“Intelligence” is one of the most common words in artificial intelligence, and that is precisely why it can hide several different questions inside one label.
A system can reason, plan, learn, adapt, use tools, and change its behavior in new environments. We may describe all of that as “intelligent.” That is useful in ordinary language, but a theory needs to ask more precisely:
What have we actually observed when we attribute intelligence to a system, and which lower-level causal processes produced those observations?
Connection Theory does not treat intelligence as a mysterious substance located inside one node. It treats intelligence as an observer-level attribution made from a system's integrated behavior across time and conditions. The scientific task is then to explain which organization and dynamics make that behavioral profile possible.
“Intelligence” compresses several different phenomena
When we call a person or an AI intelligent, we may be referring to several things at once:
- stored capability;
- reasoning on the current problem;
- learning from new experience;
- adaptation after the environment changes;
- transfer across situations;
- planning and tool use;
- stable performance across repeated conditions.
These are related, but they are not the same variable.
A model can perform extremely well on the first attempt and still fail to change after new consequences. Another system may begin with weaker performance but update rapidly after experience. A single intelligence score can hide that difference.
So the useful strategy is to unpack the high-level label before trying to explain it.
AI, intelligence, and AGI are not terms at the same analytical level
The industry term AI can continue to name a technological field and class of systems.
Intelligence is a theoretical attribution that should be justified from behavior and causal evidence.
AGI can likewise remain a conventional name for broad, robust, transferable artificial capability. But calling something AGI should not imply that a primitive substance called “general intelligence” has been found inside one model.
Being called AI does not by itself establish that a system satisfies a particular scientific definition of intelligence.
The labels can remain useful. The explanatory levels underneath them still need to be separated.
Read: AI, Intelligence, and AGI
Intelligence is not located in an isolated component
A neuron can contribute to intelligent behavior without containing a whole mind. A model parameter can influence an output without containing a complete reasoning process. A highly capable employee can be crucial to a company without being identical to the company's capability.
A more useful explanatory chain is:
The same components can generate very different system-level behavior when connection, timing, state, feedback, and coordination differ.
So “Where is intelligence?” is often the wrong first question. A better question is:
Which organization, operating in which environment, generates the dynamics that produce the behavior we later describe as intelligent?
Strong capability is not the same as adaptation
A system can be extremely capable while remaining unchanged throughout an evaluation.
A well-trained model that scores highly on a benchmark demonstrates substantial stored capability. But a separate question remains: when new interactions produce new consequences, can those consequences change what the system does next?
Capability describes what the system can do now. Adaptation describes how the system changes because of its relation to an environment.
The distinction is practical. Solving a novel problem on the first attempt and changing strategy after the first failure are different abilities.
Adaptation is not identical to intelligence either
The converse is also important.
A very simple system can undergo a narrow adaptive change without supporting a broad attribution of intelligence. Adjusting one parameter after one local consequence may demonstrate adaptation, but it does not by itself establish wide, robust, transferable capability.
The current framework therefore separates several levels:
- adaptation — consequence-sensitive change at the mechanism level;
- adaptive behaviour — behavior generated by such change;
- intelligence — a higher-level attribution made from an integrated behavioral profile across time and conditions.
A compact chain is:
Causal Organization → Adaptive Dynamics → Structured Behaviour in the Actual Environment → Repeated Evidence Across Time and Conditions → Integrated Behavioural Profile → Observer-Level Attribution: Intelligence
Each step can be tested separately. That is more informative than placing every relevant process inside the single word intelligence.
Why time matters
A snapshot easily confuses stored capability with active adaptation.
If a system answers correctly the first time, it may be using capability acquired long before the current interaction. If it answers incorrectly and then changes after informative consequences, a different property has become visible: experience is entering the future.
This is why an intelligence assessment based only on one static performance can miss an important dimension—whether the system is changed by what happens to it.
But this should not be turned into a stronger claim than the evidence supports. An adaptive system does not need to predict every long-term consequence, and a second attempt need not succeed. The more realistic requirement is that credible new evidence can alter later state and action.
That distinction is developed in Adaptation Is Not Foresight.
Why this matters for AI research
Once the levels are separated, more discriminating questions become possible:
- Does strong performance come from stored capability or from adaptation during interaction?
- Did an experience change later behavior, or did it remain temporary context only?
- Can a change persist across interaction cycles?
- Does previously stored capability remain applicable after the environment changes?
- Does the decisive capability belong to the declared system, or is it imported from tools, hidden prompts, or human operators outside the boundary?
- On a matched task, does an organized system show system-level capability that a simpler single-node explanation cannot account for?
These questions are closer to explanation than one aggregate score.
The current Connection Theory position can therefore be summarized simply:
“Intelligence” can remain a useful macroscopic behavioral description. The scientific explanation lies in the causal organization, retained capability, and adaptive dynamics that generate the behavior supporting that description.