index

Artificial Intelligence

Predictive AI

Predictive AI uses past data and smart statistical modeling to give you a glimpse of what’s coming next. By spotting patterns in historical events, it turns that knowledge into practical insights you can act on. 

Generative AI (Gen AI)

Artificial intelligence that generates new content (text, images, audio, video, code) by learning patterns from huge datasets and producing new variations.

Agentic AI

Agentic AI doesn’t just generate an output; it works toward achieving a goal and takes several steps on its own. Think of it like a newly hired executive assistant: it has limited power to act in your name.

LLM & Tokens

Large Language Model (LLM)

The engine of generative artificial intelligence, trained on massive amounts of text to absorb patterns.

Tokens

Small units of data (words, word pieces, characters, pixel groups) that AI models process and generate.

Tokens are both the operating language and the billing unit of generative AI.

What Tokens Look Like

Source: OpenAI Tokenizer (opens in a new window)

What the AI is working with

Training knowledge

What the model absorbed when it was built. This always comes with a cutoff date.

Many AI tools can look things up online, which helps fill in recent gaps.

Context window

How much the model you’re using can pay attention to at once.

Each time it responds, the model rereads the whole window from the top. For best results, start a new chat for every new task.

What goes into the window:

• Prompt

  What you type into the chat box: your question, request, or instructions.

  The more specific it is, the less the model has to guess.

• Instructions (system prompt)

  Directions the tool loads before you type anything. Some come from the company that makes the tool, and some you can set yourself as custom instructions.

  This is why the same model can behave differently in Copilot than in Claude.

• Memory

  An optional feature that saves details from past conversations and loads them into new ones.

  It’s useful, but it’s a summary, not a transcript. You can usually view, edit, or turn it off in settings.

• Files and connected data

  Documents you upload, or work data the tool can pull in, such as your email or SharePoint in Copilot. Pulling in outside information like this is sometimes called grounding.

  A very long file can take up a lot of the window, or only part of it may get read.

• Search results

  When a tool looks something up online, the results go into the window like anything else.

  The model is only as reliable as what it found.

Types of AI Accounts

Consumer vs Enterprise

Consumer accounts (free or paid) are personal accounts that can learn from what you share.

Enterprise accounts are institution-issued and come with data protections, meaning the model can’t learn from your input.

Hallucinations and bias

AI tools can be wrong in two different ways. A hallucination is a specific false statement, such as a made-up fact, quote, or source. Bias is a pattern in the output that consistently leans one way, often in ways that are harder to spot than a single wrong fact. Both come from how these tools are built, so you can’t fully turn either one off. Learn to notice them and reduce their effects.

Hallucinations

A hallucination is when the AI makes something up and states it as fact, often with complete confidence and in polished language.

Why it happens

An AI language model generates text by predicting what is likely to come next, based on patterns it learned from a very large amount of writing. It is built to produce text that sounds right. Sounding right and being right are not the same thing.

Several conditions make hallucinations more likely:

• Gaps in what it learned. If a topic is obscure, very local, or very recent, the model has less to draw on. It may fill the gap with something that has the right shape, like a plausible name, date, or policy number.
• Knowledge cutoffs. Each model’s training data stops at a certain date. Without a search or file to check against, it may present outdated information as current.
• Pressure to answer. Models are tuned to be helpful, and they tend to produce an answer instead of saying “I don’t know.”
• Requests for specifics. Exact figures, citations, URLs, and quotations are high risk because the model can generate something that looks exactly like a real citation without one existing.
• Long or complex inputs. When you give it a long document or a lot of instructions, it can misread, blend details together, or attribute a point to the wrong section.

Common forms

• Invented sources, including articles, court cases, or studies that don’t exist, or real titles with fake details
• Quotes attributed to the wrong person, or quotes no one said
• Wrong numbers or math presented with confidence
• Summaries of a document you provided that include claims the document doesn’t make
• Confident descriptions of a policy, product feature, or procedure that are outdated or invented

Grounding helps but doesn’t eliminate the risk

Tools that search the web or work from documents you provide, such as Copilot with access to your files or Claude with web search, hallucinate less because they have real text to check against. They can still misread a source, cite a page that doesn’t support the claim, or mix details from different sources. A citation shows you where to check. It doesn’t prove the claim.

Always fact-check

Pay particular attention to:

• Names
• Numbers
• Quotes
• Citations, especially by opening the source and confirming it says what the AI claims
• Policies and procedures
• Dates and anything described as “current”

Bias

In AI, bias is a systematic tendency for outputs to favor certain groups, viewpoints, assumptions, or answers over others. A hallucination is one wrong statement, while bias is a slant that repeats across results.

Why bias appears in AI results

• Training data reflects the world as it has been written about. Models learn from huge collections of text that overrepresent some languages, regions, demographics, and perspectives. Historical inequities and stereotypes in that writing get absorbed along with everything else.
• Defaults fill in what you didn’t specify. If a prompt leaves something open, the model fills it with the most common pattern it has seen. For example, it might assume a nurse is a woman, a “professional” tone means a particular dialect, or “the student” fits a traditional profile.
• Human tuning adds its own preferences. After initial training, models are adjusted using feedback from human reviewers and company guidelines. This makes them safer and more useful, but it also builds in the reviewers’ and designers’ judgments about what a good answer looks like.
• What’s missing matters too. Topics, communities, and institutions with less written about them get thinner, more generic, or less accurate treatment.

Bias you bring to the conversation

Some bias comes from how people interact with AI, not just from the model:

• Leading prompts. “Why is this policy a good idea?” gets a different answer than “What are the strengths and weaknesses of this policy?” The model tends to go along with the framing it’s given.
• Sycophancy. Models tend to agree with the user, praise drafts, and back down when challenged, even when they were right the first time. Agreement from an AI isn’t independent confirmation.
• Automation bias. People tend to over-trust output that comes from a computer, especially when it’s well formatted and confident.
• Confirmation bias. We tend to accept AI answers that match what we already believed and question the ones that don’t.

Where it matters most

Be especially careful when AI output affects people, such as:

• Hiring materials, job descriptions, or candidate screening
• Performance feedback or evaluations
• Summaries of student, staff, or stakeholder input, since some voices may be flattened or dropped
• Communications meant for diverse audiences
• Generated images of people

Reducing bias

• Write neutral prompts, and ask for multiple perspectives or counterarguments.
• Specify your audience and context instead of letting the model assume them.
• Ask, “What assumptions did you make here?” or “Who might this miss or misrepresent?”
• Compare AI output against your own knowledge and real sources, not just its own reasoning.
• Keep a person accountable for any decision that affects individuals. AI can help draft or organize, but it shouldn’t make the call.

The bottom line

AI output is a first draft from a fast, widely read, but unreliable collaborator. It can save real time, but you are still responsible for what you publish, send, or decide based on it.

You decide, judge, and own the outcome.