Where AI stands in 2026: the big picture in plain English

AI in one minute

Artificial intelligence is software that learns patterns from examples instead of following rules a person wrote by hand. Show a machine learning system millions of emails marked “spam” or “not spam” and it learns to spot spam on its own.

The current boom comes from one kind of AI: large language models (LLMs). They are trained on huge amounts of text to predict the next word, and at a big enough scale that simple trick lets them write, summarise, translate, answer questions and write code. Add images, sound and video, and you get generative AI: tools like ChatGPT, Gemini and Claude that create new content on request.

AI is not new. The term dates from 1955, and you have used it for years in spam filters, maps and phone cameras. What changed around 2022 is that anyone can now talk to it in plain language. See the AI timeline for how we got here.

What AI can do today

• Write and explain. Draft emails and reports, summarise long documents, translate, and explain hard topics simply.
• Code. AI writes and fixes code well enough that many developers use it every day. On a standard test of real coding tasks, top models went from 60% to nearly matching human experts in a single year (Stanford AI Index).
• Do tasks, not just answer. AI agents can take a goal, such as booking, researching or filling in forms, and work through the steps. Their success rate on real-world tasks jumped from 20% to 77% in a year (Stanford AI Index).
• Create images, video and voices from a short description. This is also why AI scams have become so convincing.
• Speed up science and medicine. AI predicted the shapes of nearly all known proteins, ran a full weather-forecasting pipeline on its own for the first time, and helps doctors spend up to 83% less time writing notes (Stanford AI Index). See more in mysteries AI helped solve.

What it still gets wrong

• It makes things up. Chatbots can state invented facts, quotes and sources with total confidence. This is called a hallucination, and it has cost lawyers, newspapers and consultants real money. See AI fails.
• It is uneven. The same systems that ace PhD-level exams succeed in only 12% of everyday household tasks when they control a robot (Stanford AI Index).
• It can be unfair. AI learns from past data, including its biases, so decisions about people need human checks.
• It’s often a black box. AI companies are sharing less about how their top models are built and tested: a transparency score tracked by Stanford fell from 58 to 40 in a year (Stanford AI Index).

Who is building it

• The big chatbots: OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude), Microsoft (Copilot, built largely on OpenAI’s models) and xAI (Grok). See the best AI chatbots.
• Open-weight models that anyone can download and run: Meta’s Llama, France’s Mistral, and Chinese labs such as DeepSeek and Alibaba’s Qwen. See what an open-weight model is.
• The chips: most AI is trained on GPUs, a market dominated by Nvidia.
• The US and China lead the race. US companies invested about 23 times more than Chinese ones in 2025, yet the performance gap between the best American and Chinese models has shrunk to just 2.7% (Stanford AI Index).

How it’s changing work and daily life

• At school: four out of five US high school and college students use AI for schoolwork, but only 6% of teachers say the rules on using it are clear.
• At work: AI is taking over routine parts of many jobs, and the first effects are showing at the entry level. In the US, employment of software developers aged 22 to 25 has fallen nearly 20% since 2024, with similar drops in other AI-exposed jobs such as customer service.
• Mixed feelings: worldwide, 59% of people are optimistic about AI’s benefits, while 52% say it makes them nervous. Only a third of Americans expect AI to improve their own jobs.

All figures in this section are from the Stanford AI Index 2026.

The big debates

• Energy and water. Training and running AI takes vast data centres. The International Energy Agency expects their electricity use to roughly double by 2030, to about as much as Japan uses today (IEA).
• Jobs. Will AI mostly help workers do more, or replace them? Both are happening, and how it balances out is the most watched question in economics.
• Safety and trust. Deepfakes, scams, misinformation and AI that makes mistakes in high-stakes settings such as health, law and hiring.
• Rules. Governments are moving at different speeds:
  • Europe: the EU AI Act is the most detailed AI law so far. Rules for the biggest general-purpose models apply since August 2025, but in 2026 the EU pushed most of its strictest “high-risk” rules back to December 2027, and some to 2028 (Cooley).
  • India: the government’s AI Governance Guidelines (November 2025) favour innovation, relying on existing laws and voluntary safeguards rather than a new AI law.
  • United States: there is no single national AI law, and trust in the government to regulate AI is the lowest of the countries Stanford surveyed, at 31%.

Where it’s heading

• From chatbots to agents: AI that does whole tasks for you, not just answers questions. Agentic AI is the word you’ll hear most.
• Cheaper and smaller: capable models now run on phones and laptops, not only in giant data centres.
• AI everywhere you already work: built into email, documents, search and phones, so you use it without opening a separate app.
• More science: AI-related research in the natural and life sciences is growing by more than a quarter every year.
• More rules, phased in: Europe’s law takes full effect over the next two years, and other countries are watching.

How we keep this page current

AI changes fast, so we review this page every quarter and update the numbers when new reports come out, such as Stanford’s annual AI Index each spring. Every figure links to its source. Spotted something out of date? Tell us. For the latest news, see our weekly issues.