Research Partner, Not All-Knowing Oracle

Problem

The same tool that finds sources can invent them

I would never have been able to research and quote sources the way I can now. AI has changed who can do research. A consultant, a project lead, or a small-business owner can find studies, pull figures, and trace an argument back to where it started in an afternoon - work that used to need a research team or a week in the library.

But the same fluency that finds real sources can produce convincing fake ones: a study that doesn't exist, a quote nobody said, a statistic with the decimal in the wrong place, and a real source that says something different from what the summary claims. It all reads exactly like the real thing.

The last one is the hardest to catch. The source exists, the link works, and a quick check passes - but the source doesn't say what you're about to repeat.

The cost is public and rising. In 2023, a lawyer filed a brief built on AI research without reading the cases; several of the cases didn't exist, and the court sanctioned the lawyers (Mata v. Avianca). In 2025, Deloitte Australia partially refunded a AU$440,000 government report after a university researcher found references to research papers that don't exist and a made-up quote from a federal court judgment (AP, October 7, 2025). Deloitte didn't say AI caused the errors, but the revised report disclosed that it had used Azure OpenAI.

The failure isn't using AI for research. It's treating its answers as the source instead of as a lead to the source.

Value

Research you'd defend, done in a fraction of the time

Research that used to be out of reach: finding, reading, and comparing sources across a topic, without a research team.

Every claim traceable to where it came from: so you can answer "where did you get that?" in the room.

Errors caught before they leave your hands: the fabricated reference, the wrong figure, the source that says something else - found by you, not by your audience.

Credibility that compounds: people learn that your numbers check out.

Measure it: two counts - sources you dropped because you couldn't confirm them, and errors someone else found after you published (the goal is zero).

Context

When this pattern applies

This pattern applies whenever AI helps you gather evidence that you or others will rely on: articles, presentations, proposals, business cases, policy and legal work, anything with a statistic or a quote.

It matters most when the claim will be repeated - published, presented, put in front of a client or a court - because that's when a fabricated source does the most damage. The highest stakes: anything with legal, financial or regulatory weight, and anything published under your organization's name.

It is not a reason to avoid AI for research. It's how to use it so the speed doesn't cost you your credibility.

Solution

Let AI find the leads, then check every one at the source

Step 1

Use a tool that searches, and ask for sources, not just answers.

For research, use an AI tool that actually searches and shows its sources (Pattern 17). For every factual claim, ask where it comes from: author, title, date, and a link. A claim with no source is a guess - and a citation from a tool that can't search is a lead to look for, not a source.

Step 2

Check each source at the original.

Open the actual study, article, filing, or report - not the AI's summary, and not a search snippet about it. Confirm it exists, it says what's claimed, and the numbers match. A source is checked only when you've found the exact passage on the original page - a working link isn't enough.

Step 3

Mark what you verified, and drop what you couldn't.

Label each source verified or unverified. If you can't confirm it, leave it out or say plainly that it's unconfirmed. A smaller set of real sources beats a long list of maybes. The minimum rule under deadline: anything you'll publish or present gets checked; everything else can stay a lead.

Step 4

Follow the evidence, even when it disagrees with you.

If the original says something different from what you hoped, change the claim - not the source.

Step 5

Keep the receipts.

Save the sources you used - the link, the date you checked, the exact passage - with the work itself, not in someone's browser history, so anyone can retrace it later (Pattern 12, and The Chat Is Not Your Memory).

Notes

A partner, not an oracle

This library is built this way. Every statistic on this site was checked at its source. An October 15, 2026 source check found four errors on the live site and fixed them, and claims that couldn’t be confirmed were left out. Pasted sources are saved word for word, so anyone can retrace them.

Read the sources you rely on yourself. AI can now produce far more than anyone can check: when Anthropic studied its own workforce, reading everything employees’ agents produced would have taken half the work week (Wilmers, Catipon, Massenkoff and McCrory, “How Employment is Changing at Anthropic,” Anthropic, October 16, 2026). When output outruns review, fabricated sources ship. The sources a decision rests on are the ones to read in full, yourself.

Related. Pattern 03, Test Adversarially, attacks the AI’s reasoning; this pattern is about its evidence. Pattern 10, Models Are Not Good at Math, is the same idea for numbers. Pattern 17, Right Tool, Right Job: use a tool that searches and cites for research.