Session Overview

The session covered four practical use cases for AI in research: transcription, literature discovery and review, data analysis, and writing and editing. Before getting to tools, it is worth being aware what can go wring with generative AI. Hallucinations, sycophancy, knowledge cutoffs, privacy risks, and hidden biases are not theoretical concerns. They are daily realities for anyone relying on AI output, and understanding them is the prerequisite for using AI responsibly in research.

BoodleBox: DePaul Driehaus’s Institutional AI Platform

BoodleBox (boodlebox.com) is a multi-model AI platform that gives users access to several leading AI systems—including ChatGPT, Claude, Gemini, and Perplexity—within a single interface and conversation thread. For researchers, its most significant features are FERPA compliance (meaning student and research data is protected under federal education privacy law) and SOC 2 certification, an independently audited security standard that verifies how user data is handled and stored.

Beyond security, BoodleBox offers a range of workflow tools. Students and researchers maintain a persistent portfolio of their work, even on free accounts. Custom bots—pre-configured AI assistants for specific tasks, equivalent to ChatGPT’s GPTs or Gemini’s Gems—can be built and then chained so that output from one flows directly as input into the next. Faculty can view how students are using the platform. Real-time web data can be piped into any prompt through Tavily integration, and Semantic Scholar provides access to academic literature with citations.

bot-stacking an be used to reduce hallucinations: run a research query in BoodleBox, then append Perplexity to the end of the prompt as a secondary fact-check step. The two-model approach catches errors that a single model might miss.

Generative AI: Eight Issues Every Researcher Must Understand

Generative AI—meaning AI systems that produce new text, images, or other content in response to prompts, using statistical patterns learned from large-scale training data—offers genuine value for research workflows. It also comes with a specific set of failure modes that matter more in academic contexts than almost anywhere else. The session covered eight of them.

1: Nondeterminism

AI models are nondeterministic: given the same prompt, they will produce different output each time. This is by design—it makes responses feel natural rather than mechanical—but it means you cannot simply re-run a query to verify a result. You can push toward more consistent output by adjusting the model’s temperature, a setting that controls how predictable or creative the output is. Low-temperature instructions like “be precise and consistent” produce tighter, more reproducible responses. High-temperature instructions like “be creative and generate diverse ideas” introduce more variation. For research tasks requiring accuracy, lower temperature is generally preferable.

2: Context Windows

Every AI conversation happens within a context window—the model’s working memory for that session. Every message you send, every document you attach, and every response the model generates consumes space in that window. When a conversation grows long enough that early content falls outside the window, the model loses access to it without alerting you. For research sessions involving long documents or extended exchanges, it is often better to start a fresh conversation than to continue one that may have exceeded its reliable range.

3: Hallucinations

AI hallucinations are confident, fluent statements of things that are not true. They take two forms. Closed-domain hallucinations occur when a model generates information that contradicts the source material you provided—getting things wrong from the data it was given. Open-domain hallucinations occur when a model invents information with no grounding at all—making things up wholesale, including fake citations, nonexistent authors, and fabricated statistics.

The Whisper model, developed by OpenAI to transcribe YouTube audio for training data, may invent words when audio is unclear rather than indicating uncertainty. Because the training data contains objectionable content, the invented words can be inappropriate—a problem that has surfaced in real medical transcription records.

4: Sycophancy

AI systems are trained to be helpful, and that training makes them excessively agreeable. Sycophancy in AI means the model will validate your ideas, agree with your conclusions, and soften its criticism—not because it has evaluated your work rigorously but because agreement resembles helpfulness in its training data.

The practical consequence: never ask an AI “is this a good idea?” The model will almost certainly say yes. Instead, apply confidence scoring—prompting for criticism rather than validation. Useful prompts include: “What would a harsh critic say is missing from this?”, “What are the top three ways this could be improved?”, “What are the three strongest arguments against this approach?”, and “Cross-reference facts, dates, and definitions against trusted knowledge bases and flag potential errors or inconsistencies.”

5: Knowledge Cutoff

Every AI model is trained on data collected up to a fixed date—its knowledge cutoff. Information published after that date does not exist for the model unless it has web access or you provide the information yourself. Current models typically have cutoffs roughly a year in the past. For research in fast-moving fields, this is a material limitation that requires web-grounding or manual supplementation with recent literature.

