Artificial Intelligence (AI)

Artificial intelligence (AI) is the branch of computer science that builds systems able to perform tasks associated with human intelligence: perceiving images and sound, understanding and generating language, reasoning, planning, and learning from experience. The OECD definition, which the EU AI Act closely follows, describes an AI system as a machine-based system that infers from its inputs how to generate outputs such as predictions, content, recommendations, or decisions. Most AI in use today learns that behavior from example data through machine learning (ML).

A Brief History of AI#

Alan Turing's 1950 paper "Computing Machinery and Intelligence" asked whether machines can think and proposed the imitation game, now called the Turing test. John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon coined the term "artificial intelligence" in their 1955 proposal for the 1956 Dartmouth summer workshop, which founded the field. For the next three decades symbolic AI dominated with logic, search, and rule-based expert systems, and funding collapsed twice, in the "AI winters" of the 1970s and late 1980s, when results fell short of promises.

Statistical AI methods that learn from data took over in the 1990s. In 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov with special-purpose hardware that searched about 200 million chess positions per second. In 2012, the AlexNet convolutional neural network won the ImageNet image classification challenge by a wide margin and started the deep learning (DL) era. The transformer architecture, introduced in 2017, led to large language models (LLMs), and ChatGPT's release in November 2022 brought generative AI to hundreds of millions of users.

How AI, Machine Learning, and Deep Learning Relate#

AI is the umbrella field. Symbolic AI and machine learning are its two main approaches: the first encodes knowledge as explicit rules, and the second learns patterns from examples. Deep learning is the subset of machine learning that uses neural networks with many layers, and nearly every modern breakthrough in vision, language, and image generation is built on it.

Machine learning systems learn in a few main ways. Supervised learning fits labeled examples, such as images tagged with the objects they contain. Unsupervised learning finds structure in unlabeled data, self-supervised learning creates its own training signal from raw data, the method behind LLM pretraining, and reinforcement learning learns from rewards gained by acting in an environment.

Types of AI by Capability#

• Artificial narrow intelligence (ANI) or weak AI: AI built for a specific task or set of tasks. Every deployed system falls here, including systems that are superhuman within their task.
• Artificial general intelligence (AGI): A hypothetical system that matches humans across virtually all cognitive tasks, often called strong AI.
• Artificial superintelligence (ASI): A hypothetical system that surpasses the best humans in nearly every domain, the scenario behind the technological singularity.

General-purpose LLMs blur the first boundary. One model can write code, translate, and describe images, yet its skills are uneven across tasks, a pattern called jagged intelligence.

Applications of AI#

• Computer vision: Detecting objects, segmenting scenes, and tracking motion for autonomous vehicles, factory inspection, retail analytics, and medical image analysis.
• Language and speech: Chatbots, machine translation, document summarization, and speech-to-text transcription.
• Prediction and ranking: Fraud detection, demand forecasting, and recommendation systems for streaming and shopping.
• Science: DeepMind's AlphaFold predicts protein structures and earned its creators a share of the 2024 Nobel Prize in Chemistry.
• Agents and robotics: AI agents that plan multi-step tasks and call software tools, and robots that combine perception with control.

Running a pretrained vision model is the quickest way to watch a deep learning system work. Ultralytics YOLO26 is an object detector that has learned 80 everyday object classes from labeled COCO images:

from ultralytics import YOLO

# Load a pretrained YOLO26 object detector
model = YOLO("yolo26n.pt")

# Run inference on an image the model has never seen
results = model("https://ultralytics.com/images/bus.jpg")
result = results[0]

# Save the image with predicted boxes, class labels, and confidence scores
result.save(filename="bus_detections.jpg")

The saved image marks each detected object, such as the bus and the people beside it, with a box, a class label, and a confidence score, all of it learned from examples.

Risks, Ethics, and Regulation#

AI systems inherit the flaws of their data and objectives. Models can encode bias against groups that were underrepresented or mistreated in their training data, LLMs can state falsehoods fluently, generative models can produce deepfakes, and opaque models are hard to audit without explainable AI (XAI). AI ethics and AI safety research address these risks, and the voluntary NIST AI Risk Management Framework gives organizations a process for managing them. The EU AI Act, in force since August 2024, bans a short list of practices such as social scoring outright and sets strict requirements for high-risk uses such as hiring, credit scoring, and AI in medical devices.