A Librarian's Guide to Evaluating Sources in the Age of AI

También disponible en español: Guía para Bibliotecarios: Cómo Evaluar Fuentes en la Era de la IA

The volume of AI-generated content on the web doubled in 2025. By the time a student or patron asks you to help verify a source today, there is a real chance that article was written by a language model that has never read a primary document, never verified a claim, and has no professional accountability for what it produces.

That is the new baseline. This guide gives you a working framework for source evaluation in this environment — one that you can teach, use yourself, and build into library instruction without starting from scratch.

Why AI Changes Source Evaluation (But Not the Principles)

Before 2023, most misinformation had identifiable patterns: poor grammar, obvious bias, unknown publishers, broken links. Students could be trained to spot these signals. AI-generated content has eliminated most of those surface-level cues.

A language model can now produce a 1,500-word article with:

• A plausible author name and institutional affiliation
• Correctly formatted citations (that may or may not exist)
• A neutral, authoritative tone that reads like journalism
• Internal coherence — the argument flows logically from premise to conclusion

None of those signals tells you whether the underlying claims are true. This is the central problem: AI has decoupled fluency from accuracy. A source can be well-written and completely fabricated at the same time.

The good news is that the principles of source evaluation remain sound. Authority, accuracy, currency, purpose — these still matter. What changes is where you look for evidence of them.

The SIFT Method: Your First Line of Defense

For most library instruction settings, the SIFT method developed by Mike Caulfield is the most practical starting point. SIFT stands for:

Evaluating AI-Generated Content Specifically

When you suspect a source may be AI-generated, add these checks to your standard evaluation:

Check the Citations

AI language models frequently hallucinate citations — they generate plausible-looking references to studies, articles, or books that do not exist. Before trusting any factual claim, look up the actual cited source. If it does not exist, or if the article misrepresents what the source says, the entire piece is unreliable regardless of how well it is written.

Look for Temporal Inconsistencies

AI models have training cutoffs. A 2026 article written by a model with a 2024 cutoff may discuss ongoing events as though they have not yet happened, or omit developments from the last two years. If the article feels slightly out of date in specific ways, that is a signal.

Check the Author

Search the author's name independently. Does this person have a professional presence — academic publications, professional affiliations, social media history, other bylines? A completely blank web presence for a claimed expert is a red flag. A rich, consistent presence is not proof of authenticity, but its absence is evidence of a problem.

Run It Through an AI Evaluation Tool

Tools like Sabia evaluate source credibility systematically — checking authority signals, assessing bias indicators, and flagging consistency issues. These tools are not infallible, but they surface the same questions a trained librarian would ask, faster, and they show their reasoning so students learn the evaluation process alongside the result.

A Classroom-Ready Evaluation Workflow

This five-step workflow is designed for a 15-20 minute library instruction session. It works for both AI-generated and traditionally produced content.

1. Identify the claim you are evaluating. What specific assertion does this source make? Be precise — "this article says vaccines cause autism" is evaluable; "this article is about vaccines" is not.
2. Investigate the source before reading. Use lateral reading: search the publication name + "bias" or "reliability." Read what others say about the publisher, not what the publisher says about itself.
3. Find the primary source. If the article cites a study, a government report, or a news event, find that original source directly. Does the article represent it accurately?
4. Check who else is reporting the claim. If the claim is true and important, multiple credible outlets will have covered it. If only one source makes the claim, apply extra skepticism.
5. Assess the purpose. Why was this content produced? Who benefits if you believe it? Advertising disguised as editorial, advocacy disguised as reporting, and AI-generated content disguised as journalism all share a common feature: their purpose is different from what they appear to be.

What to Teach vs. What Tools Handle

One of the most practical questions for library instruction in 2026 is the division of labor between human judgment and AI evaluation tools.

Teach humans to do this

• Recognize the type of source and what standards apply to it
• Perform lateral reading and find primary sources
• Assess purpose and potential conflict of interest
• Make judgment calls when evidence is mixed or incomplete
• Recognize when they lack domain knowledge to evaluate a claim

Let tools handle this

• Systematic credibility scoring across multiple criteria simultaneously
• Flagging known unreliable publishers and propaganda outlets
• Generating follow-up questions students can pursue independently
• Demonstrating what a structured evaluation process looks like (modeling)

The goal is not to replace human judgment with AI — it is to use AI to demonstrate and scaffold the evaluation process until students can run it independently. Sabia is designed specifically for this: it evaluates the source, shows the reasoning, and generates Socratic follow-up questions that push students to investigate further rather than accepting the tool's score.

Adapting for Different Library Contexts

School Libraries (K-12)

Focus on the CRAAP test and SIFT for younger students. Prioritize the habit of pausing before sharing — the Stop step in SIFT. For older students, introduce lateral reading as a practical research skill. The most common failure mode in K-12 is accepting the first result that confirms what a student already believes. Your instruction should specifically address confirmation bias.

Public Libraries

Health misinformation and news literacy are the highest-priority areas for public library patrons. Focus on tracing medical claims back to published studies (and knowing the difference between a preprint and a peer-reviewed article). For news literacy, teach the distinction between news, opinion, editorial, and sponsored content — these are increasingly difficult to distinguish visually.

Academic Libraries

The most pressing issues in academic settings are AI-generated academic content and the explosion of predatory journals. Students need to know how to verify a journal's legitimacy, evaluate preprints responsibly, and recognize when a citation trail goes cold (a sign of fabricated sources). The ACRL Framework remains the standard — focus especially on "Scholarship as Conversation" and "Authority Is Constructed and Contextual."

The Long Game: Building Habits, Not Checklists

The research on information literacy instruction is clear: one-off sessions don't produce lasting change. Students who attend a single library workshop perform marginally better on source evaluation tasks immediately afterward. Six months later, the effect is statistically indistinguishable from the control group.

What works is repetition embedded in regular practice. This means:

• Integrating evaluation exercises into ongoing class activities, not standalone workshops
• Building source evaluation into research assignment rubrics so it is graded and therefore practiced
• Creating regular touchpoints — newsletters, reference desk interactions, course integrations — that keep the habits active
• Tracking behavioral change over time, not just workshop attendance or satisfaction scores

The librarian's role in 2026 is not to evaluate sources for people. It is to build the capacity for people to evaluate sources without you. Every interaction is a teaching opportunity, and every teaching opportunity should leave the patron slightly more capable than before.