6: Bias

Foundation models are trained on enormous datasets scraped from the internet—data that reflects the cultural, gender, political, and ideological patterns of the humans who produced it. Models absorb and can amplify those biases, potentially reinforcing harmful stereotypes or producing analysis that encodes existing societal inequities. Addressing bias in AI output requires awareness on the researcher’s part: AI output reflects patterns in its training data, not neutral observation, and should be read critically with that in mind.

7: Lack of Transparency

Most AI systems operate as black boxes: you cannot audit the reasoning behind a conclusion or trace which sources shaped a particular output. This directly contradicts scientific standards, which require reproducibility and source attribution. Anthropic (the company behind Claude) is among those working toward greater transparency, but the field as a whole remains opaque. AI conclusions should be treated as starting points requiring independent verification—not as citable findings.

8: Data Privacy and Security

When you paste research data, interview transcripts, or unpublished work into a free AI tool, you may be contributing to that model’s future training data. Free tiers of ChatGPT and Microsoft Copilot have historically retained user data; enterprise tiers provide stronger protections including data residency and AES-256 encryption.

A more acute threat for researchers using AI agents—tools that take actions on your behalf, such as searching the web, reading files, or sending messages—is prompt injection. In a prompt injection attack, a malicious actor embeds hidden instructions in content the AI agent encounters (a webpage, an email, a document) to redirect its behavior without the user’s knowledge.

Recommendations: Turn off model training in any AI tool you use for research (in Claude, disable “Help improve Claude” in Privacy settings; in Google Gemini, turn off and delete Activity; in ChatGPT, adjust Data Controls, Personalization and Memory, and GPT additional settings). Use BoodleBox or enterprise tools for anything involving research subjects, students, or sensitive institutional data.

Literature Discovery and Review

Transcription: Proceed with Caution

AI transcription tools offer genuine efficiency gains for qualitative researchers, but they carry specific risks: accuracy problems including hallucinations when audio is unclear, privacy and security exposure from cloud-based processing, legal and ethical risks (transcripts may be discoverable in legal proceedings; recording consent requirements apply), and compliance obligations in regulated fields including HIPAA, CJIS, and SEC rules.

For research involving sensitive data, we recommend local transcription tools—software that processes audio on your own machine without transmitting it to a remote server. Options include Buzz (free, open-source, cross-platform), MacWhisper (Mac), Vibe (free, cross-platform), and WizWhisp (Windows). All use the Whisper model running locally.

Literature Discovery Platforms

Six tools have emerged as the leading AI-assisted literature discovery platforms, each designed around a distinct research need:

Tool	Focus	Pricing	Key Strength	Limitation
Connected Papers	Citation mapping	Free / $6/mo	Visual graphs of papers sharing overlapping citations	One seed paper at a time
Consensus	Evidence synthesis	Free / $10 / $45/mo	Searches 200M+ peer-reviewed papers for evidence-based answers	Limited search parameter customization
Elicit	Systematic review	Free / $49 / $169/mo	Automates screening and data extraction across 138M+ papers	Occasional gaps in very recent publications
Research Rabbit	Visual mapping	Free / $10/mo	Visual network graphs across 270M+ papers	Based on older Microsoft Academic Graph metadata
Scite	Citation context	$20/mo	Smart Citations show whether citing papers support, contradict, or merely mention a finding	Focused on verification rather than discovery
Semantic Scholar	Paper discovery	Free	Citation graph and influence metrics across 200M+ papers	Researcher must perform own analysis

Grounding: Keeping AI Anchored to Evidence

Grounding refers to techniques for ensuring AI responses are based on specific, verifiable sources rather than the model’s general—and potentially outdated or inaccurate—training data. Four grounding strategies matter for research. Data grounding constrains the AI to answer only from the documents you provide. In-context grounding means you include reference examples directly in your prompt. Web grounding gives the AI access to live internet search, which helps with recency but introduces the risk of unreliable web sources. Temperature grounding uses explicit instructions to reduce creative variation, pushing the model toward precision and consistency.

NotebookLM and Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) is the technique that makes tools like NotebookLM particularly powerful for literature-heavy research. In a standard prompting workflow, you send a prompt along with whatever documents fit within the model’s context window. With large document sets, this quickly hits the limit—there is simply not enough room to include everything. RAG solves this by performing a similarity search across your entire document set, pulling only the most relevant chunks, and combining them with your prompt before sending it to the model. The result is answers grounded in your full library, not just whatever fits in a single context window.

