Skip to Main Content

Artificial Intelligence: For Students

Methods for Evaluating AI

Lateral reading is done when you apply fact-checking techniques by leaving your initial source and consulting other sources to evaluate the initial source's credibility. You can think of this as “tabbed reading,” moving laterally away from your initial source to sources in other tabs rather than just proceeding “vertically” down the page based on the characteristics of the initial source alone. 

While you can typically reach a consensus about online sources by searching for a source’s publication, funding organization, author or title, none of these bits of information are available to you when assessing AI output. As a result, it is critical that you read several sources outside the AI tool to determine whether credible, non-AI sources can confirm the information the tool returned.

Here's how to fact-check something you got from an AI tool using lateral reading: 

❏ Break down an AI-generated response into individual claims.

❏ Open a new tab and look for supporting pieces of information. Here are some good sources to start with:

  • When searching for specific pieces of information: Google results or Wikipedia (a starting point and with caution)
  • When seeing if an article, book, or other resource exists: Google, Google Scholar, or WorldCat 
  • Some things to watch out for: Is the AI putting correct information in the wrong context, for example, writing in an essay that a professor works at Dartmouth when they really work at MIT? Is it attributing a fake article to an actual author, or vice versa?

❏ Next, think deeper about what assumptions are being made here

  • What did your prompt assume? Did you assume that AI will understand that you're asking about a particular demographic, geographic area, or time period? 
  • What did the AI assume? Did they give you news sources when you were seeking academic ones, or summarize an article at a middle school reading level rather than a college reading level? 
  • Who is knowledgeable about this topic? Would someone knowledgeable about this topic have a different perspective than what the AI is offering? What resource would you use to determine those perspectives?

❏ Finally, make a judgment call. 

  • Is the generated content true, misleading, and/or factually incorrect?
  • Can you re-prompt the AI to try and fix some of these assumptions and/or falsehoods?  
  • Did this evaluation process lead you to discover new sources or perspectives?
  • Remember, you’re repeating this process for each of the specifics generated from AI – go back to your list from the first step and keep going!

 

Text adapted from "Assess Content: Assessing AI-Based Tools for Accuracy" by the University of Maryland under the Creative Commons Attribution NonCommercial 4.0 International License Creative Commons

The ROBOT method for evaluating sources applies to information about AI in particular, but can be used for other types of sources as well.

Being AI Literate does not mean you need to understand the advanced mechanics of AI. It means that you are actively learning about the technologies involved and that you critically approach any texts you read that concern AI, especially news articles. 

Reliability

Objective

Bias

Ownership

Type

Reliability
  • How reliable is the information available about the AI technology?
  • If it’s not produced by the party responsible for the AI, what are the author’s credentials? Bias?
  • If it is produced by the party responsible for the AI, how much information are they making available? 
    • Is information only partially available due to trade secrets?
    • How biased is they information that they produce?
Objective
  • What is the goal or objective of the use of AI?
  • What is the goal of sharing information about it?
    • To inform?
    • To convince?
    • To find financial support?
Bias
  • What could create bias in the AI technology?
  • Are there ethical issues associated with this?
  • Are bias or ethical issues acknowledged?
    • By the source of information?
    • By the party responsible for the AI?
    • By its users?
Owner
  • Who is the owner or developer of the AI technology?
  • Who is responsible for it?
    • Is it a private company?
    • The government?
    • A think tank or research group?
  • Who has access to it?
  • Who can use it?
Type
  • Which subtype of AI is it?
  • Is the technology theoretical or applied?
  • What kind of information system does it rely on?
  • Does it rely on human intervention? 

Text adapted from "The Robot Test" by The LibrAIry under the Creative Commons Attribution NonCommercial 4.0 International License Creative Commons

Mike Caufield’s SIFT technique is a reflective technique for evaluating internet sources. For new and rapidly-evolving topics like AI, there may not be many peer-reviewed articles available on your research question. 

Image of SIFT process: stop, investigate source, find better coverage, and trace claims

Image Credit: Mike Caulfield, 2019. https://hapgood.us/2019/06/19/sift-the-four-moves/ All SIFT information on this page is adapted from his materials under the Creative Commons Attribution NonCommercial 4.0 International License Creative Commons

S – Stop

The SIFT method recommends two different crucial stopping points in your internet research:

  1. When you first encounter a website or internet source, stop and consider:

    • Do you know this source?

    • Is this source reputable?

    • Are the claims on this website reputable to the best of your current knowledge?

If you can’t answer these questions, approach the source with caution as you move through the other SIFT steps.

