Evaluating AI output

University of Minnesota Libraries

17 Evaluating AI output

GenAI output should always be evaluated for validity, accuracy and bias.

Before you prompt GenAI:

Think about the questions (prompts) that you want to pose to a GenAI tool and how will you benefit from the capabilities that it offers?
GenAI is not a search engine – so don’t use it like a search engine. Sometimes Googling, or using a library database is a more effective step to find an answer.
Recognize that GenAI has limitations. It may not be the tool that you need for the task at hand.

Evaluating the response

Evaluating AI output is a critical skill when using GenAI. Because these models are designed to generate natural-sounding language, they can produce responses that seem convincing while being factually wrong, outdated, biased, or off-topic. Every GenAI response (especially for academic work or background research) must be verified against available sources and your independent logic and judgment.

Evaluate validity & relevance

Ensure the output actually addresses your specific prompt and follows sound logical steps.

Check prompt relevance: Confirm whether the GenAI answered your exact question, drifted into off-topic details, or provided generalized knowledge when you asked for specifics.
- Example: If you ask for recent policy changes in renewable energy taxation and the model returns a general history of solar panels, the answer may be factual, but it is invalid for your prompt.
Verify GenAI’s logic and reasoning: If a model refuses or is unable to explain its steps, treat the output as unreliable.
- Example: When solving a math or coding problem, execute the code in a sandbox or manually test each step of the equation rather than assuming the final line is correct.
Double-check calculations: Manually check unit conversions, raw percentages, and arithmetic using standard calculators.
- Example: If GenAI states that an increase from 40 to 60 is a 75% increase, run the calculations yourself to see if it is correct.

Fact-check everything

Never assume a cited source, headline, or data point is real without independent verification.

Verify Scholarly Sources:

Test DOIs & links: Paste AI-provided DOIs directly into dx.doi.org to find the article.
Search exact titles: Query exact article titles in quotes on Google Scholar or the University’s library databases.
Confirm authorship: Check that the listed author actually published that specific article in that specific journal.
Example: If GenAI cites “Bostrom, M. (2026). Quantum Neural Networks. Nature“, search Nature’s journal index to ensure that Bostrom published that specific paper in 2026.

Verify Web, News, & Media Sources:

Search exact headlines to locate the original reporting.
Check claims vs. sources: Compare what the original source actually says against the GenAI’s summary because they frequently misinterpret conclusions.
Example: If a GenAI summary claims a study “proved coffee causes zero sleep issues,” open the original PDF to verify what the study actually said.

Assess Timeliness & Training Cutoffs:

Compare AI-provided information against current sources to catch out-of-date practices, discoveries, or policy changes.
Example: Asking GenAI for tax filing deadlines or medical guidance might yield outdated guidelines based on older training data.

Investigate Visual Media (Images & Charts):

Look for visual glitches in images and videos like distorted text, extra fingers, and unnatural shadows. Run a reverse-image search to verify claims about historical events or locations.
Verify that axis scales, labels, and visual bar heights match the raw data numbers provided.
Example: If an AI-generated infographic shows a bar labeled “25%” that is taller than a bar labeled “50%”, the visual data representation is inaccurate.

Check for Bias & Missing Perspectives

Generative models often reflect biases present in their training data or offer partial explanations.

Identify crucial omissions in perspectives, missing variables, ignored counterarguments, or skewed framing.
Consult diverse, external sources beyond the GenAI tool’s output to gain a full picture of an issue.
Example: If you ask GenAI for the “economic impacts of urban gentrification” and it only highlights rising property values, you’ll need to actively search for literature covering community displacement and housing affordability.