Chapter 16: How to Model Other Minds

Core Idea

From early mammals (brain under half a gram, seventy million years ago) to early apes (about 350 grams, ten million years ago), the brain grew almost a thousandfold. Most of that growth was scaling of the same old parts. Two regions were truly new: the granular prefrontal cortex (gPFC) and the primate sensory cortex (PSC). They are new because of what they connect to, not because of new circuitry. The gPFC gets most of its input from the older agranular prefrontal cortex (aPFC), so it builds a model of the mammal’s own simulation, which is a model of its own mind. The same regions light up when we think about ourselves and when we infer what others want and know. Bennett’s proposal follows social projection theory: primates model their own mind first and then use that model to simulate other minds.

Frameworks Introduced

• Scaled versus new brain regions (four kinds of growth):
  • Some structures scale with body size and keep their function. A bigger body has more nerves and needs more somatosensory cortex.
  • Some structures add neurons for better performance and keep the same function. Primate visual cortex is far larger than a rodent’s but does the same job.
  • Some areas are “fuzzy”: new hierarchical layers of the same kind of processing, such as visual areas that respond to faces.
  • Some regions are truly new, with unique connections and new functions. In primates these are the gPFC and the PSC.
  • Figure 16.2 marks the new areas: the granular prefrontal cortex at the front, and the superior temporal sulcus and temporoparietal junction (the PSC) along the side. They appear in early primates and in macaques and humans, and are absent in early mammals and rats. In the human brain they cover a much larger share.
  • Figure 16.1 shows the scale: mouse, macaque, chimpanzee and human brains side by side, from tiny to large.
• The layered model of the brain (first-order and second-order models) (Figure 16.3):

Layer	Structures	What it does
Bilaterian foundation	Reflexes	Hard-coded valence responses
Vertebrate addition	Basal ganglia, amygdala, hippocampus	Model-free learning through reinforcement from the reflexes
Mammalian addition	aPFC and sensory cortex	First-order model (intent). The aPFC tries to explain the basal ganglia, amygdala and hippocampus. The sensory cortex tries to explain external sensory input.
Primate addition	gPFC and STS & TPJ	Second-order model (mind). Each tries to explain the aPFC and the sensory cortex.

• When to use: to place any behavior at its level of explanation.
• Why it works: each layer learns a model of the layer below, so each one is “bootstrapped” on the others. The primate layer needs no new computation, only neocortex aimed at a new input.
• Social projection theory (simulation theory):
  • Claim: we understand others by imagining ourselves in their situation, with their knowledge and history. The idea goes back to Plato.
  • How: to explain why someone acted, imagine yourself in their place (“I know I yell more when I am stressed”). To predict them, imagine what you would do with their knowledge.
  • Failure mode: self and other get mixed up. A sad or happy mood makes you see that mood in others. Thirsty people think others are thirstier than they are. People project their own personality traits onto others.
• False-belief tests: the Sally-Ann test and its variants check whether you can hold someone else’s knowledge apart from your own. Humans pass by age four.

Key Concepts

• Granular prefrontal cortex (gPFC): frontal neocortex that evolved in early primates. It wraps around the older aPFC and gets almost no amygdala, hippocampal or direct sensory input.
• Primate sensory cortex (PSC): the author’s name for new primate sensory areas, including the superior temporal sulcus (STS) and temporoparietal junction (TPJ). It models the knowledge in the older sensory cortex.
• The “riddle” of the gPFC: patient K.M. lost a third of his frontal cortex in the 1940s with no loss in IQ. Damage to the gPFC looks mild until you test self-reference and theory of mind.
• Self-reference: in a 2001 fMRI study, both “How does this make you feel?” and “indoors or outdoors?” activated the aPFC, but only the feelings question activated the gPFC.
• Projecting yourself into simulations: in a 2015 study, gPFC patients described rich scenes but left themselves out. Hippocampus patients placed themselves in scenes but couldn’t build the surroundings.
• Mirror-sign syndrome: some gPFC patients no longer recognize themselves in a mirror.
• Metacognition: the ability to think about thinking.
• Mentalizing: building a model of one’s own inner simulation (see the summary at the end of Ch 18).

