Towards an Intelligible Conversation about Intelligence
At UC Irvine, I specialized in Intelligent Systems. Which is rather strange because I don’t believe there’s much consensus, even among those with the specialization, about how we should define intelligence. We do touch on the ambiguity of it. We go through the classic thought experiments that probe at consciousness, intelligence, originality, subjectivity, and sentience: John Searle’s Chinese Room, Alan Turing’s Imitation Game, Ned Block’s Chinese Nation, Hilary Putnam’s Twin Earth, Frank Jackson’s Mary’s Room, David Chalmer’s Philosophical Zombie, Thomas Nagel.
So I can look at a human-computer system and tell you how the concept of intelligence might apply or how intelligent algorithms might improve or threaten it. But I don’t necessarily know by what definition of intelligence I can call myself, a dog, a swarm of ants, and an algorithm all intelligent. I can certainly explain why it’s a hard problem though.
Consider a point Bryan Cantrill (for whom I have a perfectly normal amount of parasocial admiration) brings up in his talks about the definition of “software”: he notes it is very difficult to find a definition of software that doesn’t include Euclid’s algorithms from thousands of years ago, yet no one in daily life really means to discuss software that broadly. More specific definitions that try to capture our recent imaginations can take us to very strange places with implications we might not consider intuitive, especially if the purpose of the definition is to substantively draw the distinctions between software and hardware or software and pure mathematics.
Let’s try for a moment then to do this with “intelligence” as a term. I asked some technical peers to offer a definition off the top of their heads.
“Intelligence is an emergent decision-making capability that arises from communication between neurons.”
If we want to scope out “intelligence” as just something belonging to organic neurons, we would omit all supernatural entities or deities of all religions. Note that we’re not discussing the existence of these entities: just that if they existed, we couldn’t call them intelligent under this definition.
And what of a person who is pronounced brain dead, but whose body can be kept alive through machines? They may have neurons firing and have intelligence distributed throughout their nervous system to some degree, but we probably would not call this person capable of making intelligent, or even intelligible, decisions. Their decision-making might be as random as a fly picking stocks.
What about a single-cellular organism that can transmit genetic instructions, evolve while reproducing, and exhibit conditioning to external stimuli like vibrations and taps? Intelligence may not require neurons at all when other sets of molecular networks within the cell can perform signal processing. How far can we take that? Are there molecular networks outside of cells that can process signals, perhaps silicon-based systems? What about cyborgs? Silly hypotheticals are often a really good shortcut for identifying edge cases.
“Intelligence is passing the Turing Test.”
While it’s everywhere in pop culture, and Turing himself certainly has warranted the attention, the Turing Test in practice is not a very helpful metric. Under this test, an AI chess player that can defeat every human but cannot chat has the same intelligence as an algorithm to discover new proofs, which is to say none as neither pass the test of seeming like an intelligent conversational partner.
Meanwhile, a paralyzed human in a coma couldn’t pass the Turing Test, but they could wake up tomorrow with full control and suddenly pass it. Was this person for a while lacking in intelligence? Was the intelligence dormant but not lacking? We can take a radical posture like David Hume and treat the mind as a different entity every time it changes in any way at all, including the picosecond-by-picosecond update to our memories and neurons, so perhaps we’re talking about two different minds altogether: one intelligent, one not.
“Intelligence is the capacity to learn and make decisions.”
This might actually have compounded the complexity we had before, because now we have the burden of defining learning. Maybe it’s as simple as updating a prior belief in the face of new information where the belief is a probability attached to a hopefully falsifiable proposition, or maybe that means creating new models of reality through innate deductive reasoning ability and real-world empiricism.
But then what about a perfect reasoning agent that never needs to update its belief? Maybe it never needs to learn, or maybe we don’t even know if it can since it never has to. What about an algorithm that has 100% accuracy in mapping inputs to outputs? It will never be wrong yet also performs no reasoning. Is it a snapshot of what an intelligent system looks like? Is it a piece of intelligence itself? Is it simply an artifact, a byproduct, downstream of an intelligent agent’s execution of its tasks?
Or let’s go back to religion: an omniscient God could never learn because it already knows all, but it would be strange not to consider that God intelligent.
