You have a document to write, or check. Here’s a guide to AI-assisted research that will save you (some) time and keep you honest.
There’s a tightrope we walk in these strange days at the end of 2026.
We know that using generative AI tools like Claude or ChatGPT is fast and easy. And that doesn’t matter when what we want is a recipe or a diagnosis of an ailing potplant (in my house, that would be all of them if they ain’t already dead).
But it matters when we are doing research, or writing, or editing, or dealing with any mission-critical documents. That’s when all sorts of doubts and worries set in.
Is the AI hallucinating? Have nuances been missed? Is this information reliable? Are these real or invented sources?
My training partner Anne Taylor and I recently did a webinar for a group of editors precisely about this: Using generative AI tools for fact-checking, source finding and general research.
After the workshop, Anne generated a guide to AI-assisted research (using her trusty sidekick ChatGPT). We liked it so much we thought it would make a useful blog post, to spread more widely.
Here’s that guide.
General recommended workflow:
- Use AI to help investigate. Think of it as an intern, ready and willing to help
- Use evidence to establish what is true.
- Apply your own judgement.
All the things you need to tell your ‘AI intern’
- What you want: An investigation, source list, comparison, verification, editorial query.
- Why you want it: The context, audience, intended use.
- What evidence matters: Original research, government records, datasets, independent reporting, supplied documents.
- What you need back: Links, explanations, uncertainty, specific passages, suggested corrections.
- What the boundaries are: No invented information, no unsupported assumptions and no rewriting.
Start simply. You can always follow up if necessary. It’s often more useful to investigate one question thoroughly than to request a comprehensive answer to everything at once.
Prompt ‘ingredients’
Remember, you’re asking AI to find and interrogate evidence, rather than just to give you information, like this: Investigate [topic]. Find original, authoritative sources, check for conflicting evidence and explain what the sources establish. Provide links so I can verify the findings myself.
Ready-to-use prompt example
- Fact-check this entry in a table that lists strategic moves by countries to improve their position in the critical-minerals supply chain: <INSERT CLAIM>
- Find the original, authoritative source(s) and check whether the statement is accurate as written.
- Explain any errors or misleading wording, with links to your sources.
- Precision is key. Do not rely on sources that simply repeat the claim.
- Suggest a concise correction if needed. Mark any changes in bold.
Adapt it: Replace the table context with your document’s subject. Add dates, geographical scope or other requirements where relevant.
Trace a statistic
- Find the original source for this statistic: <INSERT STATISTIC>
- Trace the figure back to the original dataset, research report or publication.
- Provide a direct link and identify the publication date, geographical coverage, relevant definitions and methodology.
- Tell me whether the original source supports the figure exactly as stated.
- Do not give me sources that merely repeat the statistic. Flag anything you cannot verify.
Watch for: Outdated figures, different reporting periods, misleading comparisons, incorrect units and confusion between estimates, forecasts and observed data.
Find sources for research
I’m researching <TOPIC> for <PURPOSE/AUDIENCE>
- Find five reliable, relevant sources to help me understand the subject.
- Prioritise original research, official statistics, academic publications and authoritative organisations. Include independent reporting where useful.
- For each source, provide a direct link, publication date, a brief summary of what it covers and an explanation of why it’s useful.
- Identify significant disagreements, gaps in the evidence and questions requiring further investigation.
Refine the results: Ask for South African sources, a particular date range, dissenting research, recent developments or different disciplinary perspectives. Perplexity is a good tool to use for working with sources.
Investigate conflicting information
These two sources appear to disagree about <TOPIC>: <INSERT LINKS OR EXTRACTS>
- Compare the evidence and identify possible reasons for the disagreement, including differences in dates, definitions, methodology, geographical coverage or interpretation.
- Explain what can be established with confidence and what remains uncertain.
- Do not assume either source is correct.
Interrogate a collection of documents
For a source-based research tool such as Gemini Notebook. Load your selected, verified sources, then ask:
- Identify the main findings concerning <TOPIC>
- Highlight agreements, disagreements and gaps between the sources.
- Cite the relevant document and passage for each important finding.
- Distinguish between what the sources explicitly state and what can only be inferred.
- Identify possible questions that the supplied material cannot answer.
Remember to first read and verify your sources. And always check the tool’s interpretation.
Diving in the deep end
Jeremy Caplan published a sophisticated prompt in his Wonder Tools newsletter on 8 November 2026. It was contributed by Ole Reissmann, formerly the director of AI at Der Spiegel. Unlike our simple example, it asks AI to extract every checkable claim, classify it, investigate its evidence and create a checklist for the editor.
What’s clever about it:
- It separates individual claims within complex sentences.
- It distinguishes concrete, verifiable assertions (hard) from more interpretive claims (soft).
- It flags claims with potential legal implications (sensitive).
- It requires the AI to consider whether apparently different sources are genuinely independent.
- It asks for original evidence and contradictory findings.
- It distinguishes unverified information from demonstrably false information.
- It identifies what the editor still needs to investigate personally.
His complete, customisable prompt is available here: Help me fact-check: Wonder Tools prompt library
Caplan’s prompt library also includes tools for identifying weaknesses in writing, challenging an argument, improving clarity and interrogating assumptions. Worth checking out.
Research techniques
Follow the evidence upstream: A published claim may have a long chain of repetition behind it. Each additional layer creates opportunities for information to be simplified, exaggerated or misrepresented. Try to work backwards through that chain.
Read laterally, not just vertically: Lateral reading is a verification technique used by professional fact-checkers. Instead of evaluating a website solely by reading what it says about itself, leave the website and investigate it elsewhere. Ask who runs it? What is their expertise? How do independent sources describe the organisation? Is there evidence of a particular agenda or conflict of interest?
Search for contradictory evidence: AI tools tend to respond to the question they’re given. If you ask for evidence supporting a claim, you may receive just that – even when credible contradictory evidence exists. A useful follow-up is: What credible evidence contradicts this conclusion? Are there alternative interpretations or important qualifications that I should consider?
Distinguish between evidence and interpretation: Ask what the evidence establishes, not merely whether it supports the general direction of a statement.
Top tip: Keep a verification trail
For substantial reports and technical documents, consider recording the evidence you track down. For sensitive or high-stakes work, save copies or archival references where appropriate. Gemini Notebook would be a good place to store your evidence.
Be transparent about your AI use
Before uploading a manuscript, report or client document to a gen AI tool, consider:
- Does the client permit the use of AI tools?
- Is the information confidential, commercially sensitive or unpublished?
- What are the tool’s data retention and training settings? Do you have a paid account?
- Can the task be completed using publicly available information or a non-confidential extract?
- Does fact-checking fall within the agreed scope of the assignment?
Other verification resources
> UK House of Commons Library — How to evaluate sources: A briefing on evaluating information, including source credibility, evidence, purpose and lateral reading.
> SIFT — A practical verification approach: Developed by digital literacy specialist Mike Caulfield: Stop; Investigate the source; Find better coverage; Trace claims, quotations and media to their original context.
Final word
This has been a list rather than a blog post – but we hope some or all of it is useful in some way. Please let us know if you used any of it, or have any tips of your own!
Main picture: Sandip Karangiya, Unsplash
Previous Sensible Guide articles
Abandon trust. We all have to be journalists now – In our strange new world, there’s a tried-and-tested way to deal with the information overflow…
How do you actually fact-check something that Gen AI told you? – It’s now a known thing that tools like ChatGPT “hallucinate” (ie make things up). But how do you fact-check the stuff they say?
If any of this is relevant to your work…
You might find one of these useful:
