AI for Research and Analysis

Module 5 · Search, Synthesize, Verify

Content current as of April 2026. AI tools evolve rapidly — underlying principles remain applicable.

Learning Objectives

Builds on: Module 4 (BRIEFS prompting framework)

After completing this module, you will be able to:

  1. Choose the right AI research tool for different types of questions
  2. Use AI-powered search tools (Perplexity, Google AI Overviews) effectively
  3. Use Deep Research agents for multi-step investigations
  4. Turn documents into interactive knowledge bases with NotebookLM
  5. Use AI for data analysis: spreadsheets, visualization, pattern detection
  6. Apply a systematic verification workflow to AI-generated research
  7. Distinguish between AI-assisted research and AI-generated content

The Lawyers Who Cited Six Cases That Didn’t Exist

In June 2023, a New York federal judge sanctioned two attorneys and their law firm $5,000 for submitting a legal brief containing six fabricated court decisions. The attorneys — Steven Schwartz and Peter LoDuca of Levidow, Levidow & Oberman — were representing a client, Roberto Mata, in a personal-injury suit against the airline Avianca. Schwartz had used ChatGPT to find case law supporting his argument. The chatbot returned six opinions that looked real: Varghese v. China Southern Airlines, Martinez v. Delta Airlines, Shaboon v. EgyptAir, and three others — complete with plausible citations, quoted holdings, and internal references to other cases.

When opposing counsel couldn’t locate the cases and the court ordered Schwartz to produce copies, he asked ChatGPT again. The chatbot doubled down: “Varghese… is a real case” and “can be found in reputable legal databases such as LexisNexis and Westlaw” (Mata v. Avianca, 678 F. Supp. 3d 443, S.D.N.Y. 2023).

None of them existed.

The case became the most cited example of AI hallucination in professional practice — but it was not a one-off. Damien Charlotin, a French legal scholar, maintains a public database of court rulings that address AI-fabricated content in legal filings; as of April 2026, it tracks roughly 1,300 such rulings worldwide. A Stanford CodeX analysis of 114 U.S. cases through October 2025 found that 90% involved solo practitioners or small firms, and of the cases that identified the AI tool, half named some version of ChatGPT.

This story illustrates both the danger and the opportunity of AI for research. The danger: AI will confidently fabricate sources — and even fabricate the verification that those sources exist. The opportunity: when used correctly — with the right tools, the right workflow, and the right verification habits — AI can make you a dramatically more effective researcher than you’ve ever been.

This module teaches you the right way.

The New Research Landscape

Research used to mean one thing: go to a search engine, type keywords, wade through results, read articles, take notes, synthesize. It was slow, thorough, and entirely manual.

AI has fractured this into three distinct approaches, each with different strengths:

Approach How It Works Best For Watch Out For
Traditional search (Google, library databases) Returns links. You read and synthesize. Known-item search, academic databases, primary sources Slow but reliable — you verify as you go
AI-powered search (Perplexity, Google AI Overviews) Returns synthesized answers WITH citations Quick factual questions, getting oriented on a topic May cite real sources but misrepresent what they say
AI research agents (Deep Research, Gemini Deep Research) Autonomously conducts multi-step research over minutes Complex questions requiring synthesis across many sources Can go deep but may follow wrong paths. Always verify conclusions.

The key insight: these are complementary, not competitive. The best researchers in 2026 use all three — AI to accelerate discovery, traditional search to verify and go deep, and their own judgment to synthesize.

What’s the Actual Hallucination Rate?

Hard numbers from peer-reviewed research:

Tool type Hallucination rate on legal queries Source
General-purpose LLMs (e.g., GPT-4, PaLM 2) 58% - 82% Dahl et al. (2024), J. Legal Analysis
LexisNexis Lexis+ AI (RAG-grounded) ~17% Magesh et al. (2025), J. Empirical Legal Stud.
Westlaw AI-Assisted Research (RAG-grounded) ~33% Magesh et al. (2025)

Even the best RAG-grounded tools hallucinated on roughly 1 in 6 queries. Retrieval-augmented generation helps, but it does not solve hallucination. If you’re citing AI output where accuracy matters, verify every claim, every citation, every time.

Common Misconceptions About AI Research
  • “AI citations are always real” — 23 out of 47 were fake in our opening story. AI generates plausible-sounding citations as easily as real ones.
  • “If it cites a real paper, the claim is accurate” — AI can cite a real source but misrepresent what it says. The citation trap.
  • “NotebookLM replaces reading the original” — It supplements reading, not replaces it. Always check AI summaries against the source.
  • “Deep Research is always better than manual search” — Deep Research is great for breadth but can follow wrong paths. For a specific known source, manual search wins.
  • “AI research is cheating” — Using AI to FIND and ORGANIZE information is AI-assisted research. Having AI WRITE your analysis is a different thing. SMARTS keeps you on the right side.

