What is commonsense reasoning in AI?

Commonsense reasoning examples

Commonsense reasoning covers knowledge about the physical world, about time and about people. None of it is difficult for a person, which is exactly why it is so seldom spelled out.

• Physical: a laptop in a bag travels with the bag.
• Time: if someone is on holiday until Monday, a reply on Saturday is unlikely.
• People: a buyer who asks for references is probably close to a decision, not starting one.
• Social: thanking someone for a meeting that has not happened yet reads as odd.

What a Winograd schema tests

A Winograd schema is a pair of sentences that differ by one word, where that word changes what a pronoun refers to, so answering correctly needs common sense rather than grammar. In "The director turned down the vendor because they were too expensive", they means the vendor. Change expensive to cautious and they means the director.

Researchers proposed these puzzles as a test of machine understanding because word statistics alone should not solve them. Language models now handle many such sentences, which shows how much everyday knowledge text carries, though performance on new, carefully built cases is a better guide than performance on well-known ones.

Why AI lacks common sense

AI lacks common sense mainly because common sense is the knowledge nobody bothers to state. Text describes the surprising, not the obvious, so a system that learns from text sees far more about rare events than about the ordinary facts that frame them.

There is also no lived experience behind a language model's words. It has read about deadlines, holidays and handshakes but has never missed one, so it can fail when a situation combines familiar facts in an unfamiliar way.

Commonsense knowledge bases vs learning from text

There are two main ways to give AI commonsense knowledge: write it down explicitly or let a model absorb it from large amounts of text. Cyc is a long-running project to encode everyday facts and rules by hand in a logical form. ConceptNet is a network of everyday concepts and the relations between them, gathered largely from crowdsourced contributions.

Hand-built knowledge is precise and checkable but never complete. Learned knowledge is broad but patchy and hard to inspect. Research on hybrid systems combines the two, for example by training a language model on structured commonsense knowledge so it can produce new, plausible inferences about everyday situations.

What it means for AI reading sales email

For a sales team, common sense is what keeps an assistant's reading of an inbox sensible. In an illustrative case, a December reply that says "circle back after the holidays" means a follow-up in January, not next week. A contact who has just started a new job is unlikely to sign a renewal this week.

A model that misses these cues writes a follow-up that is too early, too pushy or plain odd. Read AI-suggested next steps with these everyday checks in mind, especially around dates, roles and tone.