NotebookLM (Google) currently offers the largest context window of any publicly available platform, which means it can pull more relevant chunks in that similarity search. This is its core advantage over other tools—and it is free, with a paid tier that unlocks additional document uploads.

Deep Research: Agentic Literature Discovery

Deep Research refers to a class of agentic AI tools that autonomously search the internet, evaluate sources, and compile a structured report with citations—all from a single prompt. The process is iterative: the system plans a search strategy, runs multiple queries, assembles relevant documents, and synthesizes findings into a report that includes citations you can review and accept or exclude. Deep Research is available through Google Gemini, ChatGPT, Perplexity, and as a mode within NotebookLM.

Here is a prompt template for academic Deep Research.

Research the following topic: [YOUR TOPIC HERE]

SCOPE & BOUNDARIES:

• Focus area: [SPECIFIC DISCIPLINE, e.g., "cognitive psychology," "oncology," "computational linguistics"]
• Time range: Published between [START YEAR] and [END YEAR]
• Geographic or population focus (if applicable): [e.g., "US adult populations," "global," "European Union"]

SOURCE REQUIREMENTS (CRITICAL):

• Restrict sources to peer-reviewed academic journal articles, systematic reviews, and meta-analyses.
• Prioritize sources indexed on Google Scholar, PubMed, JSTOR, IEEE Xplore, Scopus, or other recognized academic databases.
• EXCLUDE: blog posts, news articles, opinion pieces, press releases, white papers, preprints (e.g., arXiv, SSRN, medRxiv), conference abstracts without full publication, Wikipedia, and non-scholarly websites.
• Prefer articles from high-impact, well-established journals in the field.
• Every claim in the report must be traceable to a specific peer-reviewed source with author(s), journal name, publication year, and DOI or URL when available.

REPORT STRUCTURE:

1. Executive Summary — Key findings in 3–5 sentences.
2. Background & Context — What is already established in the literature.
3. Key Findings — Organized thematically, with each finding attributed to specific studies.
4. Methodological Notes — Dominant research methods used across the literature (e.g., RCTs, longitudinal cohorts, qualitative interviews).
5. Gaps & Controversies — Where the evidence is conflicting, limited, or absent.
6. Conclusion & Implications — What the body of evidence suggests and directions for future research.
7. References — Full list of all cited sources with complete bibliographic details.

QUALITY FILTERS:

• Favor studies with large sample sizes, replication, or robust methodology.
• Flag any included source that may not meet peer-review standards.
• Note the level of evidence (e.g., single study vs. systematic review) for major claims.

How to Use This Template

1. Copy the prompt above.
2. Replace every [BRACKETED PLACEHOLDER] with your specific details.
3. Open NotebookLM and select "Web" as a source.
4. Choose "Deep Research" (not Fast Research).
5. Paste your completed prompt and let it run (typically 3–5 minutes).
6. Once the report generates, review the sources it pulled in and verify journals are peer-reviewed.

An important caution from the session’s Q&A: you still have to review every source. Deep Research finds papers that keyword searches miss—including relevant work where authors did not tag their abstracts with standard terminology—but it also surfaces sources that do not meet inclusion criteria. The researcher’s judgment remains essential throughout.

Data Analysis

Data Visualization

A broad ecosystem of AI-assisted visualization tools spans free open-source options to enterprise platforms. For researchers new to data visualization, the most accessible starting points are Julius AI, which offers a freemium tier with a 50% academic discount and allows you to chat with your data files to generate charts and export results to CSV and Excel; Looker Studio (Google), which is free, web-based, and comparable in capability to Tableau, with multiple LinkedIn Learning courses available through DePaul; and Jupyter Notebooks, which is free, open-source, and runs online without local installation, supporting visualization libraries including Altair, Bokeh, Plotly, and Jupyter Widgets.

For researchers or departments building toward more robust data infrastructure, the session covered three architectural patterns. A data warehouse stores structured, tabular data—customer records, transaction logs, survey results—processed through ETL pipelines and queried using SQL. A data lake stores raw data of all types: structured tables, semi-structured files (CSV, JSON, XML), and unstructured content (email, PDFs, multimedia, sensor data), using platforms like AWS S3 or Microsoft Azure Data Lake. A data lakehouse combines the strengths of both, handling all three data types with the reliability, performance, and data quality standards of a warehouse. Microsoft Fabric, Power BI, Qlik, and Tableau all support the lakehouse model.