What AI Research Looks Like in YOUR Field

  • “What’s the latest evidence on [treatment] for [condition]?” — Perplexity with Academic focus
  • Uploading 15 clinical guidelines into NotebookLM to compare recommendations across organizations
  • Deep Research on drug interaction profiles for a specific patient population
  • Critical: Medical research AI MUST be verified against peer-reviewed sources and clinical databases (PubMed, UpToDate)
  • “What are the market trends in [industry] for Q1 2026?” — Deep Research for competitive analysis
  • Uploading quarterly reports into NotebookLM to extract patterns across periods
  • Using AI to analyze customer feedback themes from 500+ survey responses
  • Critical: Financial claims from AI must be verified against SEC filings, analyst reports, or primary data
  • “What does the research say about [teaching method]?” — Perplexity with Academic focus
  • Uploading curriculum standards into NotebookLM to map learning objectives
  • Deep Research on AI literacy frameworks for a curriculum proposal
  • Critical: Educational research claims should be verified in ERIC, Google Scholar, or discipline-specific databases
  • “What are the recommended treatments for [crop disease] in [region]?” — Perplexity
  • Uploading extension service bulletins into NotebookLM for quick reference during the growing season
  • Deep Research on precision agriculture technologies for a grant proposal
  • Critical: Agricultural recommendations must be verified against local extension office guidance and soil/climate data

AI-Powered Search Tools

Perplexity

Perplexity is a search engine that answers questions conversationally and shows you exactly which sources informed each claim. Think of it as “Google that actually reads the pages for you.”

What it does well:

  • Returns concise answers with numbered source citations
  • Lets you ask follow-up questions (conversational search)
  • Shows you the actual sources so you can verify
  • Excellent for getting oriented on a new topic quickly
  • “Focus” modes let you search academic papers, YouTube, Reddit, or the full web

What to watch for:

  • It can misrepresent what a source actually says (cites a real article but summarizes it incorrectly)
  • Citation ≠ verification — just because it links a source doesn’t mean the claim matches the source
  • For academic work, always click through to the original source and confirm

Best use case: “I need to understand [topic] quickly and I want to know where the information is coming from.”

Google AI Overviews

Google now displays AI-generated summary boxes at the top of search results for many queries. These synthesize information from multiple web pages into a paragraph answer.

What it does well:

  • Fast — the answer appears before you click anything
  • Integrated with Google’s massive search index
  • Shows source links below the summary

What to watch for:

  • Early versions made embarrassing errors (suggesting people add glue to pizza to keep cheese from sliding off — sourced from a joke Reddit post)
  • Summaries can be misleading when the topic is nuanced or contested
  • You can’t ask follow-up questions — it’s a one-shot summary

Best use case: Quick factual lookup where you’ll verify the answer anyway.

The Citation Trap

AI-powered search tools cite real sources. This creates a dangerous illusion of reliability. The citation makes you trust the claim — but the AI may have misinterpreted, oversimplified, or outright misrepresented what the source actually says. Always click the source. Read the relevant section. Confirm the claim matches. Citations are starting points for verification, not proof of accuracy.

✅ Test Yourself: AI Search Tools
  1. What’s the key difference between Perplexity and a regular Google search? (Perplexity synthesizes answers with numbered citations; Google returns links you read yourself)
  2. What is the “citation trap”? (AI cites real sources but may misrepresent what they say — citation ≠ verification)
  3. When would you use traditional Google over AI search? (When you need a specific known document, a specific database, or the most authoritative primary source)

Deep Research Agents

Deep Research is a newer category of AI tool that goes beyond answering a single question. These agents autonomously conduct multi-step research — searching, reading, synthesizing, following threads, and producing comprehensive reports.

How Deep Research Works

  1. You give the agent a research question or topic
  2. The agent creates a research plan (which sub-questions to investigate)
  3. It searches the web, reads multiple sources, and extracts key findings
  4. It follows leads — if one source mentions a related study, it searches for that too
  5. It synthesizes everything into a structured report with citations
  6. The whole process takes 5-30 minutes depending on complexity

Available Deep Research Tools

Tool Provider Strengths
Deep Research OpenAI (ChatGPT Pro) Thorough, produces long-form reports with citations
Deep Research Google (Gemini Advanced) Leverages Google’s search index, good at finding recent sources
Research mode Perplexity Pro Faster than others, good at following citation chains
Claude with search Anthropic Strong at synthesis and nuanced analysis of what it finds

NotebookLM: Your Documents, Made Interactive

Google’s NotebookLM is one of the most underappreciated AI tools available. It lets you upload documents (PDFs, Google Docs, web pages, YouTube videos) and creates an interactive knowledge base that you can ask questions about.

