The AI glossary 2025 demands of anyone in a product meeting, a pitch, or a boardroom has grown into something closer to a second language than a corporate cheat sheet. The acronyms have multiplied, the concepts have deepened, and some of the standards governing how AI systems talk to each other have, for the first time, become genuinely settled.
Start with the foundation that underpins almost everything else. A large language model, or LLM, is the engine inside ChatGPT, Claude, Google’s Gemini, Meta’s Llama, and Microsoft Copilot: a deep neural network built from billions of numerical parameters, trained on vast bodies of text, that generates responses by predicting the most probable continuation of whatever you typed. When you prompt one of these systems, you are not querying a database; you are nudging a probabilistic map of language toward an answer.
Above the LLM sits a term that generates more argument than almost any other: AGI, or artificial general intelligence. OpenAI CEO Sam Altman has described it as ‘the equivalent of a median human that you could hire as a co-worker,’ while the company’s own charter defines AGI as ‘highly autonomous systems that outperform humans at most economically valuable work.’
Google DeepMind takes a different view. In its paper ‘Levels of AGI: Operationalising Progress on the Path to AGI,’ summarised by the Montreal AI Ethics Institute, the lab defines the threshold as an AI system that is at least as capable as a human at most tasks, a framing that emphasises breadth of capability over economic output. The definitions do not agree, which is itself informative: AGI remains a contested benchmark, not a measurable destination.
The AI Glossary 2025 Starts Here: Agents, Standards, and Who Controls Them
One level below AGI, and considerably closer to commercial deployment, is the AI agent: a system that can take a sequence of actions on your behalf, from filing an expenses claim to writing, testing, and debugging code. Coding agents are a specialised variant, handling the iterative trial-and-error of software development with minimal human oversight, though a human still needs to review the output.
The practical challenge for anyone deploying agents is how they connect to the rest of a business’s software stack. That is the problem Model Context Protocol, or MCP, was designed to solve. Introduced by Anthropic in 2024 as an open standard, MCP gives AI models a single, consistent interface for connecting to external tools and data sources, from file systems and databases to applications such as Slack or Google Drive, without requiring a bespoke integration for every pairing.
The governance structure around MCP became considerably more formal on 9 December 2025, when Anthropic announced it was donating the protocol to the Linux Foundation and co-founding the Agentic AI Foundation (AAIF) alongside Block and OpenAI. Google, Microsoft, Amazon Web Services, Cloudflare, and Bloomberg joined as supporting organisations. The Linux Foundation also confirmed founding project contributions from Block’s open-source agent framework ‘goose’ and OpenAI’s ‘AGENTS.md,’ extending the AAIF’s scope well beyond any single protocol.
By the time of the donation, MCP had accumulated more than 10,000 active servers, client support across most leading AI platforms, and over 97 million monthly SDK downloads, according to the AAIF formation announcement. That adoption rate, achieved in roughly a year, explains why rival platforms moved to support MCP rather than build competing standards of their own.
Hallucination, Training, and the Economics of Inference
Perhaps no term in this AI glossary 2025 matters more to anyone spending real money on the technology than hallucination: the industry’s word for a model generating information that is simply incorrect. Hallucinations arise from gaps in training data and remain a fundamental quality problem. The response has been a push towards specialised, domain-specific models that narrow the knowledge surface where gaps can appear.
Behind every model output sit two distinct phases that shape cost. Training is the compute-intensive process of feeding data into a model and adjusting its weights, the numerical parameters that control how much importance the system assigns to different inputs, until outputs align with a target. Inference is what happens afterwards: running the trained model to generate responses. Training costs are one-off, though often enormous; inference costs scale with every query a product serves.
The unit of exchange in that scaling equation is the token: the small chunk of text, often a fragment of a word, that an LLM breaks language into before processing it. Most AI providers charge on a per-token basis, which means token throughput, the number of tokens a system can process per unit of time, directly determines both operating cost and the number of users a deployment can handle simultaneously.
The vocabulary in this AI glossary 2025 will keep expanding. The formation of the AAIF signals that the industry has decided certain foundational questions, how agents connect, how models share context, how autonomous systems are governed, are better settled in common than fought over commercially. Whether that consensus holds as the commercial stakes rise is the question worth tracking through the rest of the year.
