AI has changed how people start researching. Instead of opening twenty tabs, you can ask an assistant, or a deep research feature, to survey a topic and write a summary with citations. That is a real time saver. It is also a trap if you treat the summary as the finished product.
Updated September 2026: we added sources on deep research tools, fabricated citations, AI sycophancy and Google’s research notebook.
Here is a workflow that uses AI for speed and people for judgment.
Step 1: Scope the question
Vague questions produce vague research. Before searching, ask the assistant to help you sharpen your question:
Scoping prompt I want to research [topic] because [purpose]. Help me turn this into three to five specific research questions, and suggest what kinds of sources would be most authoritative for each.
Step 2: Get the lay of the land
Use an AI search tool or a deep research feature to produce an overview with citations. Ask it to identify:
- The main viewpoints or schools of thought
- Key terms and concepts
- Where evidence is strong and where it is contested
- The most important primary sources
Step 3: Go to the sources
This is the step people skip, and it matters most. A 2023 study in Scientific Reports found that 55% of citations generated by GPT-3.5, and 18% of those from GPT-4, were fabricated. Open the cited sources and check:
- Does the source exist and say what the summary claims?
- Who published it, and do they have expertise and accountability?
- Is it primary evidence, such as research, official data or original documents, or commentary on it?
- How recent is it?
Our hallucinations guide explains why this step is essential.
Step 4: Search beyond the AI
AI tools draw on what their search step retrieves, which may miss important material. Supplement with:
- Traditional search and specialist databases, as compared in our AI vs traditional search piece
- Academic sources such as Google Scholar or your library’s databases
- Official statistics and government publications
- Expert organizations in the field
Step 5: Stress-test your findings
Ask the assistant to argue against your emerging conclusion, identify what evidence would change it and point out gaps. This counters confirmation bias, including the tendency of assistants to agree with how you frame the question. Researchers call this sycophancy; a 2023 study found it in five leading AI assistants and traced it partly to training on human preferences.
Step 6: Organize your notes
Upload your verified sources to a notebook or project tool, such as Google’s Gemini Notebook, and ask questions across them, a practical use of retrieval. Keep a record of which source supports each claim so your final work is traceable.
Standard to hold yourself to Every important claim in your final work should trace back to a source you have personally read, not to an AI summary of it.
Write it yourself
Use AI to help outline and edit, but write the conclusions in your own words and reasoning. You understand the material better when you have to explain it, and your readers can tell.
Where AI fits in good research
AI is a superb research starting point: fast scoping, broad surveys and organized notes. Reliable research still depends on reading primary sources, looking beyond one tool and challenging your own conclusions.
Sources
- Introducing deep research, OpenAI, February 2025
- Fabrication and errors in the bibliographic citations generated by ChatGPT, Scientific Reports, 2023
- Towards Understanding Sycophancy in Language Models, arXiv, 2023
- NotebookLM is now Gemini Notebook, Google, July 2026