Or what about a damaged brain which can store and retrieve memories in response to the environment it is placed in, but cannot evaluate which of two choices it would prefer? It can still feel pain and pleasure while recalling both through memories. Most people would be upset if you gave an entity like this nothing but painful experiences for no obvious reason. It’s processing information but not able to demonstrate any performance on any cognitive task.
Intelligent Quotients
“Wait, Mank, we measure intelligence all the time. We have IQ tests, standardized tests, holistic grading, Gardner’s theories, h-index, Mensa and other elite intellectual organizations, certifications and degree programs, and entire specialties within biology, psychology, and education dedicated to studying intelligence and patterns of influence. That can’t all be a waste of time right?”
We are certainly measuring something with these tests that is repeatable and often even useful when it helps us predict other variables or properties. For instance, a good SAT score might predict a higher likelihood of completing a 4-year degree. Because admissions officers generally believe in the predictive power of SAT scores (after all, it’s how they demonstrated competency at one point), the scores might also predict a higher chance of being accepted into college.
Does that mean intelligence is specifically what is being measured? Rather than a factor, or set of latent factors, that those external achievements have in common? Maybe SAT scores actually reflect the access to SAT preparation material that a student has, which suggests they would have access to more resources across the board in their lives. Someone who can hire an SAT tutor or afford elite bootcamps for their children might also be able to hire tutors for general K-college education.
Consider how Columbia University did not allow women to be admitted until 1983. If we were to measure the average profile or highest scorers in the student body of Columbia at 1982, a set of results discovering that 100% of those students are men actually says nothing at all about men’s readiness for college compared to women’s. It might even cause more harm than good to publish a study like that given how easily it can be misconstrued, especially in the modern media environment that proposes sensationalist social theories based on a handful of statistics without context.
In this case, the study’s findings only reflect a gendered allocation of resources rather than the competitiveness of male applicants relative to female applicants. In fact, we might find that this kind of gendered allocation of resources had existed along the entire sociopolitical context our subjects operated in, from K-12 education to community projects. Any assessment given at any level of the education system in 1982 might reflect a gendered discrepancy of resources allocated to people.
All attempts to isolate intellectual performance given a sociopolitical distortion like this should be evaluated with a lot of scrutiny, especially when the motivation to measure intelligence is sociopolitical iself. Sometimes it’s the result of a disagreement about progressive education reforms and how to measure discrepancies between student populations. Sometimes you end up creating effective rhetoric for eugenicists looking to justify their selection as natural.
g Factor
If you’ve been grinding your teeth reading this so far, my guess is that you want to pop off about how the g factor proves that some property called intelligence does exist that cuts across our cognitive assessments.
Charles Spearman (1863-1945) was a psychologist who produced the term “g factor” as part of his theory that there’s a single general intelligence factor underlying variations on cognitive tests. Spearman himself said:
When asked what g is, one has to distinguish between the meanings of terms and the facts about things. g means a particular quantity derived from statistical operations. Under certain conditions the score of a person at a mental test can be divided into two factors, one of which is always the same in all tests, whereas the other varies from one test to another; the former is called the general factor or g, while the other is called the specific factor. This then is what the g term means, a score-factor and nothing more. But this meaning is sufficient to render the term well defined so that the underlying thing is susceptible to scientific investigation; we can proceed to find out facts about this score-factor, or g factor. We can ascertain the kind of mental operations in which it plays a dominant part as compared with the other or specific factor. And so the discovery has been made that g is dominant in such operations as reasoning.
So basically, Spearman’s idea of a two-factor model tried to explain the noisy variation in test scores by proposing a general factor that could be used as baseline intelligence of a subject, and then a second specific factor that explains the variation on any particular test.
Say a student scores a B one day and B+ on an identical assessment another day; perhaps they have a strong baseline g factor that brings them to a B, and then a second specific factor determines whether it rises to a B+. After administering a battery of cognitive tests, we tend to find positive correlation across vocabulary development, spatial reasoning, working memory, quantitative reasoning, information processing speed, etc. We can fairly reason that someone consistently good at one of these tasks has a greater chance at being good at the other tasks.
