How to tell a sourced claim from an assumption
6 minutes read
Every AI draft reads with the same confidence, whether a line is backed by a real source or filled in by the model. Here is how to tell the two apart before review.
A sourced claim in an AI draft traces to a specific document, dataset, or approved source you could open and check right now. An assumption is something the model filled in because the source material did not say it directly. The two read identically on the page. That is the entire problem this guide solves: how to separate them before a reviewer signs off and a client sees the work.
Why AI drafts blur the two together
Fluent writing is not the same as grounded writing, and a draft that reads well gives no clue which one you are looking at. A model under a deadline-shaped prompt will often complete a sentence the way a confident answer sounds, not the way the evidence actually supports. The gap between "the source states this" and "this is a reasonable guess" disappears in the prose, and a reader has no way to recover it after the fact.
This is not a defect unique to any one tool. It is what happens whenever a draft's confidence is decoupled from its evidence. A firm publishing that draft is the one who inherits the risk, because the client sees a finished paragraph, not the reasoning behind it.
Three checks that separate a sourced claim from an assumption
Run each claim in a draft through these three checks before it goes to review.
| Check | Sourced claim | Assumption |
|---|---|---|
| Can you open the source? | Yes, it points to a specific document or dataset | No source, or a vague reference like "industry data" |
| Does the source say this exact thing? | The source states the claim, not something adjacent | The claim extends or generalizes beyond what the source says |
| Is it marked? | Shown as sourced, with the source attached | Shown as an assumption, flagged for the reviewer to resolve |
The middle row catches the failure mode that matters most. A source can be real and still get stretched. A report that says usage grew in one region becomes, a paragraph later, a claim that usage grew everywhere. The citation is genuine. The claim it is now attached to is not what the citation said.
What a properly flagged assumption looks like
An assumption is not a problem by itself. Every draft has to fill gaps the source material does not cover, and a firm's own judgment often belongs there. The problem is an assumption dressed up as a fact, with no way for the reviewer to tell the difference without redoing the research themselves.
A properly flagged assumption states plainly what is being assumed and why, in the same place the claim appears, not buried in a footnote or left implicit. "Assuming the client's current headcount holds through Q3" is a sentence a reviewer can approve, adjust, or reject in seconds. A paragraph that quietly builds on that same assumption without naming it forces the reviewer to reconstruct the reasoning from scratch, which is slower than writing the section themselves.
A worked example
Take a competitive-landscape section in a market research draft. The source material is a set of company websites and two industry reports the team approved for this engagement.
A sourced claim, drafted correctly: "According to [Report X], three of the five largest vendors in this category added a mobile app in the past year." That sentence names the source and states only what the source states. A reviewer can open the report, find the line, and move on.
An assumption, drafted correctly: "Assuming smaller vendors follow the same pattern within 12 to 18 months, based on how this category has adopted features historically." Nothing here is presented as a fact from the report, because it is not one. It is a forecast the team is choosing to make, named as such.
The failure mode sits between those two. A draft that writes "the market is rapidly moving toward mobile-first products" reads more confidently than either version above and is backed by neither. It borrows the report's authority for a claim the report never made. That sentence is the one a reviewer has to catch, and it is exactly the kind that slips through when nothing in the draft distinguishes a citation from a forecast.
A short check before a draft goes to review
Before any AI-assisted draft moves to review, run it through this sequence:
- Pull every specific claim: a number, a date, a named fact, a comparison.
- For each one, ask whether it traces to a real source you can open, or whether it is filling a gap.
- Confirm the source actually says what the claim says, not a nearby or broader version of it.
- Check that assumptions are marked as assumptions, in the text, not just in your head as the reviewer.
- Resolve or flag anything that fails steps 2 through 4 before it reaches the client.
This is quick when a draft was built to support it and slow when it was not. A draft with no evidence attached forces a reviewer to do this from memory or from a fresh search, for every claim, every time. That is the cost most firms are already paying without naming it.
Where this fits in a firm's workflow
None of this replaces expert review. It changes what review actually is. A reviewer checking evidence that is already attached to each claim is doing a check. A reviewer starting from an unsourced paragraph is doing the research again, under the model's confident phrasing instead of their own. The reliable client work guide covers the full loop this fits into: approved sources in, visible evidence throughout, expert review before delivery.
Once a firm has a review pattern that works for one type of deliverable, it is worth turning into something repeatable rather than rebuilding it from scratch on the next similar brief. That is the subject of reusable AI workflows.
