A clear structure for how AI-assisted search differs from traditional search.

Traditional keyword search workflow:

1. Think of keywords -> Search database -> Screen 500+ results -> Read abstracts -> Select papers -> Read full texts

AI-assisted search workflow:

1. Form research question -> Initial exploration (AI-powered) -> Understand landscape -> Targeted discovery (AI-powered) -> Paper screening (AI-assisted) -> Organization -> Full-text review

Key differences:

• AI-assisted workflow emphasizes understanding the landscape before deep diving
• Uses semantic search to find papers you wouldn't search for with keywords
• Uses AI to help screening (though human still judges relevance)
• Results in fewer, higher-quality papers for full-text review

Different stages of your literature search have different goals and use different tools.

Exploratory search:

• Goal: Understand what exists in this research area
• Question: What papers exist on this topic? What are the main themes?
• Tools: Semantic Scholar, Elicit, general-purpose LLM for brainstorming
• Volume: Many papers to get breadth of field
• Example: "Papers on neurofeedback for ADHD in children"

Foundational search:

• Goal: Identify the key papers that define your field
• Question: Which papers are most cited and influential?
• Tools: Connected Papers, citation mapping, research databases
• Volume: Moderate number of high-impact papers
• Example: "Core papers on neurofeedback mechanism of action for ADHD"

Targeted search:

• Goal: Find papers on specific aspects of your question
• Question: What about [specific methodology/population/outcome]?
• Tools: Semantic Scholar with specific query, Elicit for methodological search
• Volume: Focused set of relevant papers
• Example: "Neurofeedback studies using real-time fMRI with sample sizes over 20"

Contradiction search:

• Goal: Understand disagreement in the literature
• Question: Why do some papers find X while others don't?
• Tools: Scite for citation context, Consensus for contradiction mapping
• Volume: Papers with opposing findings
• Example: "Studies with contradictory findings about neurofeedback efficacy"

How to phrase questions that AI tools understand well.

Key points:

• Concept-based queries work better than keyword lists:
• Poor: "ADHD neurofeedback children efficacy"
• Better: "How effective is neurofeedback for reducing ADHD symptoms in children?"

• Include population, intervention, and outcome:
• Poor: "neurofeedback research"
• Better: "Neurofeedback training for attention deficits in adolescents with ADHD"

• Questions work better than keywords:
• Poor: "anxiety cognitive behavioral therapy adolescents"
• Better: "How do cognitive behavioral therapy effects on anxiety compare between adolescents and adults?"

• Specific comparisons work well:
• Poor: "depression treatment"
• Better: "How does medication compare to cognitive behavioral therapy for depression in adolescents?"

• Iterative refinement: Start broad, narrow based on what you find

How to use AI to help you screen large numbers of papers without losing your critical thinking.

Key points:

• AI can read abstracts and help categorize relevance
• You make the final judgment about inclusion
• This reduces screening burden without eliminating human judgment
• Process:

1. Paste title + abstract into AI
2. Ask: "Is this paper relevant to [research question]?" or "What is the main methodology of this study?"
3. Use AI response to inform your judgment (but verify)
4. Decide include/exclude

• This is much faster than reading every abstract yourself but maintains your control

How to organize papers during search prevents chaos later.

Key points:

• Create a tracking system (spreadsheet, reference manager) from the start
• Track: author, year, title, why you included it, key findings, methodology
• Tags help organize: "methodology-RCT", "outcome-anxiety", "population-adolescent"
• Note contradictions as you go
• Regular review: every 20 papers, assess your categories
• Good organization now makes synthesis later much faster

—

Practical Research Use Cases

Research question: How is artificial intelligence being used in mental health interventions for adolescents? What's the evidence for effectiveness?

Stage 1: Exploratory Search (Goal: Understand landscape)

Steps:

1. Open Semantic Scholar (scholar.google.com/scholar)
2. Search: "artificial intelligence mental health interventions adolescents"
3. Review top 20 results
4. Note: Papers cover multiple angles (diagnosis, prediction, therapy, apps, chatbots)
5. Observe: Field is emerging but active (recent papers)
6. Take-away: Field exists, has multiple approaches

Stage 2: Foundational Papers (Goal: Understand what defines the field)

Steps:

1. Go to Connected Papers (connectedpapers.com)
2. Search for one key paper you found in Stage 1
3. Visualize: See which papers are cited together, which are foundational
4. Review papers identified as "core" or highly cited
5. Record: 5-10 key papers that define how this research is done
6. Note: Which methodologies are standard? Which populations are studied most?

Stage 3: Targeted Searches (Goal: Find papers on specific aspects)

Search 1: Methodology - "AI chatbots for depression treatment in adolescents"

• Use Elicit or Semantic Scholar with this specific query
• Screen 30 results, keep 8-10 most relevant
• Note: What methodologies do these studies use? RCTs? Observational?

Search 2: Population - "AI interventions for anxiety specifically in teenagers"

• Use semantic query: "How do anxiety and artificial intelligence intersect in treatment studies of young people?"
• Screen results, keep 5-7 most relevant
• Note: Are adolescents studied separately from adults?

Search 3: Mechanism - "How does AI personalization work in mental health"

• Search for mechanism-focused papers
• Keep 3-5 papers
• Note: Do papers explain WHY AI helps or just that it does?