In 2026, the consensus as I understand it is that there is observably something that behaves like a general factor g as a statistical pattern. It holds up in replication studies across cultures, age groups, task modules, etc. The most common position among those who work with psychometrics seems to be that there are genuinely some cognitive capacities and capabilities this can measure. But we do not have evidence to say with certainty that what it’s measuring is something we can distill into a single thing called “intelligence” as we’ve used it before.
Some argue that during childhood growth and development, there’s a reinforcing cycle between memory, reasoning, language, and learning which contributes to the correlations we observe through the g factor. This idea is sometimes referred to as mutualism given the mutual growth along different axes. That might imply that high scores on cognitive assessments are the result of nurturing them with exercises that promote this cycle of mutual growth rather than a predisposition at birth.
Then we have the process overlap theory associated with Kristof Kovacs and Andrew Conway. If every test involves invoking the same processes for reasoning, memory, spatial awareness, then the g factor correlation might actually reflect the effectiveness of handling the overlap. Perhaps someone with a low g factor actually has a photographic memory but can’t solve simple riddles, or can derive very complex formulas from first axioms while never finishing a test in time. A high g factor is measuring the ability for different processes to work together to finish tasks under the constraints of cognitive assessments.
Modern neuroscience in general has adopted a more nuanced paradigm than treating the brain like a collection of modules with specific tasks allocated to specific parts. Instead, neuroscientists more often work with the concept of distributed computations across interactive brain networks. What we call intelligence might emerge as the result of the coordination between multiple whole-brain networks, not a single physical property or location we can connect to intelligence.
But this is not the same as an established objective measure of intelligence. There is still debate over whether the g factor represents underlying cognitive resources, is an emergent property of brain network intereactions, reflects synergy between overlapping executive processes, or something else entirely. There is no “g center” or module in the brain where we can see a specific batch of nerves like a GPU in a PC to produce intelligence computations behind the g factor.
A quick and dirty analogy here is temperature as an emergent property of a cluster of matter. The temperature emerges as a result of constructing an average over the kinetic energy within all the particles of a system as they collide. We wouldn’t take a temperature of a single particle, that’s not meaningful or even possible with standard tools. We also wouldn’t be correct if we said that the average kinetic energy meant every particle had exactly that level of kinetic energy, or even much at all about how the particles actually interact and collide. We would need a very sophisticated understanding of molecular physics and thermodynamics to begin modeling that.
We have simply taken a useful, if blunt, measurement that can help us in specific tasks. Intelligence should, in my opinion, be taken in the same spirit. Cognitive tests from which we derive an intelligence quotient or estimate a g factor should be seen the same way. Children who score consistently high on cognitive assessments and have a high g factor might benefit from more attention to make sure they’re not understimulated.
On the flip side, children with a low g factor might benefit from different exercises that work on expanding core reasoning and memory skills. Just because 60-80% of the g factor variation has an association with inheritable genetic differences doesn’t mean 60-80% of any individual’s performance on intelligence is a fixed or predestined trait. There is no “intelligence gene”; we can only measure the g factor after a child has been through a significant amount of development to even take a cognitive assessment.
What is true as far as genetic predispositions for intelligence go are polygenic scores that aggregate information from thousands of genetic variants and have real predictive power. But once again, we do not have the evidence to conclude there’s anything deterministic or innate about the intelligence metrics.
And remember: my exposure to the field through brief research experiences aside, my opinion on all this is not an expert’s and you should definitely not be citing this page as an academic source. I try to communicate the concepts behind science and technology the best I can, and I update this page to the best of my abilities after I receive or discover errors.
In Summation
We can go on with examples and arguments and counterarguments for a long time on specific claims about intelligence, and I welcome feedback, but hopefully it’s becoming clear that we’re grappling with very complex questions here that cannot just be solved with more compute or a standardized test. Each culture has different institutions and norms that will necessarily inform what you think of terms like consciousness, intelligence, intuition, self, self-awareness, sentience, soul, humanity, creativity, and more.
The good news is that if you like learning, you’ll have a great time diving into this type of thicket. It’s a very old question with very smart people offering very thorough answers. There is great analysis still to be done in studying the historical transformations of the concept of intelligence around the world.
If you’re trying to market a probabilistic pattern-matching product as being sufficient for emergent general intelligence… well I commend your ambition, if nothing else.