For Those Confused About AI Lingo

THE HUMAN TOUCH FIELD GUIDE

Artificial intelligence was confusing enough when it had one name. Now there's AI, ML, LLM, GenAI, AGI, ASI and a growing cast of other acronyms. And with President Donald Trump directing the American executive branch to adopt the term Super Intelligence, we have SI to contend with as well.

The trouble is that these words don't all describe the same thing. Some name technologies already embedded in our daily lives. Others describe capabilities scientists are still pursuing, or futures that may never arrive.

Here's a plain-English guide to eleven ideas worth knowing. No computer science degree required.

AI — Artificial Intelligence. The big umbrella: machines performing tasks associated with human intelligence, whether recognizing faces, recommending a film, translating languages or writing prose. The field acquired its name around the 1956 Dartmouth workshop. It did not begin with ChatGPT. But ChatGPT's arrival in 2022 brought AI dramatically into everyday conversation.

ML — Machine Learning. The process through which computers learn patterns from data instead of relying entirely on hand-written rules. Fraud detection, medical diagnostics, product recommendations and many advanced AI systems all depend on it. ML isn't separate from AI; it's one of the main ways we build it.

LLM — Large Language Model. The technology behind conversational systems such as ChatGPT, Claude and Gemini. Most leading LLMs use a design called a transformer, which helps them learn intricate relationships in language. This is why a machine can now draft a letter, explain a scientific paper, translate a passage or help write code. Fluency, however, is not the same as infallibility.

GenAI — Generative AI. AI that creates new content rather than merely classifying or sorting existing information. It can produce text, images, music, software and video. An LLM is one kind of generative AI; image and video generators are others.

RL — Reinforcement Learning. Learning through feedback. A system tries actions, receives signals about what worked, and adjusts its behaviour to seek better outcomes. It helped train AlphaGo to play world-class Go, and it's now an important technique for teaching sophisticated AI systems to tackle difficult problems. NIST defines the term in the language of rewards and actions. RL does not stand for recursive learning.

Synthetic Data. Information generated artificially, often with AI, to train or evaluate other AI systems. It can provide simulated medical cases, practice problems or unusual scenarios that are difficult to collect in the real world. But synthetic information isn't automatically reliable; as MIT researchers explain, data quality and bias still matter.

AI Agents. Systems that do more than answer questions. Agents can use tools and carry out multi-step tasks: looking up information, navigating software, writing code or preparing a report. Some need close supervision; others can work more independently. They're a glimpse of AI as an assistant that acts, rather than a chatbot that merely talks.

AGI — Artificial General Intelligence. The ambition to create an AI with broadly human-level intellectual ability across many different kinds of problems. It wouldn't have to be retrained from scratch for every new task. There is still no universally agreed definition or definitive test for AGI.

ASI — Artificial Superintelligence. A hypothetical intelligence far beyond the best human minds across practically every field. Philosopher Nick Bostrom described this idea decades ago. It's a much bigger claim than saying a system can beat a chess champion or write an impressive essay. This is also why using “Super Intelligence” as a replacement name for all AI can be misleading.

Recursive Self-Improvement. The possibility that an increasingly capable AI might help design or improve its own successors, so that each generation becomes better at developing the next. Researchers are exploring pieces of self-adapting model technology. Whether that becomes a runaway cycle of ever-more-powerful intelligence is still speculative.

The Singularity. A hypothetical moment when technology begins changing so quickly, potentially because of self-improving AI, that ordinary ways of predicting the future stop working. Futurist Ray Kurzweil predicts 2045. That's a forecast about the future, not a date on which science has established that an event will occur.

Here's the simplest way to keep these terms straight. AI is the broad field. Machine learning is a family of techniques inside it. LLMs and generative AI describe important systems we already use. Agents describe increasingly action-oriented applications. AGI and ASI describe levels of intelligence that remain contested ambitions. And the singularity is a prediction about what might happen if technological progress accelerates beyond our ability to anticipate it.

The vocabulary is useful not because we need another alphabet soup, but because the difference between a tool that exists today and a future that is still hypothetical matters to anyone trying to make sensible decisions about work, education, business or society.

For the curious twist that brought this terminology into the political spotlight, read our companion column, Trump's Super Rebrand, on Donald Trump's attempt to rename AI itself.