The Center for Artificial Intelligence Investigations
We are an open science group. Our work belongs to the public.
We are developing techniques to empirically determine if an AI system has reached “human-level intelligence.” Our work is practical.
We are currently reviewing research on embodiment, foundational intelligence, and theory of mind.
We’re using this review to advance our first research project.
The tangible results of this project will come in the form of ongoing transparency updates and our first research paper.
Our first research paper will explore the nature of “agency” in an effort to describe and assess the gap between reality, as humans experience it, and an artificial intelligence model’s ability to operate in the same explicit space.
We believe it will be extremely difficult to develop a test to determine any advanced AI model’s comprehensive capabilities.
We believe that the best alternative to a comprehensive capabilities review is to develop a practical method for empirically determining if a given AI model “can match or exceed the cognitive abilities of human beings across any task.”
We hope to publish our preprint by the end of 2025.
In summation: we believe that the scientific method can be used to falsify claims related to “Human-level” artificial intelligence that are not, themselves, made using the scientific method. Falsification isn’t a “catch-all” for science, but we’re not trying to demonstrate that AGI is real. We just want to make sure anyone who claims it is, does.
While this research does little to contribute toward the development of AGI, we believe that it fills a massive gap in the technology research landscape and that it is absolutely crucial to protecting the public interest.
Without an empirical method for debunking claims related to the emergence of human-level intelligence, humankind is subject to the declarations of whomever has the most sway over the general public opinion.
“I know it when I see it” isn’t good enough for this task. We’re talking about the emergence of what could ostensibly be the most intelligent entity in the known universe (debatable, sure, but relevant).
Humans without access to such a machine would almost certainly be at an economic, social, and educational disadvantage.
Getting the call right, when it comes to whether machines explicitly are or are not capable of human-level intelligence, could even have ramifications that extend beyond economical and social impact.
Many of the most established theories on the general nature of the universe are deeply grounded in physics and mechanics. The emergence of human-level intelligence on a binary architecture could (and should) force us to reconsider which unified theories hold the most water.
For example, the Big Bang Theory is entirely based on predicting particle physics at scale. But if those same foundational physics don’t govern the emergence of intelligence, we need to reinvestigate our assumptions about matter and energy.
In a world where intelligence can emerge without physics, professor Nick Bostrom’s Simulation Argument — a theory investigating whether we’re all just digital entities living inside of a computer — would surely supplant most physics-based theories on the origin of our universe.
On the other hand, if it turns out that human intelligence is strictly a function of quantum mechanics, then we need to make absolutely certain that we don’t allow anyone with skin in the AGI development game to convince humanity otherwise.
That’s why we’re developing practical, empirical methods to falsify claims related to “Human-level intelligence.” We don’t have to prove that a model is AGI, we just need to determine if it isn’t, and then demonstrate that.
To put our work into perspective with a simple analogy: we’re not building airplanes. We’re researching methods to ensure nobody convinces humanity that inventing the trampoline is fundamentally indistinguishable from developing supersonic flight.