# Commonsense knowledge

Commonsense knowledge refers to the basic, practical understanding of everyday situations and physical world properties that humans acquire through experience, which AI systems often lack and seek to model.

Commonsense knowledge is the collection of basic, practical facts about the world that most people take for granted, such as knowing that water is wet, that objects fall when dropped, or that people can feel hunger. It underpins everyday reasoning and communication, allowing humans to make sensible inferences without explicit instruction. In artificial intelligence, commonsense knowledge has long been recognized as a major challenge because it is rarely stated explicitly in text or data, yet it is essential for robust natural language understanding, planning, and interaction with physical environments.

Unlike specialized expertise, commonsense knowledge is broad, cross-cultural, and often implicit. It includes physical intuitions (e.g., solid objects cannot pass through each other), social norms (e.g., people usually respond when greeted), and temporal reasoning (e.g., events happen in sequence). AI systems, particularly those based on [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) and [deep-learning](https://www.wikiprompt.org/wiki/deep-learning), typically learn from large datasets that may not contain such obvious truths, leading to gaps in their ability to handle novel or ambiguous situations.

## Historical Approaches

Early efforts to formalize commonsense knowledge in AI date back to the 1970s and 1980s, with researchers like John McCarthy advocating for logical formalizations of common sense. Projects such as Cyc, initiated by Douglas Lenat in 1984, aimed to manually encode millions of commonsense facts into a machine-readable ontology. While Cyc demonstrated the feasibility of explicit knowledge representation, its scalability and maintenance proved difficult. Concurrently, frame-based systems and scripts, such as Roger Schank's work on conceptual dependency, attempted to model everyday scenarios like eating at a restaurant.

These symbolic approaches faced criticism for being brittle and incomplete, as they could not easily handle the vast, nuanced, and context-dependent nature of commonsense. By the 1990s, statistical methods gained traction, but they initially focused on narrower tasks like part-of-speech tagging and parsing, leaving commonsense largely unaddressed.

## Modern Machine Learning and Neural Models

The rise of [neural-network](https://www.wikiprompt.org/wiki/neural-network) architectures, particularly [transformer](https://www.wikiprompt.org/wiki/transformer)-based models, shifted the focus toward learning commonsense implicitly from text. Large-scale language models, such as those developed by [openai](https://www.wikiprompt.org/wiki/openai) and [anthropic](https://www.wikiprompt.org/wiki/anthropic), are trained on billions of words from the internet, which includes some commonsense statements but also much noise. These models can answer simple commonsense questions, such as "What do you use to cut bread?" with high accuracy, but they often fail on adversarial or counterfactual examples, revealing that their knowledge is statistical rather than grounded in physical experience.

Researchers have introduced benchmarks like CommonsenseQA and Physical Interaction QA to evaluate these abilities. For instance, a model might be asked, "If you drop a glass on a hard floor, what happens?" and must choose between "it breaks" and "it floats." While modern models perform well on such tests, they remain susceptible to subtle errors, such as confusing spatial relations or temporal orderings. This limitation is particularly evident in [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)s when they generate plausible but factually incorrect responses, a phenomenon often attributed to their lack of a robust world model.

## Knowledge Graphs and Hybrid Systems

To complement neural approaches, some researchers have built commonsense knowledge graphs, such as ConceptNet, which aggregates facts from crowdsourcing and existing resources. These graphs contain nodes for concepts and edges for relations like "used for," "capable of," and "located at." Hybrid systems combine these symbolic structures with neural models, allowing the model to retrieve relevant facts during inference. For example, a question-answering system might query ConceptNet to verify that a knife is used for cutting before generating an answer.

Companies like [google-deepmind](https://www.wikiprompt.org/wiki/google-deepmind) and academic labs such as [mit-csail](https://www.wikiprompt.org/wiki/mit-csail) and [stanford-ai-lab](https://www.wikiprompt.org/wiki/stanford-ai-lab) have explored integrating commonsense into reinforcement learning agents, where the agent must reason about object affordances and physical dynamics. These efforts often use simulation environments to teach agents commonsense through interaction, rather than relying solely on static data. However, bridging the gap between symbolic knowledge and continuous perception remains an open problem.

## Challenges and Limitations

One major challenge is the "knowledge acquisition bottleneck": manually encoding commonsense is labor-intensive, while automatic extraction from text is noisy and incomplete. Another issue is cultural and contextual variability; what is common sense in one community may not hold in another, making universal models difficult. Additionally, commonsense reasoning often requires integrating multiple modalities, such as vision and language, which current models handle only partially.

Evaluation is also problematic. Many benchmarks are saturable, meaning models can achieve high scores by memorizing patterns rather than demonstrating genuine understanding. Researchers like [melanie-mitchell](https://www.wikiprompt.org/wiki/melanie-mitchell) and [joshua-tenenbaum](https://www.wikiprompt.org/wiki/joshua-tenenbaum) have argued for more rigorous tests that probe causal reasoning and counterfactual thinking. For instance, asking "What would happen if gravity were weaker?" requires flexible mental simulation, which most AI systems lack.

## Future Directions

Emerging approaches include using [curriculum-learning](https://www.wikiprompt.org/wiki/curriculum-learning) to train models on progressively more complex commonsense tasks, and leveraging [data-augmentation](https://www.wikiprompt.org/wiki/data-augmentation) to generate synthetic training examples that cover edge cases. Some work explores grounding language in embodied agents, such as robots from [figure-ai](https://www.wikiprompt.org/wiki/figure-ai) or [sanctuary-ai](https://www.wikiprompt.org/wiki/sanctuary-ai), which must learn commonsense through physical trial and error. Additionally, [reinforcement-learning](https://www.wikiprompt.org/wiki/reinforcement-learning) from human feedback, as used in [rlaif](https://www.wikiprompt.org/wiki/rlaif), can help align model outputs with human expectations of common sense.

Despite progress, commonsense knowledge remains a defining unsolved problem in AI. As models become more capable, their lack of true understanding becomes more apparent, particularly in safety-critical applications. The field continues to seek a unified framework that combines the strengths of symbolic reasoning, statistical learning, and embodied interaction, with the goal of creating systems that can navigate the everyday world as effortlessly as humans do.

## See Also

- [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence)
- [machine-learning](https://www.wikiprompt.org/wiki/machine-learning)
- [neural-network](https://www.wikiprompt.org/wiki/neural-network)
- [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)
- [knowledge-graph](https://www.wikiprompt.org/wiki/knowledge-graph)

---
Source: https://www.wikiprompt.org/wiki/commonsense-knowledge
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-14T04:28:56.119899+00:00