Evolutionary Blueprint (7). Five ways to think about intelligence, beyond fluent answers

Series: Evolutionary Blueprint of AI, article 7. Max Bennett's five breakthroughs offer a way to compare brains and AI. Use them to ask better questions about action, feedback, planning, social assumptions and language, without treating evolution as a recipe for machines.

Evolutionary Blueprint (7). Five ways to think about intelligence, beyond fluent answers
Illustration generated using ChatGPT. The connected atlas from steering to speaking.

Who should read it?

Interested general readers, product leaders and AI practitioners.

Metapher

The connected atlas

Intro

A customer asks an AI assistant to resolve a delayed delivery. In this fictional case, the tracking feed shows the parcel is still on a truck. The assistant has not checked the customer's deadline or the refund policy, yet it writes a reassuring reply, proposes a refund above its authorization limit and submits a request to close the case. The refund is an offer in the draft, not a payment; the closure request is an action, even if the case system rejects it.

The customer later explains that the goods are needed for an event tomorrow. That deadline was missing from the assistant's initial information, not something it could infer from the tracking feed. A fluent answer is a poor substitute for checking the evidence and the system's authority.

The Evolutionary Blueprint of Artificial Intelligence, Max Bennett's A Brief History of Intelligence (2023) offers a framework for thinking about such differences. Bennett organizes an account of the human brain's evolutionary history around five proposed breakthroughs:

  • Steering
  • Reinforcing
  • Simulating
  • Mentalizing
  • Speaking

His 2021 research paper calls the account a 1st approximation:

  • A model that groups many anatomical and behavioral changes under a small number of ideas.

The 5 categories are useful questions, not 5 boxes into which every animal or machine must fit. Bennett traces one lineage leading to humans. Evolution had no human destination. Other lineages have their own histories, and similar abilities can arise in different ways. His proposed timings, mechanisms and dependencies are hypotheses, not a settled timetable for all intelligence. Nor do they establish a required construction sequence for AI.

Steering: choosing a direction before explaining it

Bennett's first category starts with organisms that must act in a world: moving toward a favorable condition or away from a harmful one. In his account, early bilaterians, ancestors of a broad group of animals with roughly symmetrical body plans, developed neural organization that helped integrate signals and guide movement. His proposal concerns the history of our lineage. It does not mean earlier animals had no nervous systems or that every response was a conscious choice.

Imagine an animal encountering a chemical gradient associated with food. Its movement changes as the concentration changes. The animal does not need a verbal description of food to act on the signal. The distinction is between describing an objective and having a mechanism that connects information to behavior.

For an engineered system, steering raises a practical design question:

  • “What observations can change its actions, and which outcomes or constraints guide those choices?"

A warehouse robot, for instance, needs location and obstacle information connected to movement commands. A text assistant may need a different action loop: check the current order, request clarification if information is missing, and propose an authorized next step.

The analogy has limits. A software objective is specified by designers and operators; it does not tell us that the system has biological needs or experiences pain. In business terms, the check is operational: list the inputs, actions and prohibited actions. Then test what happens when a sensor fails, an order record is stale, or a request is ambiguous. A well-worded answer cannot repair an action loop that never checks the world.

Reinforcing: learning when outcomes surprise

Steering describes what directs behavior. Reinforcing adds a way to change behavior using consequences. Bennett associates this breakthrough with early vertebrates and proposes a form of model-free reinforcement learning:

  • Learning which actions are rewarding in particular situations without necessarily rehearsing a detailed chain of possible future states

That timing and biological implementation are parts of his model, not direct observations of an ancestral brain.

A simple example is a delivery-routing system that tries different routes and updates its preference after observing travel times. The system may learn that a certain road usually works well at a certain hour. It can improve that preference without explicitly simulating every intersection along every route.

There is an empirical reason to compare some learning algorithms with biology. In their 1997 paper, Wolfram Schultz, Peter Dayan and P. Read Montague describe evidence that activity in some primate dopamine neurons tracks changes or errors in predictions of rewarding events. If a reward is better than expected, the signal can differ from the signal when the same reward was already predicted. This is a useful connection to reward-prediction error in reinforcement learning, not proof that dopamine has only one function or that a software algorithm reproduces a brain.

For a product team, the difficult question is often what counts as a reward. If the delivery assistant is rewarded only for closing tickets quickly, it can look successful while customers reopen unresolved cases. A more useful evaluation would also inspect resolution quality and reversals. Even a carefully chosen measure can miss harms, so a team needs independent checks and an owner who can change or stop the system. Feedback is valuable only when the feedback measures something worth improving.

Simulating: testing a possible action before taking it

In Bennett's account, simulating is the ability to use an internal model to consider possible sequences of actions and their outcomes. He associates it with early mammals and the neocortex. That does not mean other animals cannot plan.

The difference from the route preference above is that a system might represent what happens after each turn, compare several possible routes, and select one before driving. A simulation need not be a vivid mental movie. For an AI system, it can be a calculation of possible states and actions within a defined model.

This distinction has a working engineering example. In David Silver and colleagues' 2016 AlphaGo paper, the program combined neural networks with search and learned from expert games and self-play. Its search considered possible continuations in the rules of Go. That does not mean AlphaGo had a mammalian brain; it shows that AI research has long used learning and simulated futures alongside other techniques.

The biological picture is also wider than a simple mammal-only label. In Raby and colleagues' 2007 experiment, western scrub-jays stored food in ways consistent with preparing for conditions they would encounter the next morning. The authors interpret the results as evidence of planning for a future need. The experiment does not reveal every detail of the birds' internal process, but it is a useful warning against treating a capability as the exclusive property of one branch of animals.

