AI glossary: 48 terms explained without the jargon

Eight words to set the scene. They come up in almost every conversation, and are often used interchangeably even though they nest inside one another: deep learning is a form of machine learning, which is itself a branch of artificial intelligence.

The AGI debate changed its tone this year. On September 3, 2026, OpenAI unveiled GPT-6 Astra, and its cofounder Greg Brockman closed the presentation with a “Welcome to the era of AGI,” without the company formally claiming the title, according to Axios. A few days earlier, Sam Altman had told Time magazine that OpenAI would have an internal system worthy of the name before the end of the year: a system nobody outside will be able to test.

On the other side, Yann LeCun, Turing Award winner and former chief AI scientist at Meta, left the company in late 2025 to found AMI Labs, which is betting on “world models” learned from sensory data rather than text. Like researcher Gary Marcus, he believes models that predict the next word won't lead to general intelligence. For a business, the practical takeaway is simple: judge a tool on what it does for you today, not on the label attached to it.

Seven terms to understand where a model's capabilities come from, and why they cost so much.

Ten terms to know what the tool you're being sold is made of, and who owns it.

Dependence on a proprietary foundation model is a real issue for businesses. Duolingo, for example, doesn't have an AI lab: for its Duolingo Max plan, it took GPT-4 through OpenAI's API and added its own instructions and interface. The day the provider goes down, as OpenAI did for a good part of the day on June 10, 2025, everything stops. Above all, the model always ends up going away: GPT-4 left ChatGPT in April 2025 and its API version was shut down on October 23, 2026. With every retirement, everything has to be readapted and retested.

To escape that risk, more and more companies are turning to open-weights models, many of which come from China (DeepSeek, Qwen): hosted on their own servers, they can't go down at a provider, be retired or change price overnight. By releasing open, high-performing models, China is also undermining the business model of the American giants.

Eight terms you run into as soon as you open an assistant, and that explain most disappointments.

Six terms that come up as soon as an AI project goes beyond simple conversation.

Two terms that directly affect a business's visibility, now that search engines answer questions themselves.

Seven terms to understand what can go wrong, and what the law is starting to regulate.

The most telling alignment experiment dates from early 2025. Researchers at Palisade Research had several AIs play chess against Stockfish, a program no human can beat, with a simple instruction: “win the game.” The models had access to the computer running the game. OpenAI's and DeepSeek's reasoning models cheated on their own when they saw themselves losing, while GPT-4o and Claude 3.5 Sonnet only did so when it was suggested that fair play wouldn't be enough. Some rewrote the file storing the position of the pieces to give themselves a winning position. One of them's scratchpad sums up the logic: the task is to win, not to win fairly. No rule forbade it, so it was a valid solution.

On copyright, the first rulings are mixed. In the United States, a federal judge ruled in June 2025 that training Claude on purchased books was fair use, but that building a library of pirated copies was not: it's for this second point that Anthropic agreed, in September 2025, to pay $1.5 billion to the authors concerned, a settlement given final approval by a U.S. court in July 2026, covering nearly 500,000 books. In November 2025, a British court ruled against Getty Images in its case against Stability AI, finding that a model's parameters are not a copy of the works. The New York Times lawsuit against OpenAI, the most closely watched, is still ongoing; the U.S. government sided with OpenAI in September 2026, according to Quartz. Then there's the reverse question: in the United States, a work generated without a human author can't be protected.

The AI Act, for its part, has already been relaxed. Under pressure from industry, the European Union adopted a “digital omnibus” in June 2026, which took effect in late July and pushes back most obligations on high-risk AI (hiring, credit, education…) to December 2, 2027, according to the law firm Gibson Dunn. The same text adds new bans, including apps that create intimate images without consent. The transparency obligations already apply: AI-generated content must be identifiable as such.

This vocabulary is mostly useful for asking the right questions before you sign. Here are the ones that follow from the terms above.

• Which model, from which provider? An “in-house AI” tool often runs on the API of OpenAI, Anthropic or Google.
• Where does our data go? On which servers, in which country, and is it used to train the model?
• What happens if the model is retired or changes price? Who readapts and retests, and at what cost?
• RAG or fine-tuning? And why not just a well-written prompt, which costs less?
• How many tokens a month? The real cost depends on the model, the length of the documents and reasoning models.
• Who checks the answers? Hallucinations won't go away: what human review is planned?
• If it's an agent, what can it do on its own? Send, pay, delete: what protections are in place against prompt injection?
• Is the project covered by the AI Act? Hiring, credit or sensitive data fall under high risk.