Can AI Help Researchers Work Faster Without Weakening Critical Thinking?

AI cannot teach you how to critically consume information. You have to understand how to evaluate information independently before you start using AI as an adjunct.

Stephanie Soultanian, Clinical Psychology PhD Candidate at Adelphi University

Artificial intelligence is becoming part of the daily machinery of academic research. Researchers use it to sort sources, summarize papers, draft outlines, refine prose, and move through large volumes of material more quickly. For universities and research institutions, the appeal is clear, with AI promising speed at a time when scholars face expanding literatures, rising publication pressure, and growing demands to work across disciplines.

Speed, however, is only one measure of academic work. As experts know, research also trains a way of thinking. Scholars learn how to assess evidence, recognize weak claims, identify missing context, and understand why one interpretation is stronger than another. These habits develop through reading, comparison, revision, and disagreement with the material. And although they are harder to see than a finished draft, they are often what give the draft its value.

This makes AI adoption more than a productivity question. A source map, literature review, or summary may appear useful, but its value still depends on the person reading it. The researcher has to know what the tool has missed, simplified, overstated, or arranged too neatly. Without that ability, a faster process can create the appearance of progress while leaving the harder work untouched.

To better understand the issue, we spoke with Stephanie Soultanian, a clinical psychology PhD candidate at Adelphi University, to gain the perspective of a researcher whose work depends on close interpretation. As a PhD candidate, she examines how linguistic features in therapist interviews can reveal psychological data, particularly regarding the therapeutic relationship and the end of treatment. This kind of work requires attention to language, context, and nuance. And through her research, she has gained perspective on how AI can support academic research without weakening the judgment inquiry is meant to build.

Stephanie Soultanian is a PhD candidate in clinical psychology at Adelphi University, with a background spanning healthcare operations, community mental health, education, and client-centered care. Her work sits at the intersection of clinical psychology, health systems, and organizational leadership, shaped by more than a decade of experience supporting patients, providers, and care teams.

Previously, she spent more than a decade at My Doctor Medical Group, where she held roles in patient care coordination and practice management. She also has experience in community health, legal advocacy, literacy mentoring, and behavioral research, including work with the Berkeley Free Clinic, East Bay Community Law Center, Berkeley United in Literacy Development, and UC Berkeley’s Lucia Jacobs Animal Behavior Laboratory. She holds degrees in psychology and environmental economics & policy from the University of California, Berkeley, with a minor in education.

For many researchers, AI’s clearest value is not in producing the final argument. It is largely in helping organize the work that comes before it.

Academic research often begins with an unwieldy mass of material: journal articles, competing theories, adjacent literatures, methodological debates, and notes that lack structure. Before a researcher can make an argument, they have to find patterns in that material. AI can assist at this stage by sorting sources, identifying preliminary categories, and shaping a body of research that still requires human review.

Soultanian describes her own use in those terms. She has used AI “to compile [her] sources for my literature review” and “to do deeper research,” particularly to “compile all of the studies into groups” and “categorize them” as she works through the literature.

That kind of support can be useful, especially in fields where the volume of relevant material can quickly exceed what one person can comfortably hold in mind.

“There’s information overload,” she says. “There’s so many things to evaluate all the time, it’s impossible to do it.”

At this stage, the tool has only prepared the material. The researcher still has to decide what the structure means. A set of categories may help someone begin, but it does not settle which findings deserve more weight, which assumptions should be questioned, or which connections are intellectually meaningful.

This is where AI’s role is easiest to defend. Used carefully, it can reduce the administrative burden around scholarship and help researchers move from disorder to structure. But once that structure begins to stand in for interpretation, the question changes. The issue is no longer whether AI can help manage research. It is whether the researcher has enough independent command of the material to know when the tool is wrong, incomplete, or too confident.

The risk begins when AI-generated structure starts to stand in for the researcher’s own understanding. A tool can help organize a field, but it cannot decide whether that organization is accurate. It can surface possible patterns, but it cannot know whether those patterns are meaningful. It can produce a synthesis, but it cannot tell the researcher whether the synthesis reflects the field or just sounds plausible.

In this regard, Soultanian draws the line at critical consumption. “AI cannot teach you how to critically consume information,” she says. Before using it as support, researchers need to know how to evaluate information independently: “You have to understand how to evaluate information independently before you start using AI as an adjunct.”

