What Would Need to Change for AI to Understand the World?
An AI model can describe a bicycle, explain how its gears work and suggest how to repair a loose chain.
But does that mean it understands a bicycle in the same way as someone who has ridden one, repaired one and fallen from one?
A five-part series about what may shape future AI systems and why confident forecasts often fail.
Questions about whether AI understands the world often produce two extreme answers.
One side points to fluent explanations and successful problem-solving as evidence of understanding.
The other side says the model is only predicting patterns and therefore understands nothing.
Both positions can hide an important difficulty: understanding is not one clearly defined ability.
A system may recognize objects, predict events or explain relationships without possessing human experience. It may perform well in one kind of reasoning while failing when the situation changes slightly.
More dependable world understanding may require models to connect symbols to observations, track causes and effects, preserve information over time and test predictions through interaction.
What does “understanding” mean?
People use the word understanding to describe several different abilities.
These can include:
- recognizing what something is
- explaining how it works
- predicting what will happen next
- knowing how it changes under different conditions
- using it correctly in a new situation
- connecting it to goals, consequences and experience
An AI system may show some of these abilities without showing all of them.
For example, a model may explain that glass can shatter when dropped. It may still make mistakes when asked to reason about an unusual container, surface or sequence of physical events.
This is why a single conversation cannot settle whether a model “really understands.” The more useful question is which kinds of representation and reasoning the system can perform reliably.
Language contains a surprising amount of world information
Language is not disconnected from reality.
Books, instructions, conversations and reports contain descriptions of objects, actions, causes, emotions and physical events.
By learning patterns across large amounts of language, a model can form internal representations that capture many useful relationships.
It can learn that bicycles have wheels, that chains transfer motion and that turning the handlebars changes direction.
Modern multimodal models can also learn from images, audio and video. This gives them information beyond relationships between words alone.
However, descriptions and recordings are still not identical to direct interaction. The model may learn what usually happens without having a reliable way to test whether its prediction is correct in a particular situation.
Grounding connects symbols to observations
Grounding is the process of connecting symbols such as words to things, properties or events outside the text itself.
The word “hot” can be connected to temperature readings, visual signs of steam, warnings, physical damage or human reactions.
A grounded system does not rely only on how the word appears in sentences. It can connect the word to additional evidence.
Grounding may come from:
- images and video
- audio
- sensor readings
- software states
- robotic interaction
- simulated environments
- feedback from people
More input does not automatically create perfect understanding. The model still has to learn which observations correspond to which concepts and when those relationships change.
Causal reasoning asks more than what usually appears together
Statistical patterns can reveal that two events often occur together.
Causal reasoning asks whether changing one event would change the other.
Suppose wet streets and umbrellas often appear together. A model may learn the association. But umbrellas do not cause streets to become wet. Rain can cause both.
Understanding that difference matters when predicting unfamiliar situations.
A system with stronger causal models would need to reason about questions such as:
- What caused this outcome?
- What would happen if one condition changed?
- Which event is only correlated with the result?
- Would the same explanation hold in a new environment?
Current models can answer many causal questions, especially when similar explanations appear in their training data. They can still confuse a familiar verbal pattern with a dependable model of how the situation works.
A world model must track changing states
The world does not remain fixed while a conversation continues.
Objects move. People change their plans. A container that was full may become empty. A document may be edited after the model first reads it.
A system operating over time needs to track those changes.
This is sometimes described as maintaining a world model: an internal representation of relevant objects, relationships and states.
For a simple task, the system might need to remember:
- which file is currently open
- which version has been edited
- what action was completed
- what remains unfinished
- what changed after the previous action
Without stable state tracking, the model may respond as though an earlier version of the situation still exists.
Memory provides continuity
A model’s context window allows it to process information supplied during a current interaction. That is not the same as having dependable long-term memory.
A future system may need to preserve selected information across many sessions while also knowing what should be forgotten, updated or treated as uncertain.
Useful memory requires more than storage.
The system must decide:
- what is important enough to save
- which memory is relevant now
- whether the saved information is still current
- whether two memories conflict
- who is allowed to access them
A large pile of remembered text can confuse a model as easily as it can help. Reliable memory needs organization, retrieval and correction.
A person can learn a great deal about lemons from books, photographs and chemical measurements.
Tasting a lemon adds another kind of evidence. It connects the description to direct feedback. The experience does not replace the earlier knowledge, but it grounds part of that knowledge in interaction.
Interaction allows predictions to be tested
A model trained on fixed data mainly learns from examples collected before the interaction begins.
An interactive system can make a prediction, take an action and observe what actually happens.
This creates a feedback loop:
If the system predicts that an action will open a file but the file remains closed, it receives evidence that its plan or its model of the interface was wrong.
The difficult part is learning from that evidence without making unsafe or expensive mistakes.
Robots, simulations and controlled software environments can provide opportunities for interaction, but each environment teaches only part of the world.
Would embodiment solve the problem?
Embodiment means giving an AI system a body or another way to act within an environment.
A robot can receive physical feedback about distance, resistance, balance and motion. This may help connect concepts to consequences that are difficult to learn from text alone.
However, embodiment is not a guaranteed route to human-like understanding.
A robot can collect sensor data without organizing it into useful concepts. It can also learn patterns that work in one environment but fail elsewhere.
Simulated environments may provide large amounts of practice, yet simulations always simplify some parts of reality.
Embodiment is therefore one possible ingredient, not a complete answer.
Human understanding is not the only possible standard
AI systems may develop useful forms of representation that do not resemble human experience.
A weather model can track atmospheric relationships without feeling wind. A vision system can detect patterns outside normal human perception. A scientific model may compare more variables than a person can hold in mind at once.
The important question may not be whether a machine understands exactly as a person does.
It may be whether the system’s internal representations are accurate enough to support dependable prediction, explanation and action.
What would need to improve?
More reliable world understanding may depend on several connected advances:
- Grounding: connecting symbols to observations and actions
- Causal reasoning: distinguishing causes from associations
- State tracking: representing how situations change
- Memory: preserving and updating relevant information
- Interaction: testing predictions against outcomes
- Transfer: applying knowledge in unfamiliar situations
- Uncertainty: recognizing when the internal model may be wrong
None of these alone provides a final definition of understanding.
Together, they offer a more practical way to evaluate progress than asking whether a fluent answer proves that the system thinks like a person.
AI understanding is not an all-or-nothing property. A better test is whether the system can connect information to observations, track change, reason about causes and correct its predictions when reality disagrees.
Even without human-like understanding, AI may become more useful by moving beyond the familiar chat window.
Comments
Post a Comment