emergent capabilities
in English: emergent abilities
Capabilities that no one specifically trained the language model on still emerge when the model grows large enough.
In plain terms
Training a language model on its basic task consists of a single objective: predict the next word. Yet it can translate, write poetry, fix code, or reason through multi-step tasks. No one programmed these separately; size and the vast amount of text brought them along. This is called an emergent, or "appearing," capability.
It is like an ant colony. A single ant cannot build a bridge or cultivate mushrooms, yet the colony does so, despite having no chief engineer. The capability emerges from the cooperation of many simple parts. A child learning to read is similar, gradually understanding the logic of jokes in the process.
The concept is important because it explains why new models sometimes surprise their own creators. This makes it difficult to predict what the next, larger model will be able to do.
Traditional software only knows what the programmer wrote into it. The capabilities of a language model, however, must be discovered and measured afterward. Today's chat models are further trained after initial training (for example, to follow instructions), so not all their capabilities are emergent.
How it works
- The model is first trained on a massive amount of text for a single task: predicting the next word (more precisely, text segment, token). Chat models undergo targeted post-training, but most capabilities are formed in this first stage.
- To be good at this, it is forced to develop internal patterns regarding grammar, facts, logic, and even how an argument or calculation typically unfolds.
- In small models, these patterns are incomplete, so they perform almost randomly on complex tasks. For certain tasks, above a specific size (and training data volume), the pieces come together, and measured performance jumps dramatically.
- Researchers discover what a model can do through test tasks after the fact. Often users find new capabilities after launch.
- It is debatable whether the jump is real. With finer scoring, improvement often appears gradual, meaning the sudden emergence may partly stem from the measurement method.
An everyday example
You ask a chatbot to rewrite your grandmother's handwritten recipe, full of dialect words, into English, with US units of measurement, in a vegan version. No one trained it specifically for such a task, yet it solves it because translation, conversion, and cooking knowledge combine within it.
Who is affected
A common misconception
Limits and risks
- Just because a capability has emerged does not mean it is reliable: the model may excel in the same task type at one time and confidently err at another.
- It is impossible to predict with certainty at what model size specific capabilities emerge, so unwanted capabilities (such as deception or writing harmful code) may also arrive as surprises.
- The term is often used in marketing to create a mystical atmosphere; if you read about a 'suddenly emerged' capability, ask how it was measured.
- Researchers disagree on whether these represent genuine leaps or if the choice of metrics creates this impression.
Similar, but not the same
- fine-tuning – In fine-tuning, the model is taught a task intentionally with targeted examples, whereas emergent capability appears without separate training.
- hallucination – A hallucination is a model error in which it asserts a plausible-sounding falsehood, while an emergent capability is a skill that genuinely works but was not planned in advance.
- artificial general intelligence – Emergent capabilities relate to specific, measurable tasks and do not mean the model possesses human-level general intelligence.
Related concepts
emergencia, large language model, deep learning
in-context learning, zero-shot learning, chain of thought
scaling, foundation model, performance measurement, AI safety, Interpretability
A brief history
The concept of emergence originates from 19th-century philosophy (the term was introduced by George Henry Lewes in 1875). In the natural sciences, it was widely popularized by physicist Philip W. Anderson's 1972 essay 'More Is Different'. For language models, the term became widely known through a 2022 study by a research group consisting mainly of Google researchers (Wei et al.: Emergent Abilities of Large Language Models). In 2023, Stanford researchers challenged this, arguing that the sudden jump is partly a measurement artifact.
Where things stand
At the end of 2026, the term is established in both professional and public discourse, and researchers regularly report unplanned capabilities in large language models. However, it remains an open research question whether capabilities truly emerge in sudden leaps and how predictable they are.
In the news
- AI agents are already shopping, say OpenAI, BCG and J.P. Morgan; retailers urged to prepareSep 14, 2026 · Marketing
Also known as
emergent capabilities · emergent behavior · emergent capabilities · emergencia · emergent abilities · emergent capabilities · emergent behavior · emergence
Sources
- Wei et al.: Emergent Abilities of Large Language Models (2022, arXiv)
- Schaeffer, Miranda, Koyejo: Are Emergent Abilities of Large Language Models a Mirage? (2023, arXiv)
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.