Her point is not simply that AI makes mistakes. Instead, the researcher needs a basis for recognizing those mistakes when using AI. In research, independence includes the ability to assess whether a claim is credible, whether a source belongs in the conversation, and whether a pattern identified by a tool is actually meaningful. Without that foundation, a researcher may accept the tool’s categories and output before noticing what they exclude.

Hallucination makes the problem more difficult. Researchers who know a field well can often recognize when a citation sounds wrong, when a claim is too broad, or when a summary leaves out a major debate. Newer researchers may not have that same protection. For Soultanian, “being able to detect hallucinations and really understand the landscape of the field that you’re looking at working within” is “really important.”

Her concern extends beyond individual errors. As AI becomes more common in classrooms and research settings, she worries that younger scholars may become dependent on tools before they have developed the habits needed to evaluate them. With “the issues of hallucination,” she says, “that’s the fear for me, that… we’ll lose the ability to critically consume information.”

The sequencing matters. Researchers need sufficient disciplinary knowledge to question the machine before asking it to help. Otherwise, the tool does not simply accelerate research. It can train users to accept fluency as authority.

The subtler risk is not that AI always sounds wrong, but that it can sound factual and conclusive when its output is polished fluff.

A response can be orderly, readable, and broadly accurate while still missing the texture of the field. For a researcher, that may be more dangerous than an obvious error. A false citation can be checked but a shallow AI output that is engineered grammatically to be persuasive is harder to catch.

Soultanian is especially concerned by output that appears useful because it has the surface qualities of a good answer. AI can “generate super generic responses a lot of the time,” she says, “but it sounds really good.”

The problem, in her view, is how easily users recognize pieces that appear credible. “We identify with the bits of information that sound good and that seem like they check out,” she says, “but in reality it’s not giving you the full picture.”

The gap between sounding right and being sufficient matters in academic work. A shallow synthesis may be useful as a starting point, but it can be misleading if it gives the impression that the main work has been done. The danger is not only that AI invents information. It is that it can make partial understanding feel complete.

The output, Soultanian reinforces, can also be “skewed” or “surface level.” As a result, overreliance on AI could cause researchers to miss “a lot of the depth” that comes from their own minds making connections through past experience, prior reading, and time spent with the material.

Depth is not easy to measure, but it is central to scholarship. Researchers develop instincts about what matters, what does not fit, what deserves another question, and what earlier work may have overlooked. Those instincts are built through repetition and friction. They often come from reading too much, getting stuck, changing direction, and returning to the material with a better question.

AI can reduce some of that friction, and sometimes that is useful. But if it removes too much too early, it may also remove the conditions under which original thinking develops. For Soultanian, that risk reaches the creative dimension of research, and as she thinks, “The creativity that is inherent in you or developed within you can sometimes be thrown off with AI.”

What gets lost, in that case, is not only accuracy. It is the chance for a researcher to arrive at an unexpected connection before a tool offers a ready-made path.

Responsible AI use in research starts with clearer boundaries around the work a tool should and should not perform. In Soultanian’s account, AI has a place in the process, but that place is limited. It can help organize sources, manage information overload, and refine early work. It should not become the source of the researcher’s interpretation.

Those boundaries are practical. If she were managing a research team, Soultanian would be comfortable allowing AI “for organizational purposes” and sees it as useful for “editing and organization.” The hesitation comes when the tool moves closer to authorship. She would be more cautious, she explains, about using AI for “the actual work product generation part” unless the researcher’s own understanding is already clear.

That standard places responsibility back on the researcher and the institution training them. Before AI becomes part of a project, researchers still need to show that they can think through the work without it. Soultanian would want to see “samples of their work without using AI” and would want to “talk through their research” to make sure they “really understood” the theoretical pieces behind it.

That may be the simplest test for responsible adoption. The researcher should still be able to explain the field, defend the choices, recognize weak claims, and revise the argument when the evidence requires it. The tool can assist the process. It cannot carry the intellectual responsibility for the work.

Ultimately, when used well, AI can give researchers more room to think by reducing the burden of sorting and structuring. Used poorly, it can make unfinished understanding look complete. The difference depends on whether universities and research teams treat AI as a shortcut to output or as a limited tool within a larger culture of judgment.