Research-First Content: AI Drafts, a Person Owns the Facts

The safest way to use AI in content is to do the research first, let the model draft only from that research, and keep a claim ledger where a named person owns every fact. Anthropic's and Google's own documentation both say model output can be wrong and must be checked, and a research paper explains why. This post gives the process step by step, with a ledger template you can copy.

Key Takeaways

• Google's generative AI guidance says models predict likely word sequences rather than retrieve verified facts, so output can contain errors, and all of it should be manually fact-checked before publishing.
• Anthropic's hallucination guide recommends four relevant habits: allow "I don't know", extract word-for-word quotes first for long documents, have the model support each claim with a quote or retract it, and restrict it to the documents you provide. It also says these reduce hallucinations but do not eliminate them.
• A September 2025 paper by Kalai, Nachum, Vempala and Zhang argues hallucinations persist partly because training and evaluation reward guessing over admitting uncertainty.
• A claim ledger (one row per factual claim, with source, evidence, owner and date) turns "someone checked it" into something you can audit.
• Different claim types need different evidence. A price needs the live pricing page and a date; an opinion needs a label, not a source.
• The model drafts. A person decides what is true. Do not let the tool grade its own homework as the only check.

Why research has to come first

When you ask a model to write about a topic with no sources attached, it fills gaps from patterns in training data. Google's page describes this as models predicting likely word sequences. The paper above gives a theory for why the behavior persists: evaluations that give full marks for a lucky guess and zero for "I don't know" push models toward guessing. I treat that as a plausible explanation, not a settled one, and the practical conclusion does not depend on it. Give the model the facts and tell it to stay inside them.

This also fits Google's people-first questions, which ask whether content has easily verified factual errors and whether it builds trust through clear sourcing.

The process, step by step

Step 1: Write the question brief

Before any prompt, write four lines: the reader's question, the claims you expect to make, the claims you must not make without proof (prices, outcomes, rankings, legal or health statements), and who will own sign-off. Anthropic's research guidance in its prompting best practices starts with the same idea: define what a successful answer looks like.

Step 2: Build a source pack

Collect the sources yourself, or have a person choose them. Rank them:

For each source, record the URL, the date you read it and, for anything that changes, a saved copy or screenshot in your own files. If two sources conflict, write that the sources conflict, say which you trust and why.

Step 3: Extract before you write

Ask the model to pull word-for-word quotes and figures out of the pack, with the source and location for each, before drafting anything. Anthropic recommends this for long documents (it cites over 20,000 tokens) to ground the response in the actual text. Then check a sample of the extracted quotes against the originals yourself. A model can misquote, so the extract needs its own spot-check.

Step 4: Draft from the extract only

A starting prompt, wrapped as a plain spec (see a prompt structure that survives model upgrades for the full format):

Draft the section using ONLY the material in <extract>. Do not use outside
knowledge. If a claim you want to make is not supported, write [GAP: what
is missing] instead. After each factual sentence, add the extract ID in
brackets, e.g. [E7].

The bracketed IDs are what make the ledger fast to build.

Step 5: Build the claim ledger

One row per factual claim. Copy this structure into a spreadsheet.

Step 6: Run the retract pass, then the human pass

Anthropic's pattern: after drafting, have the model find a supporting quote for each claim, and remove any claim it cannot support. Use that as a first filter. Then the owner opens the source for every row of type number, date, name, quote, price and comparison. High-risk topics get a second reader. The model's "I found support" is a flag, not a verification.

Step 7: Sign off and date it

The owner's name goes in the ledger. For anything that can change, add an "As of" date in the text and a recheck date in the ledger. Corrections go through the process in your editorial and correction policy.

What each claim type needs

A hypothetical ledger example

This is an invented illustration for a fictional "Example Academy", not a real project.

Draft sentence: "The 12-week analytics bootcamp costs 90,000 rupees and has a 95% placement rate."

Result: the published sentence states the duration, the corrected fee with its date, and says nothing about placement. The model's draft contained a confident figure with no source, which is exactly what the process exists to catch. If you work in education marketing, the ASCI claims checklist for edtech ads is a good companion for outcome claims.

What this means for marketers

• This week: pick one article in progress and build a ledger for it, even a rough one. Count how many claims have no source.
• Separate roles. One person owns the question brief and sign-off. The model drafts. Do not let the same model run extraction, drafting and verification without a human reading the sources.
• Add original research where you can. A ledger shows what is borrowed; original research on a solo budget shows how to add something that is yours.
• Keep a clear trail. For a solo workflow built around this, see my real AI content workflow.
• Leave alone: low-risk claims such as obvious definitions. The ledger is for claims a reader could act on or a competitor could challenge.
• My view: the ledger will slow your first article. By the fifth, most of the time goes to reading sources, which is the part that gives the piece value.

Common mistakes

• Asking the model to "add sources" after drafting. It may attach plausible but unverified references; the order must be sources first.
• Verifying the extract but not the claim as phrased in the final text, where numbers drift during editing.
• Accepting a secondary site that cites no primary source.
• Letting "verified by AI" count as a name in the ledger.
• Skipping the "as of" date on prices and rollout status.

Caveats

Anthropic's techniques are described by the vendor, and it says they do not eliminate hallucinations. I have not measured error rates for this process, and I am not claiming one. The hypothetical ledger is invented. The explanation of why models hallucinate comes from a single paper that I summarized from its abstract; I did not review the full text.

FAQ

Can AI do the fact-checking?

It can flag unsupported claims, and Anthropic documents having the model find a quote for each claim and retract those it cannot support. A person still needs to open the sources, because the model can misread or misquote them.

What is a claim ledger?

A spreadsheet with one row per factual claim in a draft, listing the source, the exact evidence, the claim type, who verified it and when. It makes checking auditable and shows which claims have no support.

Why do AI models state false things confidently?

Google says models predict likely word sequences instead of retrieving verified facts. A 2025 paper by Kalai and colleagues argues that grading schemes reward guessing over admitting uncertainty. Both point to supplying sources and permitting "I don't know".

Should I tell the model to use only my sources?

Yes. Anthropic lists restricting the model to provided documents as a hallucination-reduction technique. Pair it with a fallback such as [GAP] so the model has something to write when the source lacks a fact.

How do I handle two sources that disagree?

Say so in the article. State both figures, name which source you trust and why (usually the primary one), and record the conflict in the ledger so a later editor sees it.

How long does a ledger take?

It depends on the piece. In my view it adds most time to the first articles and less as your source pack habits improve. I have not timed it across projects.

Do I need to disclose that AI helped draft the piece?

Google says explaining how content was created can help readers, especially where they would ask, and suggests describing how automation was used. A short process note near the byline is a reasonable approach.

Want help with this?

I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience across edtech and startup teams. If you want help building a research-first, verifiable content process for your team, see my work and reach me through the contact form at younusfardeen.com.