Commonsense reasoning is the capacity to draw plausible conclusions from the implicit everyday knowledge that humans share about objects, people, time, space and cause and effect, even when that knowledge is never written down. Among the AI fundamentals of artificial intelligence, it names one of the oldest open problems: getting machines to know that a glass dropped on a tile floor will probably break, or that a person who is asleep cannot answer the phone.
The difficulty is not the logic itself but the sheer breadth of unstated assumptions behind ordinary situations. Commonsense reasoning (also written common-sense reasoning) therefore sits at the crossing of knowledge representation, probabilistic inference and modern language models, and it remains a central test of how far current systems are from robust, general understanding.
What commonsense knowledge covers
Commonsense knowledge is the background understanding of the world that an average adult applies without reflection. Researchers usually divide it into several knowledge domains, because each one requires different representations and fails in different ways when a system lacks it.
| Domain | What it captures | Typical inference |
|---|---|---|
| Naive physics | Gravity, solidity, containment, liquids | An unsupported cup falls |
| Temporal reasoning | Order, duration, typical schedules | Breakfast comes before lunch |
| Spatial reasoning | Location, size, relative position | A car does not fit in a drawer |
| Naive psychology | Beliefs, goals, emotions of agents | A person who lost a wallet is upset |
| Social norms | Etiquette, roles, obligations | Guests usually thank the host |
| Causal knowledge | Typical causes and effects | Rain makes streets wet |
These domains overlap constantly. Understanding the sentence “She put the ice cream in the oven by mistake” requires physics (heat melts ice cream), psychology (the action was unintended) and causal expectation (the dessert is now ruined). Human readers combine all three instantly, which is precisely what makes the problem hard to engineer.
Why commonsense reasoning is difficult for machines
The core obstacle is that commonsense facts are rarely stated in text because every human reader already knows them. This reporting bias means that corpora mention unusual events far more often than normal ones, so a system trained on text may learn that people are murdered more often than they blink.
A second obstacle is defeasibility: commonsense conclusions hold by default but can be overturned by new facts. Birds fly, yet penguins, injured birds and roasted chickens do not. Classical logic is monotonic, meaning added premises never retract a conclusion, so early researchers developed non-monotonic logics, default reasoning and circumscription to model this behavior.
- The frame problem: specifying what does not change after an action without listing every unaffected fact.
- The qualification problem: an action has endless possible preconditions, such as a car starting only if the tank is not empty and the exhaust is not blocked.
- The ramification problem: an action has indirect side effects that must be inferred, not listed.
- Context sensitivity: the same fact can be true in one setting and false in another.
These problems explain why commonsense reasoning is often described as AI-complete: solving it fully would likely require solving general intelligence, and it is closely tied to the symbol grounding problem and to Moravec’s paradox, the observation that skills easy for toddlers are hard for computers.
Approaches to building commonsense into AI systems
Three broad families of methods have been used to give machines commonsense, and each reflects a different era of the field. The comparison below shows how symbolic knowledge bases, statistical learning and hybrid neuro-symbolic systems trade coverage against reliability.
| Approach | Representative work | Strength | Main weakness |
|---|---|---|---|
| Hand-built knowledge bases | Cyc, formal axiom sets | Precise, inspectable inferences | Very costly, brittle coverage |
| Crowdsourced knowledge graphs | ConceptNet, ATOMIC | Broad, cheap relational facts | Noisy, shallow, context-free |
| Pretrained language models | Large transformer models | Wide coverage, fluent output | Inconsistent, hard to verify |
| Neuro-symbolic hybrids | Models plus graphs or solvers | Combines recall and rigor | Complex integration |
The Cyc project, started in 1984, attempted to encode millions of hand-written assertions in a formal language and showed how expensive explicit encoding is. Knowledge graphs such as ConceptNet store relations like “knife UsedFor cutting”, while ATOMIC focuses on if-then knowledge about events, intentions and reactions.
Since about 2019, large language models have absorbed a great deal of implicit commonsense from massive text corpora, and techniques such as chain-of-thought prompting expose intermediate reasoning steps. The result is far better benchmark performance, but the knowledge remains statistical rather than structured, which leads to confident errors on unusual phrasings.
How commonsense reasoning is evaluated
Commonsense reasoning is measured with purpose-built benchmarks that are trivial for people but designed to defeat superficial pattern matching. The table summarizes widely used datasets and the abilities they target.
| Benchmark | Task format | Skill tested |
|---|---|---|
| Winograd Schema Challenge | Pronoun resolution pairs | World knowledge behind reference |
| WinoGrande | Large adversarial Winograd set | Robustness to dataset artifacts |
| HellaSwag | Choose the plausible continuation | Everyday event prediction |
| PIQA | Choose a physical solution | Naive physics |
| Social IQa | Questions about motives and reactions | Naive psychology |
| CommonsenseQA | Multiple-choice questions | Concept relations |
Benchmark scores must be interpreted with care. Many datasets contain annotation artifacts that models exploit, and leading models now approach human accuracy on several of them. High scores therefore show strong pattern coverage, but they do not prove robust understanding, so researchers increasingly use adversarial and out-of-distribution tests.
Limitations and common misconceptions
The most frequent misconception is that fluent language implies sound commonsense. A model can describe a kitchen accurately yet claim that a heavy pan floats when a question is phrased unusually. The table below pairs typical misconceptions with what practitioners actually observe.
| Misconception | Reality | Practical consequence |
|---|---|---|
| Fluent text means understanding | Fluency and reasoning are separable | Verify critical inferences |
| Benchmarks prove commonsense | Scores can reflect dataset artifacts | Test on real task data |
| More data solves it fully | Reporting bias persists at scale | Add structured knowledge |
| Commonsense is universal | Norms vary across cultures | Check cultural assumptions |
Commonsense is also culturally situated: norms about meals, gestures or family roles differ between societies, so a model trained mainly on one language community may treat local conventions as universal truths. This is both an accuracy problem and a fairness concern in deployed systems.
The simple explanation
Commonsense reasoning is the everyday good sense that people use without thinking. You know that hot coffee can burn your mouth, that a closed door must be opened before you walk through it, and that a friend who is crying is probably sad. Nobody teaches these facts in school because they seem too obvious to say.
Computers do not get this knowledge for free. They only know what they are given or what they can pick up from data, and data rarely spells out the obvious. Teaching machines the obvious turns out to be one of the hardest jobs in artificial intelligence, and even today’s best systems still make surprising mistakes that a child would avoid.
How it is used in practice
A team building a household assistance robot for a care home must make sure the robot handles everyday requests safely. When a resident asks it to “bring a glass of water and put it on the table”, the robot has to infer unstated constraints: the glass should be upright, not filled to the brim, and not placed on the edge where it could fall.
The engineers combine a language model that interprets the request with a structured safety knowledge base containing rules about liquids, fragile objects and people with limited mobility. Before acting, the planner checks each proposed step against these rules. In testing with 500 scripted requests, this hybrid design caught unsafe plans that the language model alone had proposed, such as placing a hot drink within reach of a resident with tremors, and the team kept human review for any action flagged as uncertain.