What You Are Actually Talking To
When exploring artificial intelligence on campus, it helps to move past science-fiction tropes and look at the actual tools available today. While headlines often refer to "AI" as a single, human-like mind, everything currently in existence falls into specialized, task-focused categories.
To help faculty, staff, and students build a practical mental model of AI, here are three key types shaping higher education and everyday life—along with real-world examples you may already recognize.
1. Narrow AI (The Everyday Foundation)
What it is: AI designed to perform a specific, single task exceptionally well within strict parameters. It excels at its target function but cannot apply its knowledge to anything else. 100% of AI systems in existence today are forms of Narrow AI.
- Real-World Examples:
- Everyday Life: Email spam filters, facial recognition on smartphones, web search algorithms, and streaming recommendations on Spotify or Netflix.
- Campus Applications: Plagiarism checkers, auto-captioning tools in video lectures, and automated grading rubrics.
2. Generative AI (The Content Partner)
What it is: A specialized subset of Narrow AI that creates new content—text, images, code, or audio—by predicting statistical patterns learned from massive training datasets. Rather than searching a database or "thinking," it calculates what word, pixel, or sound should come next based on your prompt.
- Real-World Examples:
- Text & Code: Google Gemini, ChatGPT, Claude, and Microsoft Copilot. Used on campus to draft lesson plans, summarize lengthy research papers, brainstorm essay topics, or debug code.
- Visuals & Audio: Nano Banana (Gemini), DALL-E, Midjourney, and voice synthesis tools used to design presentation slides or course materials.
3. Agentic AI (The Autonomous Problem-Solver)
What it is: The next functional evolution of AI. While standard Generative AI waits for step-by-step human prompts, Agentic AI is given a high-level goal and can autonomously plan multi-step workflows, use external tools (browsers, databases, code execution), adapt to unexpected errors, and complete complex tasks.
- Real-World Examples:
- Autonomous Research Assistants: AI tools that independently search academic databases, evaluate source credibility, synthesize findings, run statistical code, and generate a formatted literature review.
- Administrative Workflow Agents: Systems (such as IBM watsonx Orchestrate) that can read incoming student inquiries, check schedule availability, negotiate meeting times, update calendar invites, and file forms—all without requiring step-by-step human intervention.
Exploring the Broader AI Landscape
NOTE: This section adapts concepts from insights published by IBM, the Syracuse University School of Information Studies, and Towards AI.
While Narrow, Generative, and Agentic systems represent the technologies in use today, computer scientists categorize AI across seven distinct capability and functionality tiers—most of which remain entirely theoretical.
We encourage readers to explore the full articles linked above for deeper context on:
- Theoretical Frontiers: Concepts like Artificial General Intelligence (AGI), Superintelligence (ASI), and Theory of Mind AI, including the technical hurdles researchers face in trying to achieve them.
- Why the Distinction Matters: Understanding why current tools excel at language generation but struggle with basic common-sense logic—a crucial perspective for maintaining academic integrity, critical thinking, and appropriate oversight when using AI in work and learning.
The Skills
Disappointing results usually come from missing detail, not from needing clever phrasing. A useful mental model: give the AI clear instructions the way you'd give them to a competent stranger — someone smart and capable, but with zero context about your specific situation. CAST is a simple way to check whether you've given it enough: Criteria, Audience, Specifications, Testing.
-
Criteria — set the shape and the limits. Tell the AI what form the answer should take: a paragraph, a bulleted list, a table, 200 words, five options. Then tell it what to avoid: "Don't use jargon," "Don't make up statistics," "Don't include anything I haven't given you." Constraints do as much work as instructions.
-
Audience — say who it's for. "Explain photosynthesis" and "Explain photosynthesis to a student who hasn't taken chemistry" produce genuinely different answers. Naming the audience — yourself, a specific class level, a general reader, an expert — tells the AI how much to assume and how to pitch the explanation.
-
Specifications — hand over the real material. If you want feedback, give it the actual text — your syllabus, your draft, your dataset — not a description of it. The more concrete detail and example you provide, the less the AI has to guess, and guessing is where most bad output comes from.
-
Testing — treat the first answer as a draft. Your first prompt rarely produces your final answer, and that's expected. Read the response, notice what's missing or off, and refine: add a constraint, narrow the audience, supply an example of the tone you want. Good prompting is closer to a short back-and-forth than a single perfect request.
You do not need clever phrasing or special keywords. Clear instructions to a competent stranger is the right mental model.
References:
Prompt Literacy: A Key for AI-Based Learning
Prompt engineering in higher education: a systematic review to help inform curricula
The single habit that most improves results: do not accept the first response. Tell the tool what was wrong with it.
“Too long — cut it in half.” “That is more formal than I need.” “You included three points I did not ask about; focus only on the second one.” “That example does not fit a community college context.”
This is not a workaround for a badly written prompt. It is how the tools are meant to be used. Getting to a good result in four exchanges is normal and faster than trying to write one perfect prompt.
AI tools do not signal uncertainty the way a person would. There is no hedge in the voice, no “I think,” no visible hesitation. The output reads identically whether it is solidly grounded or entirely fabricated.
Be especially skeptical when you see:
- Specific citations, page numbers, DOIs, or URLs. These are frequently invented. Assume every one is wrong until you have opened it yourself.
- Statistics with no source, or with a source that sounds plausible but that you cannot find.
- Recent events. Every model has a cutoff date after which it knows nothing, and it will often answer anyway rather than say so.
- Anything local. Maricopa policies, specific course requirements, a particular department’s process. The tool has almost certainly never seen this and will produce something that sounds right.
- Answers that are exactly what you were hoping for. These tools tend toward agreement, which feels good and is not the same as being correct.
AI is a poor choice when:
- The information must be current or authoritative. Policy, legal requirements, deadlines, prices, anything with a date attached. Go to the source.
- You would be entering information that should not leave the district. Student records, personnel matters, anything covered by FERPA, unpublished institutional data. See Responsible and Safe Use Basics.
- The work is the point. If an assignment exists so that you will practice a skill, having AI do it removes the reason for doing it at all.
- You would not be able to explain how you got the answer. If you cannot defend the process, do not submit the result.
- Judgment about a specific person is involved. Evaluation, discipline, hiring, advising a struggling student. These require someone who knows the situation and is accountable for the outcome.
If you use AI to draft an email, the email is yours. If you use it to help write a report, the report is yours. Every error in it is your error, and “the AI wrote that part” is not a defense anyone will accept.
In practice this means reading everything before it leaves you, checking anything factual, and making sure the result sounds like a person rather than like generic AI prose — which readers increasingly recognize and discount.
Try It
Pick something low-stakes that you actually need to do this week — a draft email you have been avoiding, an outline for something, a concept you want explained differently. Use a tool for it, then deliberately push back on the first answer twice. That exercise teaches more than any amount of reading about prompting.
When you have done that a few times, Critical Evaluation is the natural next step.