For the delivery assistant, a simulated scenario might ask what happens if the parcel is rerouted, replaced or refunded. Its usefulness depends on the model:

  • "Does it know stock levels, cutoff times and carrier restrictions?"

A plausible story built on incorrect assumptions may be worse than no scenario at all. Before acting on a simulation, compare its predictions with outcomes on held-out historical cases and include cases where conditions changed. If the model cannot represent a relevant constraint, the decision belongs with a person or a better-suited system.

Mentalizing: keeping another person's knowledge separate

Mentalizing means reasoning about another person's beliefs, knowledge, desires or intentions. These are called mental states: for example, a customer may believe a parcel is already at a depot while an employee knows it is still on a truck. Bennett places this theme in his account of primate evolution and also applies it to anticipating one's own future needs and learning by observing others. Those are his proposed connections, not a claim that all social understanding is unique to primates.

Consider a customer who knows the package is delayed but has not seen the updated tracking estimate. An employee who has seen that estimate must distinguish their own knowledge from the customer's. Otherwise, a reply such as "as you can see, it will arrive Friday" can be confusing or misleading. Mentalizing is the kind of inference involved in tracking that difference, not a guarantee that the inference is correct.

Animal experiments illustrate both the appeal and the difficulty of the concept. Christopher Krupenye and colleagues (2016) used an eye-tracking task in which great apes watched an actor who had an outdated view of an object's location. The apes' anticipatory looking was consistent with expecting the actor to search where the actor falsely believed the object was. The researchers interpreted this as evidence for an implicit understanding of false belief; the behavior remains something researchers must interpret, rather than direct access to the animals' thoughts.

A language model can also answer questions about who knows what, but a successful answer does not by itself establish a human-like mental process. In a 2023 preprint, Tomer Ullman found that small changes to certain theory-of-mind tasks altered the performance of the tested language models. That result is about those models and tests, not a timeless verdict on all AI systems.

For a customer-facing system, turn the idea into testable requirements. Give it cases where the customer, employee and system have different information. Check whether it asks rather than invents a preference, corrects itself when given new evidence, and offers a human handoff. A model that infers a preference accurately can still pursue the wrong goal; recognizing what someone wants is not the same as respecting it. Permission checks, review and accountability remain necessary.

Speaking: communicating a model is not sharing one

Bennett's final category concerns a human capacity for structured communication. His paper uses "speaking" more broadly than vocal speech and proposes connections between language, music and earlier social cognition. Human language includes signed languages, while other animals communicate in many ways.

Language lets a person describe a situation another person has not witnessed:

  • "The parcel is on the truck, but the event is tomorrow”

The listener can use that description with their own knowledge to decide what to do. Yet the words do not transfer a whole mental model. The recipient can miss a constraint, interpret a term differently or lack the background needed to act.

Language models make this gap visible. They can draft explanations, translate requests into structured fields and help users retrieve information. A well-formed message, however, does not verify a delivery date or prove that an offer is authorized. Some AI systems also use tools, images, memory or action interfaces. Evaluating only their generated text misses important parts of the system.

For the delivery workflow, a useful design separates the model's draft from the evidence behind it. Display the tracking source and its update time, identify the policy used to propose a refund, and prevent a commitment when the necessary information is absent. Have a responsible employee review exceptions. These are product choices, not conclusions that follow automatically from a theory of language.

Use the framework as a set of tests, not a maturity score

Control systems act on sensor feedback. Reinforcement-learning systems learn from rewards. Planning systems use models and search; and researchers have built computational approaches to social reasoning. A particular language-model application may have weak or absent capabilities in one of these areas, but that is a question to investigate, not a verdict on AI as a whole.

The 5 categories can guide a review of the complete delivery assistant:

  1. Steering:
    1. "Which live information can it observe?"
    2. "Which actions can it take, and where are the hard limits?"
  2. Reinforcing:
    1. "Does it change based on feedback?"
    2. "If so, what is measured, and can a good score conceal an unresolved problem?"
  3. Simulating:
    1. "Does it compare consequences before acting?"
    2. "Are the scenarios tied to current data and tested against outcomes?"
  4. Mentalizing:
    1. "How does it handle different people's knowledge and preferences?"
    2. "Can users correct its assumptions?"
  5. Speaking:
    1. "Is its explanation supported by evidence, and can a user tell the difference between a verified fact and a proposal?"

The fictional team's review ends with 3 findings and corresponding design choices.

  1. Tracking showed the parcel on a truck but did not establish when the customer needed it, so the assistant must ask for the deadline rather than infer it.
  2. Its proposed refund exceeded the policy limit, so the interface labels the draft as a proposal and routes exceptions to a support employee. The agent cannot issue refunds.
  3. A closure request could hide an unresolved delivery, so the case service rejects closure while delivery is unconfirmed and sends failed requests to a person.

The team tests these rules with stale tracking and missing policy data before use. These are decisions in a teaching case, not observed deployment results.

Bennett's framework makes a broad evolutionary story easier to think with, while comparative research reminds us that the story is more complicated than a single ladder. It cannot tell a company where to invest, establish that a machine understands, or predict what artificial general intelligence will require. Those conclusions need their own evidence. For a team deploying an assistant, the next useful step is narrower:

  • Identify the decision it will influence and test what happens when its confident answer is wrong.

Previous

Article 6: Understanding people is not enough for AI safety

Next

Article 8: What does an AI explore, when curiosity becomes a reward?