Your checklist to evaluate AI tools for research

Trusted content

What is trusted content?

Research literature, along with the journals and publishers that support it, are trusted if they reflect three interdependent qualities; accuracy, authority and transparency. Processes that safeguard these qualities and contribute to their robustness must also be in place.

Why does trusted content matter?

When AI tools rely on authoritative scholarly content, researchers are better equipped to verify claims, evaluate evidence and make informed decisions. Trusted content can also help mitigate the risk of misinformation, outdated findings and unsupported conclusions.

What should institutions and their libraries evaluate?

Content quality and coverage

• Is the content peer-reviewed?

• Are the content authors credible and authoritative?

• Is the content from sources that can be trusted?

• Does the tool include a strong breadth of full-text scholarly content, including journal articles and book chapters?

• Is the full text open access or subscription, or both?

• Is there broad subject coverage?

Content integrity

• Is the content updated often enough to keep up with the scientific record?

• Are retractions or corrections appropriately marked?

• When there are retractions or corrections, how frequently are they updated in the content base?

• Do the content sources have processes in place to evaluate and correct information?

Content governance

• Does the AI tool have content curation?

• Do human experts oversee the content selection or curation?

• Are there any checks on content curation, like an independent review board?

• Does the vendor provide clarity on how content is acquired and governed?

• Does content comply with copyright and licensing agreements?

Responsible functionality

What is responsible functionality?

AI tools display responsible functionality when they support transparent and verifiable work that is human-centered, scholarly and maintains credibility.

Why does responsible functionality matter?

These features make AI responses easier to verify and use responsibly, enabling researchers to control evaluation and decision-making processes. An AI tool with responsible functionality not only uses trusted content, it also provides context and guardrails so responses can be critically evaluated.

What should institutions and their libraries evaluate?

Transparency and verification

• Are responses supported by clear citations, references or source links?

• Does the tool allow users to evaluate response claims against the original sources?

• Can researchers independently verify the AI-generated conclusion using the evidence provided?

• Is it clear what steps the tool takes to mitigate bias and hallucinations?

• Does the vendor provide transparency into the search and retrieval mechanisms?

• Is there clarity around how results are weighted and prioritized?

• Does the tool show the steps taken to generate responses?

• Does the model prioritize or favor certain content types, publishers or sources? If so, is that transparent to users?

Confidence and context

• Does the tool identify uncertainty, conflicting evidence or limitations?

• Can users compare multiple perspectives or interpretations?

• Does the tool avoid presenting uncertain information as fact?

• Does the tool provide context around how the findings fit within the broader scientific literature?

Privacy and security

• Is there transparency around how the tool handles and retains user data? And are there policies in place that prevent the unauthorized sale or sharing of it?

• Is user data, the information users upload, or the prompts they enter, used to train models?

• Where is user data stored and how is that server protected?

• What kind of encryption is applied to user data?

Human-in-the-loop

What is human-in-the-loop?

Human-in-the-loop means consistent involvement of humans during the creation, maintenance and use of the AI tool.

Why does human-in-the-loop matter?

Human oversight is essential to ensure responsible design and use of AI tools, and the preservation of important research skills, such as critical thinking and originality.

What should institutions and their libraries evaluate?

Researcher controls

• Are there guardrails in place to keep users in control?

• Do features encourage critical evaluation of outputs (e.g., provide direct evidence of sources and support assessment of the evidence)?

• Do responses encourage critical thinking (e.g., stimulate ideas or highlight gaps and potential avenues for further exploration)?

Responsible AI practices

• Does the vendor have and implement responsible AI principles?

• Does the vendor routinely test their algorithm against evaluation frameworks?

• If so, are those frameworks robust and widely recognized?

• What kind of security certifications does the vendor hold?

Vendor governance

• Are subject matter experts involved in tool development, evaluation or governance?

• Does the system indicate when confidence in a response is low?

• Are any external parties involved in validating the technology and content?

Improving the research experience

What does it mean to improve the research experience?

As well as meeting the ‘research-grade’ criteria outlined above, AI tools must also enhance the research experience. In other words, they should help researchers progress by increasing efficiency, discovery, collaboration and productivity.

Why is improving the research experience important?

Supporting the research experience helps users get more value from the tools your institution selects. For libraries and institutions, this can translate into increased use and greater return on investment. For researchers and other users, it means spending less time navigating between tools and more time achieving goals.

What should institutions and their libraries evaluate?

• Is the tool effective at drawing insights from that knowledge base?

• Does the tool recommend further avenues of exploration?

• Does the tool help surface relevant grants or funding opportunities connected to a research topic?

• Can researchers identify other experts or authors working in similar areas?

• Can researchers move seamlessly from AI output to source content?

Knowledge development

• Does the tool have a reasoning engine for complex tasks?

• Does the tool provide visualizations to support analysis or discovery?

• Does the tool support iterative exploration and refinement of a research question or a manuscript draft?

• Can researchers easily compare different research articles?

• Can the tool support a multidisciplinary/interdisciplinary approach?

Research execution

• Can the tool help researchers develop and refine a research plan?

• Can users organize or export reports and outputs?

• Does the tool help researchers improve their research writing?

• Does the tool support multiple stages of the research workflow?

• Does the tool save users’ time?

Analyzing your results

While responses will vary, the more you’re able to answer “yes” to these questions, the more your tool is grounded in trusted content, responsible functionality and meaningful human oversight.

As needs may vary across institutions, focus on the needs of your users and your institution's goals.