Artificial Narrow Intelligence (ANI)

Artificial narrow intelligence (ANI), also called narrow AI or weak AI, is artificial intelligence (AI) built to perform one task or a tightly bounded set of tasks, such as detecting objects in images, recommending videos, or transcribing speech. An ANI system can match or exceed human performance inside its domain, yet it cannot carry that skill to an unrelated problem without new training data and retraining. By the standard definition, every AI system deployed today counts as narrow, from spam filters to the computer vision models that guide self-driving cars.

ANI vs. AGI and ASI#

AI is commonly described in three levels of capability. ANI is the first: competent within one domain and defined by the data and objective it was trained on. Artificial general intelligence (AGI) is a hypothetical system that matches humans across virtually all cognitive tasks and learns new ones without task-specific engineering. AGI is often called strong AI, a term the philosopher John Searle coined alongside weak AI in his 1980 Chinese Room argument, where it meant a computer that genuinely has a mind. Artificial superintelligence (ASI) is a hypothetical system that surpasses the best humans in every domain, the scenario behind the idea of a technological singularity.

Large language models (LLMs) stretch the narrow category. One model writes code, translates, summarizes, and answers questions about images, which the EU AI Act recognizes by calling such systems general-purpose AI models. Their skills remain uneven, a pattern known as jagged intelligence: the same model can solve olympiad-level math and then misjudge which of two decimals is larger, and its weights stop learning once training ends.

How Narrow AI Works#

An ANI system is a model trained on data for one objective, so its competence is bounded by two things: the distribution of its training data and the set of outputs it can produce. An object detection model trained on the 80 classes of the COCO dataset has no way to say "this is something new". Shown an unfamiliar object, it either ignores it or reports the closest class it knows:

Ultralytics YOLO26 shows this boundary directly: the COCO-pretrained model can name exactly 80 classes, and teaching it a new object, such as a specific machine part, means fine-tuning on labeled images of that part, which replaces the model's class list with the new dataset's classes.

from ultralytics import YOLO

# A COCO-pretrained detector: its whole vocabulary is fixed at training time
model = YOLO("yolo26n.pt")
print(f"Classes this model can name: {len(model.names)}")

# Every detection it returns must carry one of those labels
results = model("https://ultralytics.com/images/bus.jpg")
result = results[0]
print(result.verbose())

The first line prints the size of the model's entire vocabulary, and the second lists which of those classes it found in the image, each drawn from the same fixed set.

Real-World Applications#

Narrow AI powers most of the software people use every day:

• Recommendation systems: streaming and shopping platforms rank items for each user from viewing and purchase history.
• Speech recognition: voice assistants and dictation tools convert spoken audio into text.
• Manufacturing inspection: vision models check every part on a production line for one set of defect types.
• Medical image analysis: models flag findings such as lung nodules or diabetic retinopathy in one imaging modality.
• Autonomous vehicles: a self-driving stack combines many narrow models, each handling one job such as lane segmentation, pedestrian detection, object tracking, or path planning.

Strengths and Limits of Narrow AI#

Within its domain, ANI can be superhuman, fast, and tireless. IBM's Deep Blue defeated world chess champion Garry Kasparov in 1997, and DeepMind's AlphaGo beat Go champion Lee Sedol 4–1 in 2016. Outside that domain, the same systems have no skill at all, and performance also drops on inputs that differ from the training data, a problem known as data drift once a model is deployed. Narrow systems also have no understanding of what they process; they map inputs to outputs through statistical patterns learned from data. Even DeepMind's AlphaZero, which learned chess, shogi, and Go with a single algorithm, trained a separate network for each game, and its chess network cannot play Go.