# Emergent Capabilities Abilities that appear in AI models at a certain scale — model size, training data volume, or compute — without being explicitly trained for. A model trained to predict text might "emerge" with the ability to do arithmetic, translate between languages it wasn't separately trained on, write code, or solve logic puzzles. The capability doesn't appear gradually; it seems to turn on at a threshold, absent below a certain model size and present above it. ## The contested framing The term does significant rhetorical work in both directions: **For AI labs and the safety community:** Emergent capabilities support the narrative that scaling AI is unpredictable and potentially dangerous. If you can't predict what a model will be able to do before you train it, then frontier model development requires extraordinary caution — and extraordinary governance. This framing supports both safety-research funding and the frontier-model regulatory category. **For skeptics:** A 2023 Stanford paper (Schaeffer, Miranda, Koyejo) argued that "emergent abilities" are largely a measurement artifact — they appear as phase transitions only when evaluated with specific metrics, and look like smooth, predictable improvement when evaluated with different metrics. The implication: AI capabilities scale predictably ([[Scaling Laws]]), and the "emergence" framing mystifies what is actually engineering-accessible improvement. **The political stakes:** If emergence is real, AI development contains genuine surprises — a model trained to write emails might unexpectedly develop the ability to plan bioweapon synthesis. This supports aggressive regulation and cautious development. If emergence is a measurement artifact, AI development is predictable and manageable — supporting lighter regulation and faster deployment. Both positions serve interests. The emergence framing helps labs argue for their own importance (they're managing dangerous, unpredictable technology) and helps safety researchers argue for their funding. The anti-emergence framing helps open-source advocates and deregulators argue that AI is a normal technology that doesn't require exceptional governance. ## See Also - [[Scaling Laws]] — the predictable counterpoint to emergence - [[Frontier Model]] — the systems where emergence is claimed - [[AGI]] — the endpoint that emergence narratives point toward - [[Capability Overhang]] — adjacent safety vocabulary