Stage 4: Organize and Note Gaps (Goal: Prepare to write)

1. Create spreadsheet with columns:

• Author | Year | Title | Type (RCT/Observational/Review) | AI Type (Chatbot/Recommendation/Diagnosis/Other) | Outcome (Depression/Anxiety/General mental health) | Sample Size | Key Finding | Methodological Strength | Status (Full-text reviewed/Abstract only)

1. Enter all papers you've selected (probably 30-50 at this point)

1. Mark papers you must read in full for your review

1. Note gaps in literature:

• Are certain populations understudied?
• Certain AI approaches?
• Certain outcomes?

1. Do final targeted searches for gaps:

• "AI mental health outcomes adolescents autism spectrum"
• (if you notice ASA population underrepresented)

Stage 5: Verify Literature with Manual Check

Steps:

1. Look at your organized papers
2. Are there important names/groups in this field you haven't found?
3. Do a forward-citation search on key papers: "Who has cited this important paper?"
4. Check: Did AI tools miss important papers? Usually not, but verify
5. For systematic reviews: Conduct final keyword search in PubMed/Web of Science to ensure comprehensive coverage

—

Hands-On Exercise

Objective: Actually do a literature search using AI tools, experiencing the workflow firsthand.

Steps (Expected time: 4-6 hours over multiple sessions):

Session 1: Exploratory Search (60 minutes)

1. Choose your research question (can be for current project or hypothetical)
2. Open Semantic Scholar
3. Search using a concept-based query (not just keywords)
4. Browse 20 top results
5. Document: What themes emerge? What journals publish this research? What's the timeline?
6. Note: 3-5 papers that seem foundational

Session 2: Deepen Exploration (60 minutes)

1. Open Connected Papers
2. Enter one foundational paper you identified
3. Explore the visualization
4. Identify papers that appear central (many connections)
5. Note: Which papers should you definitely read?
6. Document: Updated understanding of field

Session 3: Targeted Searches (90 minutes)

1. Identify 2-3 specific aspects you want to search for (methodology, population, outcome, mechanism)
2. For each aspect, conduct targeted search on Semantic Scholar
3. Screen titles and abstracts from each search
4. Keep 5-15 papers per search you think are relevant
5. Create spreadsheet starting to track papers (author, year, title, why included)

Session 4: Organize and Synthesize (90 minutes)

1. Go through all papers you've collected (probably 40-60 at this point)
2. Complete spreadsheet for each paper:

• Author | Year | Title | Relevance (how directly related to your question?) | Key finding | Methodology | Notes

1. Create tags: At least organize by theme or category
2. Review completed spreadsheet
3. Identify: What's the landscape? What patterns do you see? What gaps exist?
4. Document your assessment: What are the main findings across papers? Where do they disagree?

Session 5: Verification (60 minutes)

1. List papers you plan to read in full
2. Do one manual search in PubMed or your field's database to verify you haven't missed obvious papers
3. Note: Did AI tools find papers the manual search didn't? Did manual search find papers AI missed?
4. Reflect: How did AI-assisted searching compare to how you would have done it traditionally?

Time required: 4-6 hours across multiple days

—

Common Mistakes and Misconceptions

AI-powered search engines have large databases but not exhaustive. Always do one verification search in your field's main database (PubMed, Web of Science, PsycINFO, etc.) to ensure you haven't missed papers.

No. Screen papers first at abstract level. Many papers will be peripheral to your exact question. Read full texts only for papers most directly relevant. AI helps you prioritize.

Literature discovery is iterative. Search once, understand landscape, search again with refined understanding, repeat. Each iteration finds papers previous searches missed.

Organization during search saves enormous time during synthesis. Creating a spreadsheet of papers as you find them (10 min per paper) prevents spending weeks later trying to remember which paper said what.

It's done when you've:

• Understood the landscape (exploratory search)
• Found foundational papers (citation mapping)
• Searched multiple specific angles (targeted search)
• Verified you haven't missed anything (manual verification)
• Organized papers for synthesis (spreadsheet complete)

—

Key Takeaways

• AI-assisted search workflow emphasizes landscape understanding first, deep diving second, using exploratory and foundational searches to understand the field before targeted searches
• Different search stages use different tools: exploratory (Semantic Scholar, general LLM), foundational (Connected Papers), targeted (Semantic Scholar with specific queries), verification (manual database search)
• Concept-based queries work better than keywords, and questions work better than keyword lists; iterative query refinement based on landscape understanding improves results
• Screening papers with AI assistance means using AI to summarize or categorize abstracts while you maintain judgment about relevance
• Organization during search (creating spreadsheet with metadata as you find papers) prevents chaos and enables rapid synthesis later
• Verification with traditional databases ensures AI-assisted search hasn't missed important papers, completing comprehensive literature discovery

—

Reflection Questions

1. Your research question: For your current or next research project, formulate a research question. Write out how you'd conduct an exploratory search, a foundational search, and two targeted searches for specific aspects.

1. Tool selection: For your literature search, which tools (Semantic Scholar, Connected Papers, Elicit, Consensus) would be most helpful? Why?

1. Organization system: How would you organize papers as you search? What metadata matters for your field? What tags/categories would help you later?

1. Verification approach: How would you conduct a final verification search to ensure you found the major papers in your field?