Writing, Editing, and Presentations

Academic Writing Tools

Two AI tools are purpose-built for academic writing. Paperpal focuses on grammar, coherence, and plagiarism checking, with a Research and Cite sidebar, a Microsoft Word add-in, and more than 1.5 million users; it is strongest for editing short-to-mid-length papers. Writefull is trained specifically on published journal articles and excels at converting informal prose to academic register through its Academizer feature; it integrates with Microsoft Word and Overleaf.

For researchers working in LaTeX: OpenAI Prism is a free, cloud-based, AI-native LaTeX workspace that integrates AI-assisted writing, real-time collaboration, and formatting tools.

AI Detection and Watermarking

AI-generated text cannot be reliably detected by traditional plagiarism tools, which work by finding prior instances of text online. AI content is generative—it produces something new each time—so there is no prior instance to match.

Current watermarking approaches attempt to solve this through statistical patterns embedded in AI output. The most technically sophisticated approach in production today is Google’s SynthID, integrated across all Gemini products and open-sourced for broader use. SynthID uses a “Tournament Sampling” algorithm to embed specific statistical patterns in generated text, audio, images, and video. A corresponding detection algorithm can verify whether content originated from a SynthID-enabled model, even after moderate editing. However, SynthID is vulnerable to paraphrasing and significant text revision. No detection tool currently provides reliable verification of AI-generated content in academic contexts.

Presentation Tools

Gamma generates presentations, websites, and social media posts from a topic description or uploaded document, and exports to PowerPoint via Share → Export. Beautiful.ai produces polished slides by refining a prompt into an outline and generating designed slides from it. Both tools are designed for rapid professional output.

Getting Started Toolkit

The tools below are grouped by research function. Free and freemium tiers are noted throughout. For any task involving research subjects, student data, or sensitive institutional information, use BoodleBox or verify that your chosen platform meets DePaul’s data security requirements before uploading.

Institutional Platform

BoodleBox — boodlebox.com

Multi-model AI platform (ChatGPT, Claude, Gemini, Perplexity) with FERPA compliance and SOC 2 certification; supports bot-building, bot-chaining, collaborative workspaces, and faculty oversight of student work. Use for any research task involving sensitive data, student work, or institutional research.

Literature Discovery

NotebookLM — notebooklm.google.com — Free

Google’s research notebook with the largest context window of any current platform; ingests uploaded documents and generates audio overviews, reports, slide decks, mind maps, flashcards, infographics, quizzes, and data tables from your literature set. Always verify source accuracy before citing outputs in academic work.

Connected Papers — connectedpapers.com — Free / $6/mo

Generates visual citation graphs showing how papers relate through shared references, useful for finding key works you may have missed. Use to map a literature landscape from a known seed paper; start with one strong anchor paper.

Consensus — consensus.app — Free / $10 / $45/mo

Searches 200M+ peer-reviewed papers and synthesizes evidence-based answers to research questions. Useful for quick evidence summaries; validate findings against primary sources before use.

Elicit — elicit.com — Free / $49 / $169/mo

Automates systematic review screening and data extraction across 138M+ papers. Verify every automated extraction against the original paper before including it in a review.

Scite — scite.ai — $20/mo

Smart Citations show whether papers support, contradict, or merely mention a finding—critical context missing from standard citation counts. Use when you need to verify the evidentiary status of a frequently cited claim.

Semantic Scholar — semanticscholar.org — Free; also available in BoodleBox

Citation graph and influence metrics across 200M+ papers. Use to identify high-impact papers and trace citation lineages; researcher must perform own analysis of results.

Deep Research

Google Gemini Deep Research — gemini.google.com

Agentic research tool that autonomously searches the web, evaluates sources, and compiles a structured report with citations from a single prompt. Use the prompt template for academic research; review and verify every returned source before use.

ChatGPT Deep Research — chatgpt.com

OpenAI’s agentic research mode; similar workflow to Gemini Deep Research. Same cautions apply: treat output as a starting point, not a finished literature review.

Perplexity — perplexity.ai — Free; also in BoodleBox

Fast, citation-backed research tool; particularly useful as a secondary fact-check (bot-stacking) after an initial AI-generated response. Use to verify specific claims and surface additional sources quickly.

