It's a lazy Wednesday night in Amsterdam, work is done, I’ve finished reading my latest book, couldn’t be bothered with scales on the piano and don’t particularly feel like couching the latest TV series with a bag of M&M’s. So, let’s do a bit of study and share random thoughts on the topic of intelligence.

Art work is by https://www.instagram.com/manoloide https://superrare.com/manoloide#top

Defining intelligence is both the starting point and end state of working in the field of artificial intelligence. What is it that we are trying to synthesize anyway? Can something we haven’t clearly defined be modelled?  Way back in the 50’s at the Dartmouth Conference the titans of computer science kicked off the study of AI. They generalized in the direction of what intelligence could be ,then from that generalization they determined that: If you could define our brain’s inner workings in a series of logical algorithms and models, you could (theoretically) create a synthetic version of intelligence.

 In this short blog post I’ll briefly summarize the study of intelligence over the last hundred years, talk a little about a fantastic book on this topic called “Bots and Beasts” by philosopher Paul Thagard. Then I’ll open the can or worms of cognitive bias, by way of “The Dunning Kruger Effect”. Which tackles our tendency to overestimate our own competence, which in turn raises a question on how we assess intelligence.

Back in 1904 Psychologist Charles Spearman created a series of tests that measured people’s abilities to complete common tasks. What he found was that people scoring highly in a specific area tended to have higher scores across the board. Spearman aggregated all the results and came up with a single score for general intelligence: (creatively named) “G”.  An easy metaphor for the logic behind this is” fitness”.  In broad strokes: If you are a professional footballer, there is a good chance that you’ll perform better at other sports than someone who isn’t. Even if it’s just your blunt stamina and endurance, the professional sports person is going to perform at a high level across all sports.  It’s not much of a stretch to think of how the IQ score was created.  In the 40’s there was also the addition of “fluid intelligence” and “crystalized intelligence”, the brains we are born with (fluid) and cultural intelligence (crystalized) and many others that slowly but surely created an intelligence defined not by a single score but based on features. (this went hand in hand with the advances in cognitive science, neurology and psychology)

In his book “Bots and Beasts, what makes machines, animals and people smart” ? Paul Thagard attempts to find correlations between intelligence in animals, people and machines. To do this he follows a simple concept from psychologist Greg Murphy and expands upon it. Murphy suggests you just take:

  • Examples of intelligent people
  • Find typical features in those people
  • Explanations provided by the concept

If you go down this rabbit hole you should have at least one or two features of intelligence. Let’s use it on one of my heroes Esther Perel as an example.

What makes her intelligent? Well, she seems to have a deep understanding of the subject matter, a lot of experience, she can translate that into being perceptive in her therapy sessions. She is also communicating and acting on those perceptions in a way that is understandable and relatable for people who aren’t professors in psychology. .

I’ll save you the time of figuring this out and just list the features that Thagard defined while following this logic:

  • Perception
  • Problem solving
  • Planning
  • Deciding
  • Understanding
  • Learning
  • Abstracting
  • Creating
  • Reasoning
  • Feeling
  • Communicating
  • Acting

There are some interesting features in this list, some of them spring out at you and others require a little bit of explanation. For example: Perceptual intelligence could be a feature of a talented police officer, that person has their radar tuned to what is going on around them and can quickly make distinctions on where they are needed.

Acting Intelligence is for the doers, those people that just know when to cut the chatter and step forward and do something, they can judge when the time is  (and is not) right to act.

Feeling or emotional intelligence resonates very strongly in our modern brains. We are constantly interacting and connected with other human beings and our ability to solve problems in this domain is in some ways more valuable and important than solving a mathematics puzzle. We could break it down even further into social intelligence, empathy or self-regulation, all of which are important features of intelligence that deserve their own mini-IQ scores.

In modern times we can also add the concept of “collective intelligence” to the list, the hive or swarm intelligence we have observed in bees or swallows is one I find particularly interesting. Working in teams and solving problems as a group is so rewarding and hopefully everyone reading this has been fortunate enough to experience this. I won’t go into this in too much detail, but I believe that working with AI models could be a form of collective intelligence.

Defining intelligence is tough and it’s a moving target; to make it even tougher it also has an inherent weakness in assessment: cognitive bias.

One of my favorite forms of cognitive bias is “The Dunning-Kruger Effect”. In a nutshell it says that we are lousy at estimating our own competence. The theory started with lemon juice.

Justin Kruger and David Dunning became famous with a hilarious story of a young man and an attempted bank robbery. The would-be bank robber didn’t even make it into the bank. Employees of the bank described how he stumbled toward the front door clutching his face, walked face first into the reinforced glass door of the bank, and fell unconscious on the ground. The logical  assumption of this f-up would be a balaclava impairing his sight, but on the security camera’s he wasn’t wearing a mask or stocking. So, what happened?