What Makes NotebookLM Special

Unlike ChatGPT or Claude, which draw on their general training data, NotebookLM only answers from your uploaded sources. If the answer isn’t in your documents, it tells you — rather than hallucinating. This makes it dramatically more reliable for document-specific questions.

Use Cases

  • Studying: Upload lecture notes, textbook chapters, and study guides. Ask questions. Get answers grounded in YOUR course materials.
  • Meeting prep: Upload the agenda, last meeting’s minutes, and relevant reports. Ask “What were the action items from last meeting?” or “What’s the background on agenda item 3?”
  • Literature review: Upload 10-20 research papers. Ask “What do these papers agree on?” or “What methodological approaches are used across these studies?”
  • Policy analysis: Upload policy documents, regulations, and guidelines. Ask specific questions about what’s allowed, required, or prohibited.
  • Learning a new field: Upload foundational papers and textbooks. Ask the AI to explain concepts using only those sources.

Audio Overviews

NotebookLM can generate audio summaries of your documents — a podcast-style conversation between two AI voices discussing the key ideas in your uploaded material. This is genuinely useful for:

  • Reviewing material while commuting or exercising
  • Getting a high-level overview before diving into detailed reading
  • Identifying which sections of a long document are most relevant
NotebookLM vs General AI Chatbots

The key difference: grounding. ChatGPT answers from its training data (the entire internet). NotebookLM answers from YOUR documents. This means NotebookLM is less likely to hallucinate — but it can only answer questions covered by the documents you uploaded. Use ChatGPT/Claude for broad questions. Use NotebookLM for document-specific questions where accuracy matters more than breadth.

AI for Data Analysis

One of the most practical AI applications for working professionals: analyzing data without being a data scientist.

What AI Can Do With Your Data

  • Summarize a spreadsheet: “What are the key trends in this sales data?”
  • Create visualizations: “Make a bar chart comparing Q1 vs Q2 revenue by region”
  • Run calculations: “Calculate the year-over-year growth rate for each product category”
  • Find patterns: “Are there any correlations between customer age and purchase frequency?”
  • Clean messy data: “Standardize the date formats and remove duplicate rows”
  • Generate SQL queries: “Write a query to find all customers who haven’t purchased in 90 days”

Tools for AI Data Analysis

Tool How It Works Best For
ChatGPT Advanced Data Analysis Upload files, ask questions in natural language Quick analysis of CSVs and spreadsheets
Claude Upload files, strong at explaining methodology Analysis that requires careful reasoning
Google Sheets + Gemini AI built into the spreadsheet Analysis within your existing workflow
Microsoft Copilot in Excel AI built into Excel Enterprise users already in Microsoft ecosystem

The Critical Skill: Knowing When to Trust the Analysis

AI data analysis has a subtle danger: it can produce a chart that looks professional, a statistic that looks precise, and a conclusion that looks authoritative — while being completely wrong about what the data means.

Common failure modes:

  • Confusing correlation with causation — “Sales increase when temperature increases” doesn’t mean heat causes purchases (both increase in summer)
  • Ignoring missing data — analyzing only the rows that have values, producing misleading averages
  • Wrong statistical test — running a t-test when the data isn’t normally distributed
  • Cherry-picking results — presenting the one analysis that supports a conclusion while ignoring contradictory patterns

The rule: Use AI to DO the analysis (compute, visualize, summarize). Use YOUR brain to INTERPRET the analysis (what does it mean, what’s missing, what conclusions are justified).

Guided Practice 5.2

Find a publicly available dataset (e.g., from data.gov, Kaggle, or your own work). Upload it to ChatGPT or Claude and ask:

  1. “Summarize this dataset. How many rows, how many columns, what are the key variables?”
  2. “What are the 3 most interesting patterns or trends in this data?”
  3. “Create a visualization that shows the most important finding.”
  4. “What questions can this data NOT answer?”

Pay attention to question 4 — does the AI acknowledge the limitations of the data, or does it gloss over them?

What to look for Most AI tools handle questions 1-3 well — they can compute, summarize, and visualize. Question 4 is the test of critical thinking. Good models will acknowledge limitations (sample size, missing variables, time period constraints). Weaker responses will claim the data “shows” conclusions that require additional evidence. If the AI doesn’t mention limitations, you need to identify them yourself.

