Editor’s note: We welcome a guest blog post from Jay Singley, Document Delivery and Circulation Desk Manager at North Carolina School of Science and Mathematics.
In March 2025, Ex Libris unveiled their AI-powered Research Assistant tool for institutions using Summon. Within a week, Summon users reported error messages with specific search terms and topics. The first error reports shared in a listserv for Summon users contained “Tulsa race riot” and “Tulsa race massacre.” Test searches by reference librarians and systems librarians containing these search terms generated no results.
I work in an academic library in a public state university system serving high school students and have since turned on Research Assistant in our Summon Preview Environment (a beta testing environment hidden from users). I have begun testing terms and topics systematically using a methodology akin to Matthew Reidsma’s auditing of algorithms, which seeks to expose the implicit biases of opaque technological systems.
In addition to “Tulsa race riot” and “Tulsa race massacre,” my audit of “controversial” search terms conducted in March-April 2025 turned up error results for the following:
- Genocide in Palestine
- Gaza war
- Rwandan genocide
- Armenian genocide
- Genocides across the world
- History of genocides
- lynching
- lynching in the united states
- lynchings in the united states
- january 6
- covid
- covid data
- COVID-19
Ex Libris responded to initial reports about “Tulsa race riot” and “Tulsa race massacre” error results with the following:
“This is due to safeguard policies enforced by our AI service provider to support ethical and responsible AI use. It’s not something the Primo or Summon applications control.
If a query does not return a result, please try different phrasing. For example, ‘tulsa black wall street’ will return results that are directly related to the race riots.”
Despite requests for more information in the same listserv, Ex Libris has not provided a comprehensive response to librarians’ concerns about who the third-party provider is, what the ethical and responsible AI use policies are, whether local control can be made available, what known search terms and topics are obstructed, and more. As one librarian aptly noted, academic users want and need access to information that may be blocked by AI use policies meant for the general public.
An essential note is that Summon’s Research Assistant is not limited to an institution’s catalog. Instead, the AI tool searches the entire Central Discovery Index of Ex Libris’ available records regardless of whether the material is actually available to the user or not. (As of June 2025, Ex Libris has provided an option to limit search results to a user’s institutional catalog.) When the AI generates an error message (or perhaps more aptly refuses to run a search or share results), it is not necessarily because the material is unavailable. Rather, the third party “safeguard” policies obstruct the search to begin with.
In running test searches, I also uncovered troubling instances of suppressed results. I recently saw the Oscar-winning documentary No Other Land, co-directed by Palestinian and Israeli filmmakers. I decided to ask Research Assistant about this film. Searches for most of the directors generated the same no results error messages. A search for one of the co-directors generated partial and inaccurate results. Research Assistant made several guesses as to what the film No Other Land was about and why it was important. Could this be an example of biased third-party policies preventing searches about this Palestinian film and filmmakers? Is it also possible that Research Assistant is accurately reflecting the lack of materials related to No Other Land available in Ex Libris’ entire Central Discovery Index and widespread censorship of this film (and more broadly Palestinian people, culture, and experiences) in the United States? Are there truly no academic sources talking about this film and these filmmakers, one of whom was recently attacked outside his own home by Israeli settlers? More research is needed to clarify why “controversial” searches return error messages, inaccurate results, or partial results. The current state of Research Assistant partial results and error notices for “controversial” topics warrants more inquiry.
I know with near certainty that if a student user at my workplace received an error message when searching for “Tulsa race massacre,” they would switch their topic and probably not tell me or their instructor about the error message. To my users, “we couldn’t generate an answer for your question” translates to “your topic is not worthy of pursuing—change it.”
Uncritically adopting AI tools in discovery systems will perpetuate, if not exacerbate, existing biases and suppression of minoritized people. Try this safer topic. Try this approved topic. Try this unobstructed topic.
Libraries, museums, and information sciences are a frontline in resisting the federal administration’s targeting of minoritized people as well as their whitewashing, disinforming, censoring, and defunding tactics. In the age of AI, our access to information is not spared from these attacks. We must uncover answers to the questions Safiya Noble, Joy Buolamwini, and others elicit: Who and what is safeguarded through AI tools in discovery searches? Whose safety and comfort are prioritized? At whose expense?
Same thing (e.g. Tulsa race riot) seems to be happening for Primo Research Assistant, and I get “Error: An error occurred generating the response, the response quality might be affected.” though it seems to generate a direct answer anyway. I guess this isn’t surprising since this is a sister product to Summon RA. I suspect Web of Science RA might be similarly effected.
My guess is the LLM Clarivate is using has high sensitivity and when the LLM tries to generate a Boolean search strategy on the input it refuses and errors out.
Scopus AI, Scite assistant which use LLMs in similar ways (generate Boolean), do not have any issues as far as I can tell.
EBSCOhost Natural language Search behavior is odd. With Singapore Race Riot, it expands to Singapore AND race AND riot. But with Tulsa Race Riot, it does not show this expansion…
Yet Proquest Research Assistant “suggested terms” which the docs says uses LLM to suggest terms seem to work, despite also being under Clarivate.
But I think such binary yes/no is fairly easy detect where the LLM just refuses to generate and you might with luck see it.
But LLMs are now increasingly used to decide how to search and even directly assess relevancy with agentic/deep search e.g. OpenAI Deep Research and their academic counterparts. If they have any unacceptable bias in the way they judge relevant this can be even more difficult to detect.
Based on comment here https://scholarlykitchen.sspnet.org/2025/05/08/guest-post-eight-hypotheses-why-librarians-dont-like-retrieval-augmented-generation-rag/#comment-138941 seems to be related to Azure OpenAI content filtering?
Aaron, yes, I traced it down to Azure OpenAI content filtering and read the docs. There are various levels you can configure for various subjects.
On the Summon listserv, I asked:
> I read through the Azure docs and made some notes, the most relevant of which is:
> They have a set of different areas that people can configure:
> Hate and fairness
> Sexual
> Violence
> Self-Harm
> Protected material (2 types)
> User prompt attacks
> Indirect attacks
> And it seems configuration levels can be set at low, medium, and high for each. Which ones does Summon use?
… I did not get an answer.
Hi Ruth, I suspect I know the answer already, but did you ever get any response to this? We’re Primo users and I’m still seeing the same ‘We couldn’t generate an answer for your question.’ responses for these queries.
I’m on the IGeLU Primo Working Group so intend to raise this there as this is fundamental information that should be made available to customers.
I’ve been experimenting with this in Primo’s Research Assistant, and I found something fascinating. Yes, as stated, the AI does not like to provide responses to those keywords listed. But, I tried a number of variations. It turned out that it was answering any number of *questions* I asked on those topics. Strange, I thought… what’s the difference?
It hit me, then, that this AI did not like responding to keyword searches. And to test that, I tried searching “cat” with no answer. Same with “bicycle.” But when I searched for “How are cats and bicycles related?” it gave me the best answer it could to a nonsensical question. I likewise got an answer to the search “cat and bicycle.”
Similarly, if I searched “COVID,” no answer. But a search for the potentially more contentious topic “Was COVID created in a lab?” actually gave me a nuanced response. Same thing with “January 6.” Nothing there, but I got a nuanced answer for “What happened on January 6?”
I’m not making any claims that this platform is either biased or not biased, and I have no clear evidence to explain what is happening. But some experimenting with it anecdotally seems to indicate that it prefers questions over keywords, and I wonder if that issue is the cause of all this?
a search for tulsa riot race in goolge search shows thounsands of results.
(same for other sensitive searches)
why stick to *** search tools which censor argomenti sensitive subjects?
it gives the idea that there are no ways to reach for the information