  1. As you move through each step, stop and consider:

    • What is my purpose for seeking this information out?

    • Am I doing a quick scan of internet sources to understand the internet discourse of the day, or am I doing deep academic work?

    • Both are useful kinds of research but involve a different set of questions and practices. The first involves a shallow scan across different sources, the second requires a deeper investigation of each figure, quote, etc.

I – Investigate the Source

  1. Who is the author?
    • What are their credentials?
    • What have they previously written on this topic?
  2. What is this source?
    • Is it a newspaper, a blog, or a news aggregator website?
    • Where does this source derive funding?

F – Find Better Coverage

Read laterally across the internet and library resources about this topic.

If you are writing about the prevalence of CTE in former NFL players, do not solely rely on a press release from the NFL about the topic. You can certainly use NFL statements, but be sure to read across medical journals, newspaper sources, and peer-reviewed research articles about CTE before completely forming your research argument around one source.

T – Trace Claims, Quotes, and Media Back to Original Context

Tracing involves pulling apart a resource for claims, quotes, images, videos, etc., and finding the original source. Think of this as swimming upstream, potentially to a better source. To trace:

  • Click on any links to see what sources are being used to construct an argument.
    • Are these sources reputable?
    • Are these sources news aggregators?
    • If a new item, is the source reputable?
  • Check sources in the bibliography if present.
    • Are those sources more academic than the source you are evaluating? Feel empowered to ditch your original source and move on to the linked source.
  • Reverse image search photos used in the source.
  • Search for the original full video of any embedded media.
    • Did the source cut or manipulate the video?
    • Did the article accurately summarize the video? If not, you can deduce the bias of the article and use it accordingly in your own research.

Accuracy

When working with AI, keep in mind the following best practices for evaluation:

  • Meticulously fact-check all of the information produced by generative AI, including verifying the source of all citations the AI uses to support its claims.
  • Critically evaluate all AI output for any possible biases that can skew the presented information. 
  • Avoid asking the AI tools to produce a list of sources on a specific topic as such prompts may result in the tools fabricating false citations. 
  • When available, consult the AI developers' notes to determine if the tool's information is up-to-date.
  • Always remember that generative AI tools are not search engines--they simply use large amounts of data to generate responses constructed to "make sense" according to common cognitive paradigms.

Text from "Ethics & Privacy - Artificial Intelligence (Generative) Resources" by Georgetown University under the Creative Commons Attribution NonCommercial 4.0 International License Creative Commons

When AI Gets it Wrong

When AI Gets It Wrong: Addressing AI Hallucinations and Bias (MIT Management)

Currently a typical AI model isn't assessing whether the information it provides is correct. Its goal when it receives a prompt is to generate what it thinks is the most likely string of words to answer that prompt.  AI cannot interpret or distinguish between correct and incorrect answers. It’s up to you to make the distinction.  

It can give the wrong answer and omit information.  

Sometimes an AI will confidently return an incorrect answer. This could be a factual error or omitted information.  

It can make up false information. 

Sometimes, rather than simply being wrong, an AI will invent information that does not exist. This is known as “hallucination,” or, when the invented information is a citation, a “ghost citation.”   

It may not accurately produce its sources. 

If you ask an AI to cite its sources, the results it gives may not be where it is pulling this information.  Even an AI that provides footnotes may not provide the places information is from, just an assortment of webpages and articles that are roughly related to the topic of the prompt.  

It can interpret your prompts in an unexpected way. 

AI can accidentally ignore instructions or misinterpret a prompt.  A minor example of this is returning a 5-paragraph response when it was prompted to give a 3-paragraph response. If you’re not familiar with the topic you’re asking an AI-based tool about, you might not realize that it’s interpreting your prompt inaccurately. 

Resources for Evaluation

Fact Checking Sites

  • FactCheck.org - Annenberg Public Policy Center’s nonpartisan, nonprofit “consumer advocate” for voters that aims to reduce the level of deception and confusion in U.S. politics.
  • Politifact - PolitiFact is a fact-checking website that rates the accuracy of claims by elected officials and others who speak up in American politics. PolitiFact is run by editors and reporters from the Tampa Bay Times, an independent newspaper in Florida.
  • SciCheck - Focuses exclusively on false and misleading scientific claims that are made by partisans to influence public policy.
  • All Sides - Provides multiple angles on the same story.

Image Checking Sites

Web History Checking Site

  • Wayback Machine - Web archive that captures websites over time and can be used to verify content history and edits.