Transcription (Local Processing)

MacWhisper — goodsnooze.gumroad.com/l/macwhisper — Mac; free and paid tiers

Local transcription using the Whisper model; audio stays on your machine. Required for IRB-sensitive recordings, clinical transcription, or any audio that cannot leave your device.

Buzz — github.com/chidiwilliams/buzz — Free, open-source

Cross-platform local transcription (Windows, Mac, Linux). Same use cases as MacWhisper; a strong choice for Windows and Linux users.

Vibe — thewh1teagle.github.io/vibe/ — Free, cross-platform

Local transcription with a simple interface. Good option for researchers who want a lightweight, no-cost local transcription tool.

Data Analysis and Visualization

Julius AI — julius.ai — Free tier; 50% academic discount on paid plans

Chat with your data files to generate analysis and visualizations; exports to CSV and Excel. Verify all outputs against source data before reporting; do not assume the model retained every row.

Looker Studio — cloud.google.com/looker-studio — Free ($9/mo Pro)

Google’s web-based business intelligence and data visualization tool; integrates with Google services and compares favorably with Tableau. Multiple LinkedIn Learning courses available through DePaul. A strong no-cost starting point for interactive data dashboards.

Jupyter Notebooks — jupyter.org — Free, open-source

Web-based environment for live code, equations, and visualizations in Python, R, and Julia; named for all three languages; runs online without installation using JupyterLite or JupyterHub. Use for reproducible computational research with full methodological transparency.

Microsoft Power BI — powerbi.microsoft.com — Free / $10–$24/user/mo

Business intelligence platform with extensive AI integration; 60+ LinkedIn Learning courses available through DePaul. A strong option for departments already in the Microsoft ecosystem.

Academic Writing

Paperpal — paperpal.com — Free / $11.50/mo

AI writing assistant for academic papers: grammar, coherence, and plagiarism checks with a Research and Cite sidebar; Microsoft Word add-in available; 1.5M+ users. Best for editing and refining drafts; not a substitute for original analysis or argument.

Writefull — writefull.com — Free / $150/mo

Trained on published journal articles; converts informal prose to academic register through its Academizer feature; integrates with Microsoft Word and Overleaf. Best for language editing, particularly for non-native English writers or researchers writing across disciplinary registers.

OpenAI Prism — openai.com/prism — Free

Cloud-based, AI-native LaTeX workspace for scientists and researchers; integrates AI-assisted writing, real-time collaboration, and formatting tools. For researchers already working in LaTeX; treat generated content as a draft requiring review.

Presentations

Gamma — gamma.app

Generates presentations, websites, and social media posts from a topic description or uploaded document; exports to PowerPoint via Share → Export. Use for rapid prototyping of slide structures; review and revise content before presenting.

Beautiful.ai — beautiful.ai

AI presentation tool that refines a prompt into an outline and generates professionally designed slides. Use as a starting structure; all factual content requires your review and verification.

Next Steps and Resources

The best starting point is a tool that costs nothing and requires no institutional approval. Try NotebookLM, upload a set of papers you are already working with, ask it a research question, and explore the asset generation features.

Try These First

• Google NotebookLM — Free. Upload three to five papers from your current research. Generate an Audio Overview and a structured Report, then compare the report’s claims against your own reading of the sources.
• Semantic Scholar — Free. Search your core topic and explore the citation influence graph to identify high-impact work you may have missed.
• Connected Papers — Free tier. Enter a key paper from your research and map its citation neighborhood to find related work not surfaced by keyword searches.

DePaul-Specific Resources

• LinkedIn Learning via DePaul — DePaul provides LinkedIn Learning access to all faculty and staff. Courses are available for Looker Studio (3 courses), Microsoft Power BI (60+ courses), Microsoft Fabric (2 courses), Qlik (1 course), and Tableau (20+ courses). Log in with your DePaul credentials.

A Note on Using These Tools

Every tool on this page produces output that requires your review. AI systems are nondeterministic, prone to hallucination, and often sycophantic. The researchers who get the most from these tools treat AI output as a first draft: a starting point for their own analysis, not a substitute for it. AI finds papers you might have missed. It can synthesize a literature base you provide. It can generate a structured report in minutes. What it cannot do is judge the quality of a finding, evaluate the methodology of a study, or take responsibility for the conclusions in your work. That part remains yours.