One of the most frequent and most poorly executed uses of AI is research. The typical flow of the inexperienced user: ask the model a question, receive a detailed and confident response, use it directly without verifying. The problem is not that AI always produces incorrect responses — many are correct — but that it produces incorrect responses with the same tone as the correct ones. And that means the user does not know when to trust.
Using AI for research effectively requires a different mental model: AI is not the source, it is the research assistant.
AI is not a search engine
When you search on Google, you see results with their sources. You can go to each source and verify the content directly. The verification process is integrated into the mechanism.
When you ask a language model, you receive a response that synthesises what the model “knows” — what is encoded in its parameters — without direct access to the sources. If the model states that “according to an MIT study from 2021…”, that study may exist, may exist with different results, or may not exist at all.
Models with web search access (such as Perplexity, or ChatGPT with search enabled) are significantly more reliable for recent factual information because they base their responses on verifiable sources. But even then, direct verification of the most important sources is irreplaceable.
AI is excellent for: orienting research, identifying what to search for, synthesising material you have provided, reformulating information you already know, identifying angles you had not considered.
AI does not replace: searching primary sources, verifying specific numerical data, reading original documents for important conclusions.
What AI does well in research
Initial map of the territory. When you start researching an unknown topic, AI is effective at giving you an overview: the main concepts, the currents of debate, the most relevant authors or institutions. That gives you the vocabulary to search better in primary sources.
“I am starting to research the European energy storage market. Give me a conceptual map of the main players, competing technologies and regulatory trends. I don’t need it to be exhaustive: I want to orient myself to know what to search for next.”
Identifying questions you have not asked. “I have gathered information on this topic over the past week. What important questions do I still not have answered? What angles might I be missing?”
Contextualising data you have. If you have a specific piece of data — from a verified source — AI can help you contextualise it: “This study shows a drop-off rate of 67% in the first 48 hours. Is this high or low compared to the norm for fitness apps?”
Summarising long documents
AI is especially useful for reducing the time it takes to extract value from long documents: reports, contracts, academic articles, meeting transcripts.
Effective method:
- Paste the document (or the relevant part) into the context
- Specify what you want to extract: not “summarise this,” but “identify the three contractual obligations that most affect us as a supplier” or “extract all deadlines mentioned and the context of each”
- Ask it to cite the original section for each important point
The citation instruction is critical. “For each point you mention, indicate in which paragraph or page of the document that information appears.” This lets you verify directly in the original rather than blindly trusting the summary.
For very long documents that do not fit in the model’s context, there are two strategies: divide the document into sections and process them separately, or use tools with extended context (Google’s NotebookLM, designed specifically for this case, is free and very good).
Synthesising multiple sources
When you have several sources — three articles, five interviews, multiple reports — AI is very useful for finding the common thread and the contradictions:
“I have collected these three analyses of the microchip market. [Texts]. Identify: the points where all three agree, the points where they differ, and the questions none of them satisfactorily answers.”
Or to generate structured syntheses: “Based on these documents, produce an executive summary of no more than two pages for a manager who has not read any of them. Prioritise practical implications over technical detail.”
The key is that you provide the sources. The model synthesises material that is already verified — because you selected it. The synthesis may contain interpretation errors, but it will not invent sources that do not exist.
The verification protocol
For any factual information you are going to use in something important, apply this three-step protocol:
Step 1: Identify the key claims. Not everything needs the same level of verification. Identify which claims carry the most weight in your conclusion or decision.
Step 2: Verify the highest-impact claims. For each one, find a primary source: the original article, the official report, the public database. Do not use the model to verify what the model said — that is circular.
Step 3: Treat claims without a clear source with more caution. If the model asserts something specific and you cannot find the source, treat it as unverified until you confirm it. It may be correct; it may also be a plausible hallucination.
The goal is not to distrust everything AI produces — that makes it unusable. It is to calibrate trust: high for summaries of documents you have provided, medium for general guidance on familiar topics, low for specific numerical data and concrete bibliographic references.