The young criminal had smeared his own face in lemon juice, blinding him to the closed glass door in front of him, causing him to stumble blindly into the door and wake up several hours later in a police station. How could anyone be so stupid?  When Dunning interviewed the world’s worst bank robber, he met an angry young man who certainly didn’t see himself as being stupid. He had been told by someone (confident and convincing) that the lemon juice would create a reflection that would blur his face from security cameras. This intensely stupid theory was one that, even after his arrest, he still believed.

Dunning and Kruger’s conclusion was that incompetent individuals often lack the metacognitive insight to recognize their own incompetence. They started to run evaluations to see if they were onto something. They were. In multiple tests participants who had scored the lowest genuinely believed that their competence was above average. The cognitive bias went two ways though, participants who scored in the top percentiles underestimated their abilities assuming that the “easy” task would be easy for others as well. You can expand on this theory in two ways:

We are ignorant of our own ignorance. Or to put it another way we don’t know what we don’t know, and we fill this hole in our meta cognition with confidence.

We can’t monitor our own competence. We need someone else, external indicators, teachers, managers or mentors to calibrate our self-assessments or at least burst the bubble of confidence/ignorance.

Over the last 5 years the Dunning Kruger effect has been applied to other biases as well, and it kind of holds up.

West and Eaton (2019, Personality and individual differences) suggest that this concept extends to social attitudes like racial and gender prejudice. In their study of almost 400 participants they concluded that people who have a large prejudice towards a gender or race completely overestimate how egalitarian they are, believing themselves to be much less prejudiced than the average person. Unfortunately, the dual burden of ignorance can also be applied to social biases.

Things get interesting when neuroscientists start looking at this theory especially when it is framed with the idea of biased prior beliefs about their skill. So in Jansen, Rafferty, & Griffiths (2021, Nature Human Behaviour) – “A rational model of the Dunning–Kruger effect supports insensitivity to evidence in low performers.”  In this study they start with the assumption that most people think they are somewhat above average to start with, and another variant assumed that low-skill individuals are worse at monitoring their own accuracy on each question (a genuine metacognitive deficit). Findings suggest over- and under-estimators use differing cognitive processes when assessing their performance, under-estimators rely on recollection (I thought the test was great!) and over-estimators draw upon excess familiarity (I've done this a million times before) when over-estimating their performance. Episodic memory appears to play a contributory role in metacognitive judgments of illusory superiority and inferiority

This concept is expanded upon in Muller, Sirianni, & Addante (2021, European Journal of Neuroscience) – “Neural correlates of the Dunning–Kruger effect.” This innovative study explored the brain processes associated with Dunning–Kruger judgments. The question was whether differences in how people encode and recall information might be behind why some underestimate or overestimate their performance. Researchers had participants perform an episodic memory task (an item-recognition test) and periodically judge what percentile they believed they were in relative to others. The participants’ brain activity was recorded using electroencephalography (EEG). The sample was later divided into those who overestimated their performance (low scorers with inflated self-assessment) versus those who underestimated (high scorers with modest self-assessment), to compare their brain responses. Behaviourally, they replicated the classic Dunning–Kruger pattern: the bottom 25% performers overestimated their relative percentile rank, while the top performers tended to underestimate themselves. The intriguing results came from the EEG data. The brain activity of over-estimators versus under-estimators diverged in key ways when they were making self-assessments. Those who overestimated their performance showed EEG markers consistent with relying on familiarity (a mid-frontal brain signal called the FN400 around 400–600ms after stimulus), whereas those who underestimated showed a strong late parietal component (LPC) around 600–900ms, associated with detailed recollection of memories. In simpler terms, the high-performing individuals appeared to recall specific details of the task and their mistakes (activating recollection-related neural processes), making them more cautious and less confident in self-evaluation. Low performers, in contrast, might have based their self-evaluation on a vague sense of familiarity or fluency (e.g. “I feel like I did fine”), without recollecting the details that would reveal errors. Additionally, over-estimators had faster response times when confidently (and erroneously) rating themselves highly, suggesting perhaps a quick but shallow self-assessment, whereas under-estimators took longer when downplaying their performance, perhaps due to more analytic reflection.

So we can also add memory to the list of features of intelligence and it obviously plays a crucial role in cognitive bias. So how does all of this tie back into the concept of artificial intelligence? Well to be completely honest I don’t think it does. In fact it just seems to make the case that human intelligence is not even remotely close to machine intelligence. We are much much more complicated than any machine. Sure at a meta level, we both have bias, and if you are interacting with a language model you could be forgiven for projecting sentience onto the model…Since we don’t have a one stop definition for intelligence and assesing it is just asking for trouble, maybe we could see AI as a “plug-in” for human intelligence. An augmentation technology that when combined with us reaches its full potential. I’m convinced that we are just at the beginning of the AI revolution and that its going to change everything, but I also think we need to stop seeing AI as an entity and just reframe it as a tool.

Thanks for reading,

Kai