Mental Models

• Think of the aPFC as inventing “intent” to explain the amygdala and hippocampus, and the gPFC as inventing a “mind” to explain that intent.
• Ask why a primate in a maze turned left, and each layer gives a different answer:
  • Reflexes answer that there is a hard-coded rule to turn toward the smell.
  • Vertebrate structures answer that left maximizes predicted future reward.
  • Mammalian structures answer that left leads to food.
  • Primate structures answer, “I’m hungry, eating feels good when I am hungry, and to the best of my knowledge, going left leads to food.”
• Think of knowing as simulating. If a mammal’s simulation of a path shows water at the end, it “knows” the water is there. The PSC models why the simulation is that way (“I last saw water there”).
• Use the idea of “modeling yourself to model others” when you think about how to build theory of mind into AI.

Anti-patterns

• Treating gPFC damage as harmless: the deficits are real but specific. Patients do worse on false-belief tests, at recognizing and sharing others’ emotions, at telling lies from jokes, at spotting a faux pas, at taking another’s visual perspective, and at deceiving.
• Explaining primate smarts by size alone: most of the growth was scaling, but the new abilities came from new regions.
• Calling current AI “theory of mind”: video action classifiers, ad platforms that predict purchases, and face-emotion classifiers are trained on labeled examples. Bennett calls them “a far cry” from primate theory of mind.

Rules of Thumb (author’s wording)

• “Reasoning about your own mind and reasoning about other minds is, in the brain, the same process.”
• “We understand others by imagining ourselves in their shoes.”
• If we want true humanlike AI systems, “theory of mind will undeniably be an essential component of that system.”
• Theory of mind was born “from the subtler and far more cutting dangers of politics.”

Worked Example

The Sally-Ann false-belief test (Figure 16.5, from Frith, 2003).

1. Sally has a basket; Ann has a box.
2. Sally puts a marble in her basket.
3. Sally goes out for a walk.
4. While she is gone, Ann moves the marble from the basket into the box.
5. Sally comes back. The question: where will Sally look for her marble?
6. The right answer is the basket. To give it, you must hold Sally’s knowledge (marble in basket) apart from your own (marble in box).
7. Humans pass by age four. In fMRI, the gPFC and PSC activate during false-belief tests, and performance correlates with the activation. Patients with gPFC damage do worse. Monkeys show the same pattern: the gPFC activates when they reason about another’s intent or knowledge, and damage to it impairs them.

The comic-strip task (Eric Brunet-Gouet, 2000, PET scan) tests intent rather than knowledge. Each strip has three frames and three possible endings (Figure 16.4). In strip 1, a man struggles with a barred window, then stands by a bed. The answer is B, where he sits on the bed holding a rope, because he wants to escape through the window. In strip 2, a short boy can’t reach a door handle and takes an umbrella from a stand. The answer is C, where he hooks the handle with the umbrella, because he wants to open the door. Strip 3 needs only physics: a man steps on a toy car, and the answer is C, where he falls down. The gPFC lit up only for the strips about intent.

Development follows the same order. Children pass the mirror self-recognition test at about two and begin to use words like “want”, “wish” and “pretend”. At about three they can say “I thought it was an alligator. Now I know it’s a crocodile.” Only at about four or five do they pass false-belief tests about other people. Chimpanzees raised in social isolation can’t recognize themselves in a mirror.

Key Takeaways

1. Most of the primate brain is a scaled-up mammal brain; the gPFC and PSC are the truly new parts.
2. These areas are new because of their connections. The gPFC models the aPFC, so it builds a second-order model, a model of one’s own mind.
3. The same areas handle thinking about yourself and inferring others’ intent and knowledge.
4. Larger gPFC predicts larger social networks and higher rank in primates, and in humans a thicker gPFC predicts a larger social network and better theory-of-mind scores.
5. Theory of mind likely works by modeling yourself and then projecting that model onto others, which is why self and other get mixed up.
6. AI systems that live alongside us, and especially superintelligent ones, need theory of mind to infer what we mean.

Connects To

• Ch 10–14 (Breakthrough #3, Simulating): the aPFC, hippocampus and sensory cortex build the first-order simulation that the gPFC models. aPFC damage causes akinetic mutism.
• Ch 5–9 (Breakthrough #2, Reinforcing): the basal ganglia’s temporal difference learning is the model-free layer in the stack.
• Ch 15: the political pressures that made theory of mind worth having.
• Ch 17 and Ch 18: theory of mind repurposed for imitation learning and for anticipating future needs.
• Ch 19–22 (Breakthrough #5, Speaking): why AI systems need theory of mind to understand requests.