Editor’s Note: Today’s post is by John Frechette and Heather Pierce-Lopez. John is co-founder and CEO of moara.io, and adjunct instructor at George Washington University and Stevenson University. Heather is the Research and Outreach Librarian at New England College of Optometry (NECO) in Boston.
If you spend any time on LinkedIn or X in the scholarly communications space, you’ve seen the debate. AI is either going to revolutionize research or destroy it. Every day brings a couple of hundred new takes, or more — most of them arriving at one extreme or the other.
We aimed to have a different kind of conversation. Not a debate about whether AI is good or bad, but a more practical one: which parts of the research lifecycle should be more automated, which require more of a human touch, and why? Note the emphasis on “more” — the interesting questions aren’t about whether to automate, but how much and at which stages.
What follows is a summary of our recent discussion — one informed by Heather’s decades of experience in academic libraries and John’s background in economics, technology, and building research tools.
We’re Already Comfortable with Black Boxes
Before debating the role of AI, it’s worth pausing on a point that rarely gets recognized: we already rely heavily on automation that we don’t fully understand.
Google Scholar is a black box. No one outside Google knows exactly how it ranks results. PubMed uses algorithms that surface some studies over others. Every database has its own opaque logic, its own quirks, its own proprietary ranking factors. For years, librarians have known that Google’s core product is advertising, and search results are influenced by who pays to play. And yet, for most researchers, these tools are simply the water they swim in – accepted, familiar, and often unquestioned. The broader point is that so much of search strategy is already automated. The moment a researcher types a query into any of these engines and trusts the ranked output, they have (by necessity) delegated a significant portion of the discovery process to an algorithm.
AI tools, by contrast, generate significantly more scrutiny. Some of this scrutiny is warranted — and Heather rightly emphasized that transparency around content sourcing remains a top priority, particularly for librarians. But in our discussion, we found an inconsistency worth examining. As Heather observed: “If people are embracing natural language search, they’re already embracing AI.” Many databases researchers rely on daily have quietly incorporated AI into their backends and have rarely received the same level of scrutiny. And, what’s more, a legacy vendor retrofitting AI for an existing product won’t necessarily do it better than a company built around the technology.
That said, demanding perfect reproducibility from AI tools while accepting opaque results from legacy systems is in some respects a double standard that deserves a bit more examination.
What Can’t Be Automated: The Foundational Skills
The heart of our conversation centered on a deceptively simple question: what makes great research great? Or put differently: what research skills are genuinely non-negotiable for automated solutions?
We landed on several.
The first is curating and combining inputs. Research, at its core, is about bringing together the right data, the right literature, and the right frameworks in ways that produce new insights. In economics, John pointed out, one of the core tenets of growth theory is that we don’t really create anything from nothing — we combine existing things in more interesting ways. Other fields of study work the same way. An AI can surface materials, but innovation around what belongs together and why remains a human function.
The second is ethics, risk management, and long-term institutional planning. Heather, who serves as the copyright officer at the New England College of Optometry (NECO), pointed out that many publishers now include disclaimers prohibiting the upload of their content into AI tools. Meanwhile, the distinction between open-circuit tools (which only process data you provide) and those that pull from the broader internet is widely misunderstood. For institutions in the medical field, where HIPAA and FERPA compliance are non-negotiable, these concerns often determine outright whether a tool can be adopted.
The third, and perhaps the most difficult to define, is connecting research to local institutional needs. A librarian or experienced researcher brings years of accumulated context: knowledge of historical shifts in a field, relationships built at conferences and through collaboration, and an understanding of what matters to a specific community. These are the product of lived professional experience — not the kind of thing you can prompt an LLM to replicate.
We compared this to the “vibe coding” phenomenon in software development. When people without foundational coding knowledge use AI to build applications, IT departments end up cleaning up the mess. The parallel in research is clear: without information literacy, the outputs of AI tools can be misleading at best, and dangerous at worst. The tool is only as good as the person using it.
The Benefits We’re Not Talking About
One of the patterns we noticed in the broader discourse is an overwhelming focus on the downsides of AI in research and not nearly enough attention to the benefits — particularly the ones that are harder to see.
Consider publication volume. Yes, the growth in published research raises concerns about quality and the strain on peer review. But more research is, on the whole, better. The fact that people who previously lacked the time, resources, or language fluency to participate in the publishing process are now doing so is worth celebrating — rather than dismissing the entire shift as likely slop.
