Teaching Source Evaluation When Half the Web Is AI-Generated

The five-question source checklist most of us grew up teaching (Currency, Relevance, Authority, Accuracy, Purpose — the CRAAP test) was designed for a web where a human wrote the article. That is no longer the default. A student searching for "causes of the French Revolution" today will land on a mix of human-written history, AI-summarized human writing, and fully AI-generated articles that were never edited by anyone. The old checklist doesn't distinguish between them.

This piece is a classroom-ready update: a six-item framework, a printable student handout structure, and three model lessons for middle and high school.

What changed, precisely

Three things about the information landscape are new since 2022:

1. Cheap fluency. Confident, well-organized prose is no longer a signal of expertise. AI produces it for free.
2. Confident wrongness. Generative models hallucinate — they invent citations, statistics, and quotes that sound plausible. A student who was taught "if it has sources, trust it" is now the most vulnerable, not the least.
3. Untraceable provenance. A summary of a summary of a paraphrase of a study can look identical to the original. The chain is invisible.

Source evaluation now has to test for these three things directly, not just for authority and currency.

The updated framework: PROVE

A five-letter mnemonic students can actually remember:

• P — Provenance. Who or what wrote this? Is there a named human author with a real bio, or an unattributed "our editorial team"? Is there any signal (author page, LinkedIn, prior work) that the person exists?
• R — Receipts. Does every specific claim (a stat, a date, a quote) link out to a primary source? Click at least two. Do the links go where the article says they go?
• O — Original vs. Aggregated. Is this the source of the information, or a rewrite of someone else's reporting? Aggregators aren't bad, but they should send you to the original.
• V — Verifiability. Can a claim be checked against a source outside this article? If a statistic appears only in AI-generated content, that's a red flag by itself.
• E — Editing signal. Are there small, human-written details — a photo caption, an editor's note, a correction, a comment section — that suggest a human touched this page? Fully-automated content usually has none of these.

PROVE is faster than CRAAP and maps better to what students actually encounter. It also gives them something to do — click, check, compare — rather than a checklist to read.

Model lesson 1: The three-article compare (45 min, grades 7–10)

Give students the same query ("who invented insulin," "when did women get the vote in the US," "what causes ocean acidification") and three sources: a Wikipedia article, an AI-chatbot answer, and an AI-generated blog post from a low-effort SEO farm. Have them apply PROVE to all three and rank them.

The reveal: Wikipedia usually wins on Receipts and Editing signal, the chatbot wins on Fluency but often loses on Receipts, and the SEO farm loses on nearly everything. Students see the framework paying off in real time.

Model lesson 2: Hallucination hunt (30 min, grades 9–12)

Ask any general-purpose AI a specific, obscure question: "What did the 2003 Smith and Rodriguez study on adolescent sleep find?" Most models will confidently produce a study that does not exist, complete with a page number. Have students try to find it in Google Scholar. They can't. Discuss what happened and why the model produced it.

This is the single most memorable lesson about AI trust I've seen. Once a student catches a model in a fabrication, they never fully trust one again — which is exactly the disposition you want.

Model lesson 3: Provenance archaeology (60 min, grades 8–12)

Pick one viral claim ("chewing gum takes seven years to digest," "we only use 10% of our brain," "goldfish have three-second memories"). Have students trace it back through as many articles as they can, using PROVE to identify which ones cite an original source and which just cite another article. Most viral misinformation dies at step three: the trail leads to a broken link or a source that says something different.

What to put on the student handout

One page, top to bottom: the PROVE letters with a one-sentence question under each, a small "click at least two links" reminder, and a three-line box for the student to write "trust / partly trust / don't trust — because…". That's it. Handouts fail when they're a wall of text; the framework does the teaching, the handout just prompts the reps.

What not to teach

Don't teach students that AI content is inherently untrustworthy and human content is trustworthy. That was never true and is becoming less true. Teach them to check for the specific failure modes: hallucination, missing provenance, unverifiable claims. The medium isn't the problem; the missing chain of evidence is.