Generative artificial intelligence (AI) is quickly becoming ubiquitous. Its integration is not just limited to our personal technology but now also includes library resources (see examples in “2026 Library Systems Briefing”). Meanwhile, vendors are constantly introducing new tools or integrations into existing products, and we must measure emerging AI platforms against our library’s tried-and-true resources to gauge which ones truly bring a new perspective or added value to our collections.
It’s an overwhelming time to vet and license library resources. To cut through the noise, I’ve narrowed down my queries for vendors to five questions to better understand how a new AI tool might enhance my library’s user experience.
1. What content is your tool trained on?
The future of generative AI models will heavily depend on whether the use of copyrighted content to train models violates copyright or falls under fair use. It will take years for the copyright-and-AI question to be decided, but if the courts ultimately rule that all content used to train models must be licensed, many existing models could become unusable.
While this unfolds in court, I am interested in ethical data acquisition by my vendors. AI tools typically follow one of three training approaches: training a full model, supplying a full model with a specific knowledge base, or training a small language model. Each approach to an AI model has its pros and cons. Large language models respond more predictably but aren’t trained as well for specific purposes, whereas small language models can focus on specific purposes but potentially sacrifice accuracy.
The vendor’s response to this question also provides insight into the ethics of their model. Ideally, all content should be properly licensed and clearly attributed. If user inputs or uploads are used for training, vendors should clearly notify users and warn them of the potential for copyright infringement. A vendor that is unwilling or unable to address how their model is trained and how they source model content is cause for concern.
2. What sets your tool apart from other vendors’ or from freely available models like ChatGPT?
While there are many capable generative AI tools on the market, there is a limit to the uniqueness and value that they can offer library users. This is especially true in a world of constantly shrinking budgets where content redundancy no longer makes sense for many libraries.
Whether you love or hate AI, there is an undeniable appeal in a general-purpose tool that can help with a range of everyday tasks. Kickstarting your research in ChatGPT is questionable from a copyright lens, but I guarantee that library users are already doing so. Well-trained models with sufficiently sized corpuses can already help users navigate research tasks, such as literature reviews. These tools may not have direct access to our licensed content, but they can help construct meaningful keyword searches, identify relevant databases, and summarize seminal research without specific content. When evaluating newly available AI resources, libraries must carefully consider the incremental benefits of a licensed tool and whether they outweigh the free functionality users can access elsewhere.
Fundamentally, a library-licensed model needs to bring true novelty to justify adoption. Resources that operate in limited environments, such as a model that can only return information about a single database, risk frustrating users when they are unable to answer their questions completely. Moreover, limited tools may devalue the library’s resources by conveying to patrons that library resources are less useful than free tools.
Patron privacy is a core value of librarianship. Users trust that resources provided by the library protect their best interests and that their data is safeguarded. Any adopted AI models or tools must reflect this trust.
3. How does your AI tool or platform protect the privacy of users and the intellectual property of other vendors?
Users are already overly comfortable sharing their personal information with AI models—regardless of whether or not those models are diligent about protecting it. Data privacy is rapidly evolving in a post-AI world, but users assume that most of their conversations will never see the light of day and will stay locked behind logins that they set themselves.
Patron privacy is a core value of librarianship. Users trust that resources provided by the library protect their best interests and that their data is safeguarded. Any adopted AI models or tools must reflect this trust. Our vendors must provide transparent guidelines on how user data is captured and used, especially if it is being used to retrain their tool. In a perfect world, user data is not captured by a model, remains anonymous, and cannot be used for training (i.e., a closed model). Otherwise, resources should clearly inform users of any saved data and its intended future use.
4. What do librarians and users need to know before engaging with your AI tool or platform?
There is always some level of onboarding for new library resources. Some teams will have the bandwidth to dive deep into each new tool, while some libraries may only have a few minutes to review it before they debut it for patrons. If you cannot arrange a trial for an AI platform, your conversation with the vendor is the best measurement of how painless adoption will be for your library.
AI should streamline or automate something, so platforms that are difficult to navigate, require unintuitive engagement, use unusual jargon, or return unreliable results may be deal breakers. I am wary of new resources that purport to return excellent results but are difficult to use, because they not only require a time investment from library staff to learn but also result in frustrated users. Platforms don’t have to be perfect out of the box, but I am cautious of a tool that may add minimal value and maximum stress and workload on library teams.
Fundamentally, a library-licensed model needs to bring true novelty to justify adoption.
5. What is the vendor’s long-term AI strategy?
In my experience, few vendors have an iterative multiyear plan for releasing, updating, and expanding their AI tools. The proliferation of AI is unsustainable in more ways than one, with continuous model growth and year-over-year costs of particular concern to libraries. For example, it is estimated that AI will run out of training data sooner rather than later, hindering continuous model improvement. Models may not drastically improve with time or improvement may be slow. If a platform doesn’t meet your expectations now, it may not be able to meet them later, either.
Moreover, if each library vendor platform rolls out an AI tool at an introductory rate or begins to build in AI components that can’t be separated from overall platform costs, these resource costs will quickly outpace library budgeting. Many of my library’s current resources have incorporated AI components at no additional cost. However, these integrations sit inside of the preexisting platform, and few vendors have been able to speak to whether those integrations will continue freely in the coming years. It is not a stretch to imagine that some library vendors will pivot to a platform with a non-optional AI component and then charge a premium for their AI-powered resources.
Ultimately, vendors may not have the perfect answers, but asking the right questions positions librarians to dig deep and advocate for their communities. AI will continue to evolve at a pace that libraries cannot always match. Approaching these tools with informed curiosity allows us to separate genuine innovation from fleeting hype and ensure that the resources we license continue to serve users ethically, responsibly, and well.
A version of this article was originally published on Choice 360 LibTech Insights on October 20, 2025.
