A preprint server recently began penalizing authors when it was clear they hadn't read the papers they cited. One researcher, caught off guard by this, complained publicly. The response from the broader academic community was largely unsympathetic. Still, beneath all the reactions, there is an important question: at what point does the convenience of AI-assisted research tools start to undermine the basic expectation that researchers have read and understood what they cite?
The answer has less to do with which tools researchers use and more to do with how they use them.
This Problem Didn't Start With AI
Scholars have always operated on a spectrum of reading depth. If a paper is central to your argument, you read the entire thing carefully, take notes, and question its claims. If you cite a paper to establish uncontroversial or peripheral facts, you might only skim it or limit close reading to relevant portions. Researchers at every career stage make those judgment calls constantly, since time is limited, the amount of literature available is huge, and not every citation feels as though it carries equal importance. This isn't meant to excuse a dishonest practice, but to honestly describe how research usually works. Understanding this helps us see what AI has changed and what it hasn't.
AI tools have accelerated and scaled this behavior, making the difference between simply finding a paper and truly understanding it both more consequential and more visible.
So, Where's The Line?
The dividing line is drawn based on how the paper relates to their argument as opposed to which tool a researcher uses.
Papers peripheral to your work, cited to establish context or acknowledge a tangential finding, have always been read less carefully or with deep reading limited to relevant portions. That's a reasonable allocation of limited time and attention, and AI assistance is arguably an extension of that existing practice. The problem starts when AI convenience migrates into the core of a researcher's work: when a central claim, a key methodology, or a foundational finding gets cited on the strength of a summary the researcher never verified against the actual source.
That's the line. And it's one that AI makes significantly easier to cross without realizing it.
There is a subtler problem here working in tandem. The most visible failure mode is a hallucinated reference, a citation that points nowhere. But polished, confident AI-generated language can also hide a researcher’s incomplete ideas or holes in their thinking in ways that are much harder to detect. The output sounds complete; the thinking underneath it may not be. That's a problem no citation penalty is going to catch, and it's arguably more widespread than fabricated references.
Both of these problems stem from the same source: using AI to avoid engaging with the material rather than using it to better understand it. A researcher who uses AI to orient themselves within the literature, surface relevant debates, and decide what to read closely, is using the tool effectively. A researcher who uses it to generate citations for unchecked claims is doing something categorically different, regardless of whether the resulting references happen to be real. The professional responsibility for what ends up in a paper doesn't transfer to the tool.
What Verification Infrastructure Changes
AI models are built to produce fluent, confident-sounding output, not necessarily to verify whether that output is accurate. A convincingly formatted DOI that leads nowhere is worse than no citation at all. And even when a citation is technically real, an unverified summary can still misrepresent what the paper actually argues.
Some researchers are already using AI in the following way: uploading a paper and asking the model to verify that what they've written accurately reflects the source. That's a genuine inversion of the problem, and it points toward what responsible AI-assisted research can look like in practice.
This is where tools made for discovery and analysis, not just text generation, do something meaningfully different. Scite's Smart Citations index over 1.6 billion citation statements drawn from the full text of published research, classifying each citation as supporting, mentioning, or contrasting the cited work. That context informs the practice of reading rather than replacing it. If a researcher sees that a paper's central findings have been widely supported, or that a specific methodology has been repeatedly contested, they can make a more deliberate decision about how carefully to engage with the source material. It keeps the human genuinely in the process rather than nominally so.
The Librarian's Role In Drawing The Line
AI hasn't replaced librarians. Instead, it has made the work that libraries do even more important.
Issues like citation integrity, AI literacy, and responsible research all depend on the resources and knowledge that libraries offer. This raises a broader question of standards, and libraries have long been responsible for helping establish and maintain them.
The old assumption, that “cited” means “read closely and understood,” was already under pressure before LLMs arrived. AI has made that pressure visible.
So, keep talking about where this line falls. Update research guides to address AI-assisted citation specifically. Equip faculty and teams with a framework for where and when AI assistance is appropriate. Guide researchers to use tools that support verification, not only retrieval.
The standards around how researchers use AI won't define themselves.
