AI can help with academic research but doesn’t always get it right – here are some tips to check the information you receive

Almost 25% of adults in the United States are using artificial intelligence chatbots on a daily basis, a 2026 Pew Research Center study found.

At the same time, another recent study found that 48% of the information that popular free chatbots gave in response to questions about the news contained errors.

As librarians, we know that if students are using AI chatbots to help with their academic research, they shouldn’t assume that the information they’re given is accurate.

Not all AI is the same. Here are four steps we recommend to question information AI generates, including by checking the original source and considering issues that aren’t always easy to spot.

1. Find the best fit

First, we think it makes sense to spend time choosing the right AI tool for the task at hand.

AI tools – be it ExLibris’ Primo Research Assistant, Research Solutions’ Scite, Anthropic’s Claude or OpenAI’s ChatGPT – are trained with various datasets and use different algorithms to generate results.
They should not be used interchangeably.

Some tools might be better or worse at answering specific kinds of questions, and the types of answers they generate can vary.

People might choose ChatGPT, for example, at the beginning of their research process, because it can provide overviews and summaries to help with understanding concepts.

In subsequent stages, they might choose a tool such as Primo Research Assistant or Scopus AI, which can help find and examine peer-reviewed research that has been vetted by academics.

Students conducting academic research could also see whether their library subscribes to an artificial intelligence tool like LeapSpace or Concensus, which draw their data from peer-reviewed, scholarly research.

Using these specific AI tools doesn’t eliminate the need to evaluate their answers, but it provides more certainty that the data has gone through a review process.

Tools such as Anthropic’s Claude, meanwhile, may also draw from academic research, but to what extent is unclear. According to Anthropic, Claude states that it is trained on a proprietary mix of publicly available information from online sources, public and private datasets, user data and synthetic data generated by other models.

Finally, it can be helpful to read librarians’ reviews about AI tools to help select the best fit. People can also take a simple online test such as ROBOT to help them consider which tool to use.

2. Ask more questions

Second, asking follow-up questions is an important part of evaluating the answers AI tools generate. Even if the answer an AI tool gives sounds good, it is useful to take time to learn more about it.

Checking one’s own emotions is also important, because AI chatbots have been known to flatter users by giving answers based on the user’s own biases, or what the user wants to hear, rather than just facts.

Ask the AI to tell you where it got the information, how confident it is that the answer is correct and what might be missing. Ask the AI why it believes the answer is correct, or ask it to explain how it came up with an answer.

It’s helpful to also make sure your prompt is phrased in a neutral way. For example, asking “What does psychological research say about social media and anxiety?” rather than “Tell me why social media causes anxiety” may produce different responses.

3. Check the source

Third, it is useful to trace any claims the AI chatbot makes to their original sources.

Students may be familiar with using the SIFT process to evaluate websites. This process asks people to stop, investigate the source, find better coverage and trace claims to their original source.

Start by asking the chatbot to cite its sources, if it did not do that already.

Then, go to each source and read enough to figure out whether the AI is using the information in the right context. Also, look into who authored the source and whether they have expertise in the topic. If the AI won’t cite the sources it used, or if the sources it cites don’t include the information it said they did, pull out any claims it is making and do further research.

Can you find other, credible sources that back up what the AI is saying, using Google or a library database? Can you figure out where the AI may have gotten the information and whether that source is credible?

If this seems like a lot of work, you could always go directly to library databases instead of using AI tools.

4. Not all issues are easy to spot

Fourth, there are other, less visible issues that can play out when using AI tools.

For example, it might not always be easy to tell whether an AI tool is missing important perspectives or is giving biased information.

Remember that you, not the AI tool, are the researcher. It might be easy to rely on an AI tool to generate a summary of important sources, but deciding for yourself what sources to use helps you learn about a topic more deeply.

Research shows that most AI tools focus on sources such as websites, internet forums and other online content that is freely available and written in English. Consider what expert knowledge might be missing from AI datasets and seek out those viewpoints when needed.

Last, remember that you need to cite the AI tool you used if you are quoting or paraphrasing text generated by it in your work, just as you would cite any other source.

Artificial intelligence is changing fast. As librarians, we think it is important to continually evaluate whether the tools you are using are still working for you.

These tips are a good place to start to help sharpen your evaluation skills.