The CRAAP Test in the Age of AI — A Librarian's Updated Checklist

The CRAAP test has been a librarian's best friend since 2004, when Sarah Blakeslee first published it in the CSU Chico library's quarterly. Currency, Relevance, Authority, Accuracy, Purpose. Five questions. No jargon. It worked beautifully for journal articles, news stories, and government reports — sources with authors, institutions, and publication dates you could actually look up.

Then came AI.

In 2025, ECPI University's library documented something striking: AI chatbots fail three of the five CRAAP criteria outright, and are weak in the remaining two. Not because they're malicious — but because the test was designed for sources with institutional accountability, and AI outputs simply don't have that.

If you're teaching or using the CRAAP test today, it needs an update. Here's what that looks like.

Currency: When Was This Actually Written?

Traditional currency check: When was this published? Is it recent enough for my topic?

AI-era addition: What is this AI model's knowledge cutoff?

Large language models don't know what they don't know. GPT-4o knows nothing about events after its training cutoff. Claude 3.5 Sonnet has a knowledge date. Gemini Ultra has its own. When you ask an AI about a recent law, technology, or policy, the answer may be confidently wrong — not because the AI is lying, but because it has no data on events that happened after it was trained.

Updated check: For any AI-sourced claim about laws, standards, technologies, or policies — search for the most recent source on that specific topic. If a 2026 article contradicts an AI output that may have been trained on 2024 data, trust the article.

Relevance: Could a Human Have Written This?

Traditional relevance check: Does this match my research question?

AI-era addition: Does this feel generic? Would a subject-matter expert with lived experience say it this way?

One of the clearest signs of AI-generated content is what I call the "confident vagueness" problem. The text is fluent, structured, and technically correct — but it never says anything a person who actually did the work would say. It has no rough edges, no specific failures, no real-world complications. It describes things as they should be in a textbook, not as they are in practice.

Updated check: Ask whether the source reflects the kind of nuance, specificity, or honest uncertainty that only comes from direct experience. If the content is smooth and generic with no texture, be suspicious.

Authority: Who Said This, Actually?

Traditional authority check: Who is the author? What are their credentials?

AI-era addition: Is this author's existence verifiable outside the source itself?

This is where AI content gets genuinely dangerous. AI tools cite Reddit and Wikipedia more than any other source category (Semrush, 2025). When an AI generates a claim and attributes it to a study, an expert, or a news article — you cannot assume that citation actually exists.

A 2024 study found that 30% of AI-generated citations led to papers that do not exist. Not misquoted — fabricated. The AI predicted what a citation would look like and generated one that sounded plausible.

Updated check: Every citation from an AI tool must be verified. Search for the author, the article title, the journal name. Check whether it appears in a reputable academic index (Google Scholar, JSTOR, PubMed, Web of Science). If you can't find it, the AI invented it.

When verifying an author's authority, look beyond the byline. Does the person exist in professional databases? Does their institutional affiliation check out? AI-generated bios — complete with plausible credentials and publication histories — are becoming common enough that the old "read the About page" check is no longer sufficient.

Accuracy: Is This Actually True?

Traditional accuracy check: Can I cross-reference this with other sources?

AI-era addition: Can I find the primary source, or am I only finding secondary AI summaries?

This is the hardest part. AI outputs are fluent enough that they feel verified. A well-structured paragraph with a statistic and a citation looks like a polished source — even when neither the statistic nor the citation is real.

The solution is not more fact-checking within the AI. It's leaving the AI entirely.

Updated check: Use the lateral reading method (which I covered in detail in a previous article). Leave the AI output. Open a new tab. Search for the specific claim in academic databases, news archives, or government publications. Apply the Rule of Three: find at least three independent, credible sources confirming the same fact before trusting it.

When an AI cites a study, find the actual study. When it cites a case law, pull the actual case. When it attributes a statistic, find the original data source. Do not trust summaries of sources — trust the sources themselves.

Purpose: What Is This Trying to Do?

Traditional purpose check: Is this informing, persuading, or selling?

AI-era addition: Is this output optimized for engagement or agreement, not accuracy?

Here is something the original CRAAP test didn't anticipate: AI outputs are optimized for being satisfying, not for being correct. A language model's training objective is to continue text in a way that sounds plausible to a human reader. Plausibility and accuracy are different things.

A polished, confident, well-structured AI response can be completely wrong and feel completely right. The fluency is not evidence of quality — it is evidence of statistical pattern-matching.

Updated check: Ask what the AI is optimized for. If you're using it to generate a first draft, to summarize existing material, or to brainstorm — that's appropriate use. If you're using it as a primary source of factual claims with no independent verification — that's a problem.

Quick Reference: Updated CRAAP Checklist for AI Content

CURRENCY
☐ Note the AI model's knowledge cutoff date
☐ Verify claims against sources published after that cutoff
☐ Check whether the topic has changed since the model's training date

RELEVANCE
☐ Does the content reflect domain-specific nuance?
☐ Could a practitioner with lived experience have written this?
☐ Does it include real-world complications, failures, or contradictions?

AUTHORITY
☐ Verify every citation independently — do not trust AI-generated references
☐ Check author credentials outside the source itself
☐ Cross-reference claims against domain experts and established institutions

ACCURACY
☐ Leave the AI output — don't fact-check inside the chat
☐ Find the primary source for every claim, not summaries
☐ Apply the Rule of Three: 3+ independent credible sources = confirmed

PURPOSE
☐ Is this content informing or just sounding informative?
☐ Is the AI tool being used as a search engine or a primary source?
☐ Would I trust this claim if it appeared without any citation?

How This Fits Into Practice

These checks aren't meant to make you distrust AI — they're meant to make you use it correctly. AI tools are powerful for drafting, brainstorming, explaining concepts, and synthesizing information. They fail when treated as authoritative sources of verified fact.

The updated CRAAP test gives you a vocabulary for that distinction. Currency becomes a data cutoff question. Relevance becomes a human-experience question. Authority becomes a citation verification question. Accuracy becomes a "get to the primary source" question. Purpose becomes an optimization question.

A Tool Built for This

When I evaluate sources through Sabia's evaluator, it applies these structural checks as part of the process — flagging when citations lack verification, when claims appear without corroboration, and when content has the characteristics of AI-generated text rather than documented research. If you want a systematic way to apply this checklist to the sources you encounter, sabialibrarian.com is built for exactly that.

The CRAAP test was good enough for twenty years of information literacy instruction. With a few AI-specific additions, it's good enough for twenty more.