Miller put it, “offered as evidence of its usefulness is that, if an AGI had the characteristics shown, it would be working toward all twelve of the virtues identified by techno-ethicist Shannon Vallor as core to three great human ethical frameworks.”
We argue that this or a similar framework, as well as the aforementioned requirement for moral intelligence, are just some of the many necessary prerequisites for the development of a system capable of meeting the definition of “an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult” given by the researchers, the one previously given by IBM, or even my own clumsy definition from 2022.
Ultimately, I salute the Center for Artificial Intelligence Safety and its team of world-renowned researchers for their work. They’ve modernized the current definition of AGI and given developers working on current systems (LLMs especially) a new framework for benchmarking.
It is undeniable that this represents progress in the field of artificial intelligence research. Whether you feel that this is progress towards the development of an artificial general intelligence is entirely up to you.
Ethical considerations:
- Can an AGI possess freedom of thought, conscience, or a functional equivalent?
- What standards of subjective report, self-awareness, or introspective access should AGI meet to ensure reliable, ethical participation in human institutions?
- Will AGIs be required to cultivate and apply practical wisdom (phronesis) as a guide to ethical action in the world-and how is this assessed and enforced?
The questions around AGI ethics are complex and evolving. Get thoughtful, well-researched commentary and the latest developments delivered straight to you by subscribing to AGI Ethics News.
A veteran AI journalist, Tristan was the creator and managing editor of The Next Web’s “Neural” imprint. He is one of the first and most prolific reporters to cover artificial intelligence as a full-time beat.