AI Literacy

The AI Literacy Test: Can You Answer These 5 Questions?

May 2026

By Mark Reeves — author of all 47 AI Field Guides. Written and tested in a working business.

AI-generated portrait. Pen name. Real work.

Five questions that separate people who understand AI tools from people who've just used them.

By Mark Reeves

A marketing manager pastes a statistic she found in ChatGPT into a board presentation — "$2.7 billion market size by 2027, attributed to Gartner." The Gartner report doesn't exist. The presentation does.

Using a tool and understanding it are different things. That gap — between having used something and knowing what it actually does and doesn't do — is where most of the mistakes happen.

These are five practical, diagnostic questions about how AI tools work in real use. The kind of thing that separates people who get consistent results from people who keep wondering why the output isn't quite right.

Question 1

Which type of AI tool should you be most cautious about using for research that requires current, verified facts — and what's the reason?

Think about the last time you asked an AI for a fact, a statistic, or recent information. What happened? Did you check it?

The Answer

Standard large language models — ChatGPT (without browsing), Claude, and most AI chat tools by default — should not be your primary source for current or verified facts. The reason is simple: they were trained on data up to a cutoff date, and they have no access to anything that happened after that point.

More critically, they don't "look things up" the way a search engine does. They predict plausible-sounding text based on patterns in their training data. That means they can state incorrect information — including outdated numbers, wrong dates, misattributed quotes, and outright made-up facts — in a tone that sounds completely authoritative.

This is called hallucination. The model isn't lying. It's doing exactly what it was built to do: generate fluent, coherent language. But fluency isn't accuracy. A 2023 survey of hallucination in natural language generation, published in ACM Computing Surveys (Ji et al.), documented this as a structural property of the generation process — not an edge case or a bug in the current generation of models.

Why it matters: If you paste an AI-generated statistic into a client proposal, a blog post, or a business document without checking the original source, you're taking on real professional risk. The fix is simple: treat any factual claim from a language model as a lead, not a source. Verify it before you use it.

The one meaningful exception: AI tools with live web access — like Perplexity, or ChatGPT with browsing enabled — can pull from current sources and will often cite them. That changes the equation. But even then, you should click the citation and confirm it says what the summary claims.

Question 2

What's the difference between a prompt and a brief — and why does giving an AI a brief tend to produce better output?

You've typed something into an AI chat window. How much context did you give it? Did you say what you wanted, or did you say what you wanted, why you need it, who it's for, and what format you expected?

The Answer

A prompt is an instruction. A brief is context plus instruction.

"Write a subject line for a sales email" is a prompt. A brief sounds like this: "I'm writing a sales email to small business owners who sell handmade goods on Etsy. They're tired of spending time on customer service. I'm promoting a tool that drafts responses for them. Write three subject line options — keep them under 50 characters, direct, no clickbait."

Same tool. Same model. Completely different output.

Here's why: AI language models have no idea who you are, what you're trying to accomplish, or what a good result looks like for your situation. They're working entirely from what you give them. When you give them a thin prompt, they fill in the gaps with statistical averages — which is another way of saying: the most generic version of what you asked for.

When you give them a brief — audience, context, constraints, format, goal — they have something to work with. The output narrows from "any email subject line" to "a subject line that fits this specific situation."

Why it matters: Most people who feel like AI "doesn't work for them" are handing it prompts when it needs briefs. This isn't a limitation of the tool. It's a usage gap. The more specific you are upfront, the less back-and-forth you need.

Question 3

An AI gives you a confident, specific statistic. What's the one thing you should always do before you use it anywhere?

This is related to Question 1 but more specific. The scenario: you asked, it answered, the answer sounds right, it's even plausible. What now?

The Answer

Find the original source. Not the AI's summary of it — the actual source.

This sounds obvious. Most people skip it. Here's why they skip it: the AI's answer sounds specific enough that it feels sourced. A precise number, a named study, a percentage — these signal credibility. The problem is that specificity in AI output is a stylistic feature, not a verification signal. A model will state "studies show 67% of consumers prefer X" and a "study" of that description may or may not exist.

When you go looking for the original source, a few things can happen:

• You find it, and the number is accurate.
• You find it, and the number is accurate but ten years old and no longer relevant.
• You find a similar study, but the actual figure is different from what was cited.
• You can't find the study at all.

All four of these outcomes matter. The first is the only safe one. The same ACM Computing Surveys hallucination survey (Ji et al., 2023) notes that models present fabricated specifics — named studies, precise figures, attributed quotes — with the same fluency and confidence as accurate ones. The specificity is stylistic, not evidentiary.