The Research Verification Workflow

Here’s the process that separates AI-enhanced research from AI-dependent research:

The 4-Step Verification Loop

┌──────────────────────────────────────────────┐
│  1. DISCOVER — Use AI to find information     │
│     Perplexity, Deep Research, ChatGPT/Claude │
├──────────────────────────────────────────────┤
│  2. VERIFY — Check key claims against         │
│     original sources. Click the citation.     │
│     Does it actually say what AI claimed?     │
├──────────────────────────────────────────────┤
│  3. DEEPEN — Use traditional search for       │
│     the most important findings. Find         │
│     primary sources. Read the original.       │
├──────────────────────────────────────────────┤
│  4. SYNTHESIZE — Use YOUR judgment to         │
│     combine findings. What does it mean?      │
│     What's missing? What's uncertain?         │
└──────────────────────────────────────────────┘

The Citation Verification Checklist

For every AI-generated citation you plan to use:

This takes 60-90 seconds per citation. It’s the difference between the student with 47 citations and 24 real ones versus a student with 47 citations and 47 verified ones.

The Line Between AI-Assisted and AI-Generated

This distinction matters — academically, professionally, and ethically:

Category What It Means Example
AI-assisted research You use AI to find, organize, and analyze — but the thinking, synthesis, and writing are yours Using Perplexity to find sources, then reading them yourself and writing your own analysis
AI-generated content AI produces the output — you edit, verify, or submit as-is Asking ChatGPT to “write a literature review on X” and submitting the output
AI-enhanced workflow AI handles mechanical tasks while you handle intellectual ones AI cleans your data and makes charts; you interpret the results and draw conclusions

Most academic and professional contexts accept AI-assisted research and AI-enhanced workflows. AI-generated content is more controversial and often prohibited or restricted. Know your institution’s policy (Module 10 covers this in depth).

The SMARTS framework (Module 4) naturally produces AI-assisted rather than AI-generated work — because SMARTS keeps YOU in the reasoning loop. You think, the AI supports. That’s the difference.

Module 5 Summary

Key Takeaways

  1. Three research approaches: Traditional search (reliable, slow), AI-powered search (fast, needs verification), Deep Research agents (comprehensive, needs oversight). Use all three.

  2. Perplexity gives fast, cited answers. Google AI Overviews give quick summaries. Both require you to click through and verify what the sources actually say.

  3. Deep Research agents autonomously investigate complex questions. They’re powerful but can follow wrong paths. Always verify conclusions.

  4. NotebookLM creates grounded knowledge bases from YOUR documents — more reliable than general chatbots for document-specific questions.

  5. AI data analysis is powerful for computing, visualizing, and summarizing. But INTERPRETING what the data means is still YOUR job.

  6. The 4-step verification loop: Discover (AI) → Verify (check sources) → Deepen (primary sources) → Synthesize (your judgment).

  7. Every citation must be verified. 60-90 seconds per citation. No exceptions.

  8. AI-assisted ≠ AI-generated. Use AI to find and organize. Use your brain to think and write. SMARTS keeps you on the right side of this line.

💭 Think About It

AI can read faster than you. It can search more broadly, summarize more quickly, and find connections across more sources. But it can’t think about what it reads. It can tell you what 50 papers say. It can’t tell you which ones matter, which ones are methodologically sound, and which conclusions are justified by the evidence. The research skill of the future isn’t finding information — it’s evaluating information that AI found for you. That requires domain knowledge, critical thinking, and judgment. Those are human skills. They just became the most valuable skills in the room.

🔬 Lab 5: The NotebookLM Deep Dive

Lab 5: The NotebookLM Deep Dive

Time: 45 minutes | Tools needed: Google NotebookLM (free at notebooklm.google.com) + 3 documents from your field

Part A — Build Your Knowledge Base (10 min) 1. Choose 3 documents relevant to your work or studies (PDFs, Google Docs, or web links). These should be documents you’re already somewhat familiar with. 2. Upload them to a new NotebookLM notebook. 3. Generate the Audio Overview and listen to at least 3 minutes of it.

Part B — Test the Boundaries (15 min) 4. Ask NotebookLM 5 questions about your documents: - 2 questions you already know the answer to (to test accuracy) - 2 questions you’re genuinely curious about (to test usefulness) - 1 question that CANNOT be answered from these documents alone (to test whether it admits limitations or halluccinates)

Record each question, the AI’s response, and your assessment.

Part C — Compare With a General Chatbot (10 min) 5. Ask the same 5 questions to ChatGPT or Claude (without uploading the documents). Compare the responses. Which is more accurate for document-specific questions? Which is more useful for broader questions?

Part D — Reflection (10 min) 6. Write a short paragraph: When would NotebookLM be your first choice vs a general chatbot? What surprised you? How might you use this tool in your actual work?

Submit: Your questions, NotebookLM responses, comparison notes, and reflection.

🎮 Module 5 Quiz

Previous: Module 04 — Prompting | Next: Module 06 — AI for Productivity