The AI Landscape: Vocabulary Every Executive Needs
Deliberate Academy Editorial Team
Reviewed for accuracy and professional relevance
- Define the core AI vocabulary — large language model, hallucination, fine-tuning, context window, and agent — in plain business language
- Explain why not understanding these terms leaves a leader unable to evaluate a vendor claim or a technical team's recommendation
- Distinguish what current AI systems are reliably good at from what they are structurally unreliable at, independent of vendor marketing
- Identify the two-question test for deciding whether an unfamiliar AI term matters enough to ask about before a meeting ends
A vendor tells your leadership team that their new AI platform is "powered by a fine-tuned large language model with retrieval-augmented generation, running with a 128k-token context window." Two people in the room nod. Nobody asks what any of that means, because asking feels like admitting you are behind. Three months later, the platform is producing confident-sounding answers that turn out to be fabricated, the vendor calls it "a hallucination issue we're addressing in the next release," and the executive who approved the contract is the one explaining it to the board.
None of the vocabulary in that sentence is complicated. It is unfamiliar, which is a different problem, and one this lesson fixes in the next fifteen minutes.
The Five Terms That Do Most of the Work
Large language model (LLM). The underlying AI system — ChatGPT, Claude, Gemini, and similar tools are built on one. An LLM is trained on enormous volumes of text to predict, statistically, what words plausibly come next given what came before. It does not "know" facts the way a database does; it generates plausible-sounding text, which is usually accurate and sometimes confidently wrong. That single sentence is the most important fact in this course.
Hallucination. The term for when an LLM generates a fabricated fact, citation, statistic, or quote that sounds exactly as confident as a true one. It is not a bug that gets patched out — it is a structural property of how these systems generate text. Any AI system your organization deploys needs a plan for catching hallucinations before they reach a customer, a regulator, or a board deck, not a vendor promise that it "rarely happens."
Fine-tuning. Taking a general-purpose LLM and further training it on an organization's own data so it performs better on a specific task — for example, a legal document classifier fine-tuned on a law firm's own contract library. When a vendor says their product is "fine-tuned," the leadership-relevant question is: fine-tuned on whose data, and how was that data checked for quality and bias?
Context window. The amount of text an AI system can "see" at once — the current conversation plus any documents you have given it. A larger context window lets a system work with longer documents without losing track of earlier details. It is a real technical constraint, not just a marketing number, and it is worth asking whether your actual documents fit inside it.
Agent. An AI system that does not just answer a question but takes multi-step action on its own — searching a database, drafting an email, updating a record, calling another system. Agents are more capable and meaningfully higher-risk than a simple chat interface, because a mistaken action can have real consequences before a human ever reviews it. Whether a proposed AI system is a passive assistant or an autonomous agent is one of the first questions covered in Lesson 3.
You do not need to be able to build any of these systems. You need enough fluency to ask a precise follow-up question instead of a vague one. "Is this AI safe?" gets a reassuring, empty answer. "What happens when this model hallucinates a number in a customer-facing report, and how would we catch it?" gets a real one.
A vendor describes their product as using 'a fine-tuned LLM.' What is the most useful follow-up question for an executive evaluating this claim?
Select one answer.
What These Systems Are Actually Reliable At — and What They Are Not
Current AI systems are reliably strong at: summarizing long text, drafting a first pass of routine writing, extracting structured information from unstructured documents, and handling high-volume, repetitive language tasks. They are structurally less reliable at: stating a specific fact with certainty (they will sound equally confident whether they are right or wrong), reasoning about information outside their training or provided context, and any task where a wrong answer is expensive and hard to catch downstream.
That distinction — not "is AI good or bad" but "which category of task is this" — is the single most useful mental model a leader can carry into every subsequent lesson in this course.
A Confidently Wrong Board Statistic — Regional Healthcare Network
Context
A COO asked a newly deployed AI research assistant to summarize industry benchmark data for a board presentation on operational efficiency. The tool produced a clean, well-formatted summary with a specific benchmark figure attributed to a named industry report.
Action
A board member who happened to know the cited report well asked to see the source page. The COO's team could not locate the figure anywhere in the actual report — the AI tool had generated a plausible-sounding number and attached a real-sounding citation to it, a fabrication with no factual basis. The presentation was paused mid-meeting.
Outcome
The organization instituted a rule that no AI-generated statistic reaches a board document without an independently verified source page, and the COO now opens every AI-assisted briefing with a one-line disclosure of what was AI-drafted versus independently verified. The board's confidence in AI-assisted material actually increased afterward, because the verification standard was now visible rather than assumed.
Why is 'hallucination' described in this lesson as a structural property rather than a bug that gets fixed?
Select one answer.
Exercise
Your Task
Think of an AI tool or vendor pitch your organization has recently discussed. Write down every unfamiliar term used to describe it. For each term, apply the two-question test: (1) Does understanding this term change whether I would approve this initiative? (2) If yes, what is the one follow-up question I should have asked? If you cannot currently answer either question for a term, flag it as something to clarify before your next AI-related meeting.
Your reflection
Did you complete this exercise? What did you find? (Saved locally in your browser)
- An LLM generates statistically plausible text — it does not retrieve verified facts the way a database does. This single fact explains most of the risks covered in this course.
- Hallucination is a structural property of how LLMs generate text, not a bug that gets patched out — every AI deployment needs a verification plan, not just a vendor assurance.
- Fine-tuning, context window, and agent are not jargon for its own sake — each maps to a specific, leadership-relevant follow-up question about data quality, document size, and autonomy of action.
- The useful mental model is not "is AI good or bad" but "which category of task is this" — high-volume drafting and summarization are strengths; confident, verifiable single facts are structural weaknesses.
- A precise follow-up question, not a vague reassurance-seeking one, is what separates a leader who can evaluate an AI claim from one who is relying entirely on the vendor's own framing.