At NECO, where a team of three librarians serves the entire institution, AI tools have been essential for doing more with less. Tasks that might be considered busywork — organizing, cataloging, initial literature searches, etc. — can be partially automated. This frees up time for the work that genuinely requires human engagement: connecting with students, developing programming, and conducting outreach with faculty. John discussed this as the primary motivation for building moara.io. For small teams operating under real resource constraints, the efficiency gains are what make the workload possible.
Then there’s the accessibility argument, which we think deserves far more attention. For researchers with ADHD, for example, the hardest part of engaging with a paper can simply be starting. An AI-generated summary can clear that initial hurdle and make it possible to engage meaningfully with work that might otherwise go unread. During our conversation, John offered a concrete example: a colleague posted a working paper on LinkedIn. Without AI, he would not have read it, or at most, a quick skim of the abstract. Instead, a summary made it possible to develop and write a thoughtful response, sparking a discussion that wouldn’t have happened otherwise.
The same logic extends to international researchers working in a second language, early-career scholars without deep networks, and professionals outside academia who want to engage with the literature but lack the training to navigate it efficiently. If we believe in accessibility, we should be honest about how much of current research is exclusionary by default.
Downstream: Evidence Synthesis and Writing
Where things get more nuanced is in the later stages of the research lifecycle: evidence synthesis and writing.
AI tools are increasingly good at identifying patterns across large bodies of literature. They can surface clusters of findings, highlight areas of consensus or disagreement, and even pull in gray literature that traditional databases might miss. But identifying patterns is not the same as interpreting them. A librarian or domain expert will know, for instance, that a particular subfield underwent a methodological shift a decade ago that recontextualizes earlier findings. That kind of insight comes from direct professional experience.
On writing, we were both clear: the worry that AI peer-reviewed a paper that was AI-written sounds absurd on the surface, and in some cases, it is. But peer review itself has well-documented human biases — reviewers who haven’t eaten lunch, who hold grudges in certain areas, who are simply having a bad day, etc. To suggest that peer review shouldn’t be more technology-enabled, at least in part, strikes us as reflexive and a missed opportunity.
The extreme case is worth imagining. In a future where literature discovery, data collection, statistical analysis, and initial drafting are overwhelmingly automated, a researcher’s day might look less like reading and writing and more like managing, prompting, and auditing AI workflows. The best analogy may be financial analysis, where analysts no longer grind through company reports late at night, but instead tweak algorithms and find new ways to feed data into their models. Even in that world, we believe there will still be plenty of innovation, plenty of human judgment required, and plenty of discussions like this one.
“It’s here to stay, so let’s be at the front of it. Because if we’re left behind, we have no say.” – Heather Pierce-Lopez
What Should Change
Rather than ending with platitudes, we’ll close with what we think institutions, publishers, and researchers should actually do differently.
First, try the tools. Not one tool, but several. Approach them with a clear sense of what you want to accomplish and where in your workflow you’d like more time back. NECO’s approach has been to ask faculty a straightforward question: where in your academic or clinical life do you feel bogged down by repetitive tasks? Start there.
Second, build assessment frameworks. NECO developed an AI charter and is creating an evaluation tool to ensure that any platform adopted aligns with the institution’s principles, ethics, and compliance requirements. The goal is to have a structured way to make that decision, not a blanket verdict on AI.
Third, invest in AI literacy as the new information literacy. Students entering health sciences, law, and every other field will encounter AI tools in their practice, whether their institutions prepare them or not. The choice is between guiding that adoption or being left behind by it.
Fourth, rethink how we measure research outcomes. If publication volume is about to increase dramatically, the metrics we use to evaluate research need to evolve accordingly. Impact metrics, reproducibility standards, and evidence of genuine contribution (all of which are becoming more available) should matter more than the name of the journal.
Finally, redesign curricula thoughtfully. Early evidence suggests that AI-permissive classroom policies can produce genuinely interesting outcomes. Heather shared an example of a professor whose students are engaging more deeply with course material precisely because AI tools made it more accessible — they could listen to it, interact with it, explore it in new ways. The goal shouldn’t be to remove all restrictions from learning. Some assignments should deliberately require students to engage with papers directly, just as some exams limit the use of notes. But blanket prohibitions are increasingly hard to justify.
The discourse around AI in research is loud, but too much of it is binary. The more productive way forward is to think carefully about where on the spectrum each task falls, to be honest about the limitations of both humans and machines, and to build the frameworks that help us navigate this transition with purpose. As Heather summarized it: “It’s here to stay, so let’s be at the front of it. Because if we’re left behind, we have no say.”
