For Researchers and Academics
Cognitive Sovereignty Checklist for Researchers
When you use Elicit or Claude to summarise literature, you risk missing the tensions and contradictions that point to genuine gaps in knowledge. AI tools smooth over disagreement between sources, making messy reality look cleaner than it is. Your job is to stay alert to what the AI is discarding when it distils fifty papers into five sentences.
These are suggestions. Take what fits, leave the rest.
0 / 19 applicable
Tap once to check, again to mark N/A, again to reset.
Before you search: define your question without AI
Write your research question on paper before opening any AI toolbeginner
Researchers often unconsciously shape their questions around what AI can easily analyse. Write what you actually want to know first, then decide which tools fit the question. Not the other way around.
List the contradictions in your field that sparked this researchbeginner
Note the specific disagreements between researchers that made you want to investigate. Keep this list visible when reviewing AI summaries. You will notice when the AI has flattened a disagreement into a false consensus.
Identify which findings would surprise you mostintermediate
Write down the three outcomes you would find most unexpected. This guards against confirmation bias when you review AI-selected literature. You will spot when Elicit or Semantic Scholar is returning only papers that match your prior assumptions.
Set explicit criteria for what counts as a relevant source before searchingbeginner
Decide in advance whether you need empirical data, theoretical work, critical responses, or historical context. Without this filter, you will accept whatever the AI ranks as relevant. Relevance is a choice you must make, not a fact the algorithm discovers.
Sketch your expected methodology before asking AI to suggest analytical approachesintermediate
Describe how you plan to analyse your data or synthesise literature based on your question alone. Only then compare it to what ChatGPT or Claude recommends. This prevents methodological drift where you adopt the approach the AI finds easiest to perform.
Note which recent papers you already know are important to your questionbeginner
Before using Semantic Scholar, list the three to five papers you are certain belong in your review. After the AI search, check whether these appear in the results. If they do not, the algorithm may be filtering by citation count rather than relevance to your specific question.
During synthesis: interrogate what AI is removing
Read the original abstract whenever AI provides a summary longer than two sentencesbeginner
AI summaries of papers often omit caveats, limitations, and alternative interpretations that the authors stressed. These details matter. They change what the paper actually claims.
Track which papers contradict each other and note where the AI hides these conflictsintermediate
When you input ten papers into Claude or Perplexity, mark any findings that disagree across sources. Then read the AI summary. If it presents a unified picture, you have caught the AI erasing disagreement. Go back to the originals and ask why the sources conflict.
Examine the methodologies of papers the AI groups together as supporting the same pointintermediate
AI often clusters papers by conclusion rather than method. A qualitative interview study and a large survey may reach similar conclusions but with very different evidential weight. Check whether the studies the AI treats as equivalent actually are.
Ask the AI to list papers that contradict its own summaryintermediate
After receiving a literature summary from ChatGPT or Claude, ask it to identify sources in your collection that disagree with the summary. This forces the AI to retrieve and surface conflict rather than suppress it.
Separately read papers the AI ranked as less relevantadvanced
Semantic Scholar and Elicit rank papers by algorithmic relevance. Skim at least one paper from the bottom of the results. Often these contain important critiques or methodological alternatives the algorithm downranked.
Create a contradiction matrix before writing your synthesisintermediate
Make a table with papers as rows and key claims as columns. Mark where findings agree, disagree, or rest on different assumptions. This visual record prevents you from absorbing AI summaries that hide these patterns.
Verify each causal claim the AI makes by checking the original study designadvanced
AI summaries frequently convert correlational findings into causal statements. Look at the actual methods section. If the study was correlational, the AI should not present it as showing causation.
During writing: maintain your voice and accountability
Write your first draft by hand or in a separate document without AI assistancebeginner
Composing without Claude or ChatGPT forces you to articulate your own thinking. You then use AI only to refine clarity, not to generate ideas. This preserves the distinction between your argument and AI-assisted text.
Keep a separate document listing which claims come from which sourcesbeginner
When you cite a paper, record the page number and exact finding next to your draft sentence. You will catch yourself either before or after you paraphrase, rather than submitting text you cannot trace to origins.
Use AI only for editing sentences you have already written, never for generating new paragraphsbeginner
Ask Claude to improve clarity or grammar in work you authored. Do not ask it to write a paragraph on a topic and then lightly edit the result. The first approach preserves your authorship. The second does not.
Mark every sentence ChatGPT or Claude touched before submitting your workintermediate
Highlight, underline, or use comments to flag any text the AI assisted with. Read these sentences aloud before submission to check they match your actual position and that you can defend them in detail.
Write a one paragraph explanation of why your methodology differs from what AI suggestedadvanced
If you chose a different analytical approach than the AI recommended, document your reasoning. This clarifies your independence and helps future readers understand your choices.
Conduct your data analysis before showing preliminary results to any AI toolintermediate
Complete your statistical tests, coding, or synthesis independently first. Only then ask AI to help with interpretation or visualisation. This ensures your analytical decisions are yours.
Five things worth remembering
- When Elicit or Semantic Scholar returns 200 papers, manually review the first ten and the last ten. The difference reveals what the algorithm values versus what might be genuinely novel.
- After ChatGPT summarises a paper, ask it the question: 'What would the authors say this study does not address?' This forces retrieval of limitations the AI initially smoothed over.
- Keep a running list of findings that surprised you during the literature review. If this list stays empty after reading fifty papers, ask whether the AI is returning only confirming evidence.
- Before submitting a draft, read aloud every sentence you generated versus every sentence the AI assisted with. Your ear will catch where your voice stops and the AI voice begins.
- Schedule a meeting with a colleague who has not seen your work. Ask them to identify which claims they find surprising or doubtful. If they are never surprised, your literature synthesis may be too tidy.
Related reads
Common questions
Should researchers write your research question on paper before opening any ai tool?
Researchers often unconsciously shape their questions around what AI can easily analyse. Write what you actually want to know first, then decide which tools fit the question. Not the other way around.
Should researchers list the contradictions in your field that sparked this research?
Note the specific disagreements between researchers that made you want to investigate. Keep this list visible when reviewing AI summaries. You will notice when the AI has flattened a disagreement into a false consensus.
Should researchers identify which findings would surprise you most?
Write down the three outcomes you would find most unexpected. This guards against confirmation bias when you review AI-selected literature. You will spot when Elicit or Semantic Scholar is returning only papers that match your prior assumptions.