A familiar non-AI example is a traffic jam that forms on a highway with no accident: each driver follows simple habits, and the stop-and-go wave emerges from the group. In AI, people use the term in two main ways.
First, for large models. Researchers have reported “emergent abilities,” skills such as multi-step arithmetic that seemed absent in smaller models and appeared suddenly in larger ones. Other researchers argue that many of those jumps come from how performance is measured, and look smooth with different metrics. Both positions are published research; the debate is not settled.
Second, for systems of AI agents, where agents interacting with tools, each other and people can produce patterns nobody planned.
Why it matters
Unplanned behavior is hard to predict and test in advance. That is useful when it means new capabilities, and risky when it means surprises in production.
On the podcast
In episode 63, host Conor Bronsdon says “we’re actually seeing that emergent behavior with coding agents today” at 14:42, describing agents that evolve, with little human guidance, to fill digital niches. That is his observation, not a measured result.
Related terms
Compound AI systems and multi-agent systems are where agent-level emergence tends to appear.
Sources
- Wei et al.: Emergent Abilities of Large Language Models — Defines an emergent ability as one not present in smaller models but present in larger ones, and surveys examples; checked October 24, 2026.
- Schaeffer, Miranda and Koyejo: Are Emergent Abilities of Large Language Models a Mirage? — Argues that many reported emergent abilities reflect the researcher's choice of metric rather than sharp changes in model behavior; checked October 24, 2026.