Why it matters: Specific, verifiable claims build trust with your audience when they're right, and they destroy it when they're wrong. A single bad statistic in a public piece of writing — a blog post, a LinkedIn article, a client document — can undermine everything else in it. The verification step takes two minutes. It's worth two minutes.

Question 4

When should you start a new conversation with an AI instead of continuing the one you're in?

This is a question most people never think to ask. They treat an AI chat session like a running conversation with a colleague — just keep going, keep adding context, keep building on what came before.

The Answer

Start a new conversation when:

The task has fundamentally changed. If you started writing a blog post and now you want to draft a client email, a fresh conversation tends to produce cleaner output. The model has been calibrated around the tone and context of the previous task, and that bleeds into what comes next.

The conversation has gotten very long. Language models work within a context window — a limit on how much text they can "see" and reason over at once. In a very long conversation, earlier instructions and context start to compete with each other or get weighted differently. The model can start to drift or forget specifics from earlier in the thread. If you're noticing outputs getting vaguer or off-target after a long session, a fresh start often fixes it.

You want a genuinely different perspective. If you've been iterating on something in one conversation and feel stuck, starting fresh — with a clean, well-written brief — often produces better alternatives than asking "can you try something completely different" in an existing thread. The model's probability distribution gets anchored to what it's already produced.

Why it matters: A long, messy conversation is one of the most common hidden causes of disappointing AI output. People assume the model is getting smarter as the session continues. In practice, long conversations introduce noise. Fresh context, clearly stated, almost always wins.

Question 5

You need to generate an image from text. You have ChatGPT, Claude, and Midjourney open. Which one is the right choice — and why?

This is a tool-selection question. What does each platform actually do when you ask for an image?

The Answer

Claude does not natively generate photographic or illustrative images. It can generate SVG graphics and code-based visualizations through Artifacts, and it can analyze images you provide — but it doesn't produce raster images (photos, illustrations, artwork).

ChatGPT with DALL-E integration can generate images in the same chat session. It's the most accessible option for image-plus-text workflows. The quality is capable; it's not the best dedicated option, but it's convenient.

Midjourney and other dedicated image tools (Adobe Firefly, Ideogram, DALL-E standalone) are purpose-built for image generation and will outperform chat-tool integrations on image quality.

Why it matters: Knowing what each tool is actually built to do prevents the frustration of using the wrong tool and blaming "AI" for the result. Claude is the right tool for extended writing and analysis. ChatGPT is the practical choice for text-plus-image in one session. Midjourney is the choice when image quality is the priority. Different tools, different jobs.

How Did You Do?

Here's an honest frame.

If you knew 0–1 of these: That's a starting point, not a failing grade. These aren't common knowledge — they're things that come from spending real time with the tools and hitting real walls. The AI for the Curious guide is exactly where to start. It covers the tool landscape without assuming you know anything, and it's free.

If you knew 2–3: You've been using AI tools and picking things up. You've probably hit some of the pain points these questions are pointing at. The next step is building more systematic habits — knowing which tool to reach for, when to brief instead of prompt, how to structure a working session.

If you knew 4–5: You're past the basics. You understand how the tools work at a level most people haven't reached. The question for you is whether you've turned that understanding into a real workflow — something you run consistently, not just when you remember to.

Try This Before You Move On

Take one piece of AI-generated output you've used in the last week — an email draft, a summary, a statistic, anything. Find the one claim in it that would matter most if it were wrong. Now look it up directly. Not a follow-up AI query — an actual search for the primary source.

What you find (or don't) is the result your literacy level determines.

Where to Go From Here

If any of these questions exposed a gap you want to close, the best starting point is the free AI for the Curious guide at ai-field-guide.com.

This article covered five diagnostic questions. Two things it doesn't cover that matter just as much:

How to build a working prompt habit. Knowing that briefs outperform prompts is different from having a brief format you actually use consistently. The guide has that format, with worked examples across the task types you'll actually encounter.

Which tool to use for which job. ChatGPT, Claude, and Perplexity each fail in specific, predictable ways. The guide covers the routing decision — which tool to reach for first depending on what you're trying to do, and what goes wrong when you get that wrong.

It's a practical first map — which tools matter, what they're actually good at, what to ignore for now, and how to get something working in your first session. No theory, no hype.

Related reading

• You're already behind on AI. Here's how far — and how to close the gap.
• What each AI tool is actually for

Mark Reeves is a pen name. AI Field Guide publishes role-specific, practical guides for using AI tools in real work.