Deep Research, SDG Mapping, and Bias Detection. Learn to run professional-grade AI research across multiple tools, map real organizations to the UN Sustainable Development Goals, and detect bias in AI-generated findings.
"Every big change starts as one small, deliberate action."
First Ripple
Student Access
Token provided in class or on Canvas.
Professor Meirelles van Vliet
Fall 2026 · T/R 3:30-4:45 PM · KWR 310
SE 395-2: AI Tools for Social Impact
02
Week 4 · Thursday · AI Lab
Deep Research, SDG Mapping, and Bias Detection
Learn to run professional-grade research across multiple AI tools. Map your community partner to the UN Sustainable Development Goals. Detect bias in what AI tells you and what it leaves out.
Time
75 minutes in class + homework
Tools
Gemini, Perplexity, Claude or ChatGPT
Due
Tuesday, Sep 22, 11:59 PM
Overview
What This Lab Is For
In Lab 1 you learned to have a conversation with AI: prompt, follow up, verify, challenge. That is the foundation. This lab builds on it by teaching you how professionals actually use AI for research: decomposing a question into parallel streams, running each stream through a different tool, evaluating sources by credibility, detecting bias in the results, cross-referencing across streams, and synthesizing everything with proper citations.
You will apply this research process to a real task: mapping your community partner's work to the UN Sustainable Development Goals and identifying where AI tools could genuinely support their mission. Along the way, you will practice detecting the kinds of bias that AI research tools introduce, from whose data they trained on to whose perspectives they leave out.
This is a longer lab than Lab 1. You will start in class and finish as homework.
Why three tools? Each AI research tool has different strengths and blind spots. Using only one is like reading only one newspaper. Running the same question through multiple tools and comparing results is how you build reliable knowledge.
Your Tools
Three Tools, Three Roles
You will use all three tools in this lab. Each one plays a specific role.
Tool
Role in This Lab
Access
Gemini
Deep research. Gemini Deep Research autonomously browses the web, reads dozens of sources, and produces a structured report with citations. This is your primary research engine.
Free with any Google account
Perplexity
Source verification. Perplexity returns inline citations with every answer. Use it to fact-check claims from your deep research, verify organizations exist, and confirm statistics are real.
Free tier (perplexity.ai)
Claude or ChatGPT
Bias detection and synthesis. Feed your research findings to Claude or ChatGPT and have it critique them: what perspectives are missing, whose data trained this, where is the corporate framing, what contradicts what.
Whichever you subscribe to
You must use all three tools. Submitting research from only one or two tools will result in a significant point deduction. The whole point of this lab is learning when to use which tool for what purpose.
The Exercise
Six Steps
Work through these in order. Steps 1-3 should happen in class. Steps 4-6 are homework.
1 Choose your partner and decompose
Pick one of the three community partners you have met so far: Ideas for Us, Mills Gallery, or ecoPreserve. You will research this organization for the entire lab.
Break your research into three independent sub-queries. Each one should be answerable without needing the results of the others. This is how professional research pipelines work: you split the question into pieces that can run in parallel, then merge the results.
Decomposition Template
Your three sub-queries should cover:
The organization's domain: What does the research say about AI applications in [arts/environment/sustainability consulting]? What is working, what is experimental, what is hype?
SDG alignment: Which specific UN SDG targets (not just goals) connect to [partner name]'s actual operations? What evidence links their type of work to measurable SDG progress?
Tool landscape: What AI tools are actually available and affordable for organizations of this size (2-5 people, limited budget)? What do they cost, what data do they collect, and who built them?
Write your three sub-queries down before you start researching. Having clear questions before you open any tool prevents you from going down rabbit holes. You can always add follow-up queries later.
2 Run deep research with Gemini
Open gemini.google.com and use Gemini Deep Research to investigate your first sub-query. Deep Research will browse multiple websites, read full articles, and compile a structured report.
Gemini Deep Research Prompt (Sub-Query 1)
I am a university student researching how AI is being used in [partner's domain, e.g., "small contemporary art galleries" or "sustainability consulting firms" or "environmental nonprofits"]. I need a comprehensive research report covering:
Real examples of organizations in this space using AI today (with names, not hypotheticals)
What specific AI applications have evidence of working vs. what is still experimental
Peer-reviewed studies, reports, or case studies documenting outcomes
What the main risks and limitations are for AI in this domain
Which applications are accessible to organizations with fewer than 5 staff members
For every claim, include the source. Distinguish between peer-reviewed research, organizational reports, news coverage, and marketing material. If something is unverified, say so.
While Gemini processes your first query, run your second sub-query (SDG alignment) and your third (tool landscape) as separate Deep Research sessions. You can have multiple research sessions running.
