Computer vision: how AI learns to “see”

The basic idea

Computer vision is the set of techniques that let a machine extract information from an image or video. In practice, that might mean saying what is in a photo, locating each object, reading printed text, tracking movement or spotting something unusual.

Since the 2010s, most of these systems have relied on deep learning: a neural network is trained on a very large number of labeled images (“this is a bike”, “this is a pedestrian”) until it can recognize those things in pictures it has never seen.

Everyday examples

• On your phone: face unlock, searching your photo library for “beach”, or translating a street sign by pointing your camera at it.
• At the office: scanning a receipt so the amount and date fill in automatically, or extracting text from a scanned PDF.
• In factories: a camera flags defective parts on a production line.
• In cars: driver-assistance features detect lane markings and some road signs.
• With an AI assistant: you snap a photo of instructions in a foreign language and the assistant explains them. This is called multimodal AI, because it handles text and images together.

Where it started

The ambition goes back a long way. In July 1966, Seymour Papert at MIT launched the “Summer Vision Project”, which aimed to get a program to split an image into regions and identify objects (original AI Memo 100). The problem turned out to be far harder than expected and kept researchers busy for decades.

Why it matters, and what to watch out for

Knowing that an assistant can “read” your images opens up practical uses: summarizing a table you photographed, checking the spelling on a flyer, or understanding an error message on your screen. To see which assistants accept images, check the AI assistants comparison.

A few myths worth correcting:

• “If AI recognizes something, it must be right.” It gives an estimate, not a certainty. For a medical question or any important decision, a qualified professional is still essential.
• “A photo only shows what I meant to show.” Pictures can reveal small details, such as an address on an envelope, a reflection or a license plate. Check before you share.
• “Seeing and creating are the same thing.” Image generation and deepfakes are about creating content, not analyzing it, even if the underlying techniques are related.