Guides
AI-assisted business writing

How to fact-check an AI-assisted white paper

A white paper drafted with a model needs its claims checked before it goes out under your name. We set out a method for doing that, starting with the claims that support your argument.

A firm publishes a white paper to show prospective clients that it understands a problem well. Readers take each claim in it as a reflection of the firm’s judgment. A white paper can also stay in circulation far longer than a memo or an email: it may sit on the firm’s website for years and be quoted by others long after it was written. An error in a document like that can keep doing damage for as long as people read it.

A model can handle the form of a white paper well because the conventions are familiar. The trouble is that the form is dense with claims. A fluent sentence from a model is no evidence that what it says is true. Checking a model-drafted white paper therefore takes more than a careful read-through. It helps to work through the draft in a set order.

Start by listing the claims

Before checking anything, go through the draft and mark every statement that could be false. Figures and named studies are the obvious ones. Your list should also include characterizations (“most firms in the sector”, “the standard approach”), causal claims (“this leads to”), and comparisons (“faster than”, “the first to”). In a long draft, the list can run to dozens of items. With every claim on one list, you can see how much of the paper rests on statements that haven’t been confirmed yet.

Then sort the list into the claims your argument depends on and the ones that are there for color. If your recommendation rests on a figure in the executive summary, trace that figure to its source. A passing statistic in a background paragraph needs the same treatment, but it’s often better to cut it altogether.

Read the source for every claim your argument depends on

For each of those claims, find the original document and read the passage itself, rather than the abstract, a summary, or another article that cites it. Some errors in model-drafted work are real facts described wrongly: a cautious finding presented as settled, say, or a study of one sector applied to another. You’ll only see the difference in the source.

Where the claim is a number, note the source and the year, and write down how the source defines what it’s counting. One report’s “revenue” may be another’s “turnover”, and “users” can mean almost anything. If two sources disagree, give both figures in the paper rather than picking the more convenient one. Readers who know the field are likely to know both numbers.

Treat every reference as unverified

A model-drafted white paper can cite freely, and every citation in it can look right. Some references may not exist at all. Others may be real papers that say something different from what the draft claims, or that are listed with the wrong year or authors. Open each reference and confirm both that it exists and that it supports the sentence it’s attached to. A single fabricated entry can make a reader doubt every other item on the list.

Claims that come without a reference

Some statements arrive with no citation at all: “many organizations struggle with”, “it’s widely accepted that”, “the industry has moved toward”. There’s nothing to look these up in, so they’re the hardest to verify. For each one, ask whether you know it to be true from your own work. If you do, rephrase it as your firm’s own observation; otherwise, take it out.

Gaps in the reasoning

A white paper can get every fact right and still fail if the steps between those facts don’t follow. A model can place a finding next to a recommendation without ever explaining why the finding supports it. Read the argument as a skeptical reader would, and at each step ask whether the evidence supports the conclusion drawn from it. Where a step is missing, only the authors can supply it because it depends on what the firm knows rather than on the sources.

A record of what you checked

For each claim you verify, write one line recording the statement, the source, the page, and the person who checked it. It takes a minute or two per claim. If a journalist or a client asks months later where a figure came from, you’ll have the answer on hand. The same record shows how carefully the paper was put together if anyone asks how it was produced. Our guide to what AI text watermarks mean for your documents explains what a watermark can and can’t show about a document’s history.

What to do with claims that don’t hold up

Where a claim doesn’t survive the check, the simplest fix is usually to cut it. A white paper with fewer claims, all of which stand up, is stronger than one with more claims of which some are shaky. Cutting often improves the paper as well, since it removes padding. If removing a claim leaves a hole in the argument, the draft was relying on something it couldn’t back up. Fill the hole from what your firm actually knows, or change the conclusion so that it no longer depends on the missing step.

Where we can help

When a firm sends us an AI-assisted white paper for refinement, checking references is part of the work. We confirm that each one exists and is described accurately, and flag in the margin any source we can’t verify. You decide what to do with each one. Checking that each source supports the point it’s cited for is a separate piece of developmental work with its own quote.

White paper nearly done?

Send us the draft and its sources, and we’ll agree the scope with you before quoting.

Get a quote