Gemini Deep Research Prompt (Sub-Query 2: SDG Mapping)
I need to map the work of [partner name and brief description] to the UN Sustainable Development Goals. Do not just list which SDGs are "relevant." I need:
Specific SDG targets (e.g., Target 11.4, not just "SDG 11") that connect to their actual operations
Evidence that this type of work contributes to measurable progress on those targets
Examples of similar organizations that have documented SDG-aligned outcomes
Which SDG connections are strong (backed by evidence) and which are a stretch
Cite your sources. If a connection is based on a published framework or report, name it.
Gemini Deep Research Prompt (Sub-Query 3: Tool Landscape)
I am looking for AI tools that a [2-3 person art gallery / sole sustainability consultant / small environmental nonprofit] could realistically adopt. I need:
Specific tool names with pricing (free tier, monthly cost, enterprise pricing)
What data each tool collects and where it is stored
Who built the tool (startup, Big Tech, open source community)
Whether an open-source alternative exists for each paid tool
Any documented case of a small organization successfully adopting each tool
Focus on tools that exist today and have real users, not products that are "coming soon" or in beta with no public access.
Save all three Gemini reports. You will need them for Steps 3, 4, and 5. Export or copy each report as soon as it finishes.
3 Verify with Perplexity
Open perplexity.ai. Your job now is to fact-check the most important claims from your Gemini research. Perplexity shows you exactly where each piece of information comes from.
Pick at least 5 specific claims from across your three Gemini reports. For each one, run a targeted verification query.
Verification Prompt Template
I found the following claim in my research: "[paste the specific claim from Gemini]"
Help me verify this:
Is this accurate? What is the original source?
When was this information published or last updated?
Is the source a peer-reviewed study, an organizational report, a news article, or marketing material?
Does any contradicting information exist?
Track your verification results. For each claim, note: what Gemini said, what Perplexity found, whether the claim held up, and the credibility of the original source. You will need this for your source evaluation table.
If Perplexity contradicts Gemini, that is valuable data. Do not ignore contradictions. Document them. Disagreements between tools are where the interesting findings live.
4 Detect bias with Claude or ChatGPT
This is where you put on your critical thinking hat. Open Claude or ChatGPT (whichever you subscribe to) and feed it your research findings. Your goal is not more research. It is interrogation.
Bias Detection Prompt
I just completed a deep research exercise on [partner name] and their connection to the UN SDGs. Here are my key findings:
[Paste a summary of your main findings from Steps 2-3, including any contradictions you found]
Now I need you to help me detect bias in these findings. Analyze them through these lenses:
Geographic bias: Are these findings mostly from the US/Europe? What perspectives from the Global South are missing? Would someone in Lagos or Jakarta see this issue differently?
Data representation: Whose data trained the models that generated this research? Who is overrepresented and who is invisible in the training data for this domain?
Corporate framing: Which of these findings come from companies selling AI tools? How does that framing differ from what independent researchers or the organizations themselves would say? (Think about what Iazzolino and Stremlau warned about in the PMC reading.)
SDG framing bias: Does mapping everything to the SDGs privilege certain kinds of work over others? What does the SDG framework make visible and what does it obscure?
Missing voices: If I were to interview the actual staff at [partner name], what would they likely say that none of these AI-generated sources capture?
Be direct. I need honest critique, not encouragement.
Corporate Capture Filter
Look at the AI tools I identified in my research:
[List the tools from your Sub-Query 3 findings]
For each tool, answer:
Who built it and what is their business model?
What data does the tool collect, and does the company use that data for anything beyond the stated purpose?
What happens to the organization's data and workflows if this vendor raises prices or shuts down?
Is this tool making the organization dependent on a platform they cannot leave?
Does an open-source or self-hosted alternative exist?
This is not anti-technology. This is Iazzolino and Stremlau's "corporate capture" test applied to real tools that a real organization might adopt.
"The question is not whether AI is biased. It is always biased. The question is: in what direction, to whose benefit, and at whose expense?"
Keep this frame for the rest of the course.
5 Cross-reference and map
Now bring everything together. You have three Gemini research reports, Perplexity verification results, and a bias analysis from Claude or ChatGPT. Lay them side by side.
Build two things:
A. Source Evaluation Table
Create a table with every major source your research cited. For each source, rate it:
Source
Type
Credibility
Verified?
Bias Notes
Example: Nature Communications (2020)
Peer-reviewed paper
High
Yes, via Perplexity
Western-centric dataset
Example: [AI Vendor] blog post
Marketing material
Low
Claims unverified
Selling their own product
Source types: Peer-reviewed paper, Government/UN report, Organizational report, News article, Blog post, Marketing material, Dataset documentation.
Credibility ratings: High, Medium, Low. If you could not verify a source, say so.
B. SDG Map
Map your community partner to specific SDG targets. For each connection:
Name the specific target (e.g., Target 12.6, not just "SDG 12")
Describe how the partner's actual work connects to it
Rate the strength of the connection: Strong (evidence-backed), Moderate (logical but limited evidence), or Weak (plausible stretch)
Cite the source that supports this connection
Fewer strong connections are better than many weak ones. A map with 3 well-evidenced SDG connections is stronger than one that claims alignment with 10 goals based on vague logic.
