By Brittne Kakulla, PhD | AARP | Member, Insights Association IDEA Council
When the Stacks Disappear
There is an old problem in research methodology called “the file drawer problem”. The idea is simple: not all research gets published, for a variety of reasons. Studies with inconclusive findings rarely make it through peer review. Research on underrepresented populations often goes underfunded before it even begins. Work produced outside of well-resourced universities struggles to reach the journals that shape a field, regardless of its quality. Findings that challenge the dominant narrative or complicate a tidy story end up in the back drawer of a file cabinet. The work was conducted carefully. The data were real. But these gaps reflect biases that mean the literature we see is never a complete picture to begin with.
Early in my graduate training, my method’s professor introduced this problem. He argued, when you conduct a literature review, you are not just finding sources. You are making decisions about whose work you prioritize and whose you do not. The absence of a finding is itself a finding.
I am dating myself here, but I took that lesson with me into the library stacks (and the microfiche and finally to EBSCOhost!). When I found a relevant article, I would look at the table of contents of that journal issue to see what else was published alongside it. Often, similar topics were deliberately grouped together by editors, so I could see other perspectives on a topic. That context helped me understand not just what a study found, but what larger conversation it belonged to, and what I could be missing.
The File Drawer Problem Goes Digital
We do not browse stacks anymore. Artificial Intelligence (AI) can find connections across thousands of sources faster than any human reader. But fast does not mean complete. The sources an AI tool retrieves are still shaped by what has been published, indexed, and made digitally visible.
Nearly all major AI tools now offer some form of "deep research." Deep research is an advanced, AI way of investigating a topic or question that searches, reads, and synthesizes information across many sources to produce a structured, evidence-based report rather than just an answer. But the depth of the search is only as good as the direction you send it in.
When we hand the search entirely to the tool, we can lose that layer of context. Perspectives that are present but not prominent can disappear, not because the tool is careless, but because the algorithm is doing exactly what it was designed to do: follow the path of greatest visibility. The file drawer problem has not gone away. It has just gotten harder to see, because AI-generated searches and literature summaries look authoritative. A thorough-looking report can still reflect a narrow slice of what actually exists.
For researchers working with topics or populations that are underrepresented in mainstream literature, that is a real risk. And it starts before the search, in how we frame the question. We must intentionally stop and ask whose perspective could be missing, or what approach am I overlooking? Researchers must acknowledge the digital file drawer problem, because the tools will not do it for us. On LinkedIn, I have been exploring prompting patterns through a researcher's lens. I am bringing that series here each quarter, pairing one AI tool with prompt patterns that researchers can put into practice to harness the value AI can add to their workflow while understanding its limitations. This is the second in the series.
Tool Callout: Perplexity Deep Research
What it does: As noted earlier, nearly all major AI tools now offer some form of “deep research”. What keeps me coming back to Perplexity is that citation is its DNA. It was built from the ground up as a search tool, not an AI assistant that added search as a feature. That origin matters in practice: every claim in a Perplexity search and deep research report links directly to a traceable source, which makes verification faster and makes it easier to notice when your sources are clustering around the same institutions or perspective.
Research use case: Deep research is a great starting point for landscape scans and literature reviews at the front end of a project. It can help you understand what is already known about a topic, identify key debates in a field, and quickly surface themes, gaps, and opportunities that a standard search might miss. The structured, cited output also makes verification more straightforward.
The benefit: Depth and synthesis at speed. A literature scan that might take a researcher a full day can be turned around in minutes, with a report organized well enough to actually use.
The limitation: Deep research follows the path of your original question. If that question reflects dominant-narrative assumptions, the tool will go very deep on a narrow path and produce a thorough-looking report that still misses whole bodies of work. The output looks complete, and that is part of what makes the bias harder to catch. It also draws primarily from what is indexed and digitally accessible, which means some research can be systematically underrepresented regardless of how deep the search goes.
Prompt Pattern: The Alternative Approaches Pattern
One technique that can help address this limitation is a prompt pattern called Alternative Approaches.
What it does: The Alternative Approaches Pattern can counter researcher bias before it shapes a search. The Alternative Approaches prompt pattern helps you to look at the same data or question through different lenses that you identify. Before committing to a single search path, you ask deep research to generate multiple different ways to approach the topic (3 is ideal).
The pattern works on the assumption that our default framings reflect what we already know and who we already study. By requiring the tool to surface alternatives first, you create a checkpoint that can expose blind spots before they determine what evidence enters your analysis.
Research use case: This pattern is useful when you suspect your default framing may be too narrow but you are not sure where the gaps are, or if you want to see your topic from a different perspective. By seeing several approaches side by side, you can make a deliberate choice about which path to follow, or instruct the tool to pursue more than one.
Why this prompt pattern works: Deep research is built to follow a structured path. The Alternative Approaches prompt inserts a decision point before that path is set, giving you more control over where the depth gets applied.
Limitation: This pattern generates options but does not guarantee those options are comprehensive. A researcher who already knows the field will still need to evaluate whether the alternatives generated are genuinely distinct or just variations on the same narrow frame.
Practice Prompts
Here are two prompts you can adapt for your own research workflows.
Prompt 1: Map the search landscape before committing
"Before you begin searching, generate three different approaches to researching this topic: [insert topic]. For each approach, describe what types of sources it would prioritize, which populations it would most likely capture, and what it might miss. Then ask me which approach to pursue, or whether to combine more than one."
Prompt 2: Surface the less-indexed paths
"Identify three search paths for this topic that would be less likely to surface through a standard academic literature search: one focused on community-based or advocacy organization research, and one focused on grey literature or government reports, one focused on pop culture or media representations. Describe what each path would add to a standard literature scan before you begin. [insert topic]"
Closing Reflection
The sources we start with shape the questions we ask. A literature review that consistently surfaces the same institutions, the same populations, and the same frameworks can start to look like a complete picture, even when it is not.
AI deep research tools are powerful precisely because they go further than a single search. But they still follow the direction we set them in. The file drawer problem did not originate with AI, and it will not be solved by it. But as these tools get faster and more capable, our responsibility to ask harder questions at the start of a search only increases.
________________________________________
ABOUT THE AUTHOR
Brittne Kakulla, PhD, is a researcher at AARP focused on technology adoption among adults 50-plus. She serves on the Insights Association IDEA Council and writes about AI literacy for research professionals.
Prompt patterns come from White, J., Fu, Q., Hays, S., Sandborn, M., Olea, C., Gilbert, H., Elnashar, A., Spencer-Smith, J., & Schmidt, D.C. (2023). A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT. arXiv:2302.11382. https://arxiv.org/abs/2302.11382