6 Synthesize and reflect
Write your final synthesis. This is not a summary of what the AI told you. It is your analysis of what you found, how reliable it is, and what it means for your community partner.
Your synthesis should answer:
What did you learn about your partner's domain that surprised you?
Where did the three tools agree and where did they disagree?
What biases did you detect in the AI-generated research? How would those biases affect recommendations made to your partner?
If you were advising your partner today, which 1-2 AI tools (if any) would you recommend and why? What would you warn them about?
What questions do you still have that no AI tool could answer? What would you need to learn directly from the partner?
Every factual claim in your synthesis must have an inline APA citation. This is not optional. If you cannot cite a source for a claim, it does not belong in the report.
APA inline citation format: (Author, Year) for paraphrased ideas. (Author, Year, p. X) for direct quotes. For websites without a named author, use the organization name: (Stanford HAI, 2026). Your reference list at the end should include the full citation for every source.
Rubric
What Strong Work Looks Like
Strong (full credit)
Weak (minimal credit)
You used all three tools for their designated roles and can explain what each contributed
You used one tool for everything or used tools interchangeably without purpose
Your source evaluation table rates credibility honestly and catches marketing disguised as evidence
You list sources without evaluating them or rate everything as "High" credibility
Your SDG mapping connects to specific targets with evidence and rates connection strength
You list SDG goals without targets, evidence, or honest strength ratings
Your bias analysis identifies specific biases (geographic, corporate, data) with examples
You write "AI can be biased" without identifying specific biases in your own findings
Your synthesis includes inline APA citations for every factual claim
You make claims without citations or dump a reference list with no inline links
You document contradictions between tools and explain what they mean
You ignore contradictions or pretend all three tools said the same thing
Your reflection connects to the readings (Dhaliwal/Hou, Iazzolino/Stremlau)
Your reflection is generic and does not reference course material
Deliverables
What to Submit
Upload a professionally formatted document to Canvas as a PDF or Word file (.docx). This report will be longer than Lab 1: aim for 5-7 pages. Include all five sections below.
Section 1: Research Design (from Step 1)
Which community partner you chose and why
Your three sub-queries, written out as clear research questions
A brief explanation of why these three queries are independent of each other
A table with every major source cited in your research (minimum 10 sources)
Each source rated by type, credibility, verification status, and bias notes
At least 5 claims explicitly verified through Perplexity with results documented
Section 3: SDG Map (from Step 5)
Your partner mapped to specific SDG targets (not just goals)
Evidence for each connection with a cited source
Honest strength ratings: Strong, Moderate, or Weak
Section 4: Bias Audit (from Step 4)
At least three specific biases you detected in your AI-generated research
For each bias: what it is, how you detected it, and how it would affect recommendations
Your corporate capture analysis of at least 2 AI tools you researched
Section 5: Synthesis and Reflection (from Step 6)
Your overall findings with inline APA citations throughout
Tool comparison: what each tool got right, wrong, or missed
1-2 tool recommendations for your partner (or an honest argument for none)
What you still need to learn from the partner directly
APA reference list at the end
Research Logs (All Three Tools)
You must also submit research logs from all three tools. This is not optional.
Gemini Deep Research
Export or screenshot your three Gemini Deep Research reports. Include the sources Gemini cited.
Perplexity
Export or screenshot your verification queries and results, showing the inline citations Perplexity provided.
Claude or ChatGPT
Submit your bias detection conversation using one of the methods from Lab 1: ask the AI to generate a conversation document, share a link, or copy-paste into a document.
I need to see how you used each tool. A report with no research logs cannot receive full credit.
Grading
How This Gets Graded
Source evaluation is thorough, honest, and catches unreliable sources
20%
SDG mapping uses specific targets with evidence and honest strength ratings
20%
Bias audit identifies specific, concrete biases with examples from your research
20%
Synthesis uses APA citations, connects to readings, and shows original thinking
20%
All three tools used purposefully with research logs showing the process
15%
Research design is clear with well-decomposed sub-queries
5%
Due: Tuesday, September 22 by 11:59 PM. Upload your report and all three research logs to Canvas.
Length: The written report should be 5-7 pages, professionally formatted (PDF or .docx), with APA inline citations and a reference list. Research logs can be any length.
On honesty: Do not fabricate sources, do not invent APA citations for articles that do not exist, and do not claim you verified something you did not. If Gemini cited a source you cannot find, say so. If a claim could not be verified, label it as unverified. Honesty about what you do not know is worth more than a polished fiction.
★
Exemplary Submission
This report uses Second Harvest Food Bank (not one of our actual partners) to demonstrate what an excellent Lab 2 submission looks like. Use it as a model for structure, depth, and critical thinking.
★
Exemplary Conversation Logs
These logs show how Gemini, Perplexity, and Claude were used for their designated roles: primary research, source verification, and bias detection. Notice the annotations between exchanges where the student flags issues for follow-up.
