Ground: where each claim in an AI answer comes from
Every claim shows its sources, how sure it is, and how fresh the source is.
An AI answer sounds equally sure about everything, even the parts it made up. Ground splits the answer into claims so you can check each one. Claims with no source get flagged, and sources that disagree sit side by side. It's a live prototype, set at a fictional company.
Live prototype: Ground. Data is synthetic.
The problem: you can't tell which claims have a source
In a normal AI answer, a made-up sentence reads exactly like a well sourced one.
So people trust all of it, until a confident number turns out to be invented. A disclaimer at the bottom doesn't help. You need to see the support next to each claim.
| Normal answer no sources shown | Ground sources on every claim | |
|---|---|---|
| Where claims come from | Unknown, sources are hidden if there are any | Each claim links to its sources |
| How sure it is | Same tone for everything | A meter on each claim |
| How fresh | No idea if a source is new or years old | "Updated 3 days ago" or "16 months · stale" on each claim |
| Claims with no source | Look the same as the rest | Flagged "no source found," one click to soften the wording |
| Sources that disagree | Blended into one answer | Shown side by side, you pick one or leave it flagged |
The idea: check the answer one claim at a time
Ground treats the answer as a set of claims, not one block of text.
Click a claim to see its sources, how recent they are, and how sure the claim is. A claim with no source says so.
One claim and its sources, built in the same design system as the app above.
How sure and how fresh, on each claim
One score for the whole answer would hide the weak claims. So each claim gets its own meter and its own date.
Two claims in one sentence can differ a lot. One source was updated three days ago, the other is 16 months old. Showing both lets you spot the weak one.
How sure and how fresh each claim is, built in the same design system as the app above.
Sources that disagree, shown side by side
When two good sources disagree, AI tends to blend them into one confident sentence. Ground shows both instead.
A banner says the sources disagree, and you decide. Pick one, or leave the claim flagged.
Two sources that disagree, built in the same design system as the app above.
Try it: soften a claim and watch the two meters
Click the claim with no source and soften its wording. Then settle the conflict by picking a source.
Softening lowers the count of claims that say more than the sources do. It doesn't raise the count of claims with a source, because rewording isn't evidence. The review log lists every claim you checked.
How the design changed from v1 to v6
It started as a list of links under the answer. Each version moved the sources closer to the claim they support.
-
1
A list of sources at the end
Links under the answer. Nobody read them, and none pointed to a specific claim.
-
2
+ Footnote numbers
Numbers in the text. Better, but a number can't tell you if the source really says that.
-
3
+ Clickable claims
Each claim could be clicked to show its own sources.
-
4
+ How sure & how fresh
A meter and a date on each claim, so one old source stopped looking as solid as several new ones.
-
5
+ "No source found" flags
Claims with no source got a red flag and a one-click way to soften the wording.
-
6
+ Disagreements & two meters · current
Sources that disagree shown side by side. Two meters count claims with a source and claims that say too much.
The other three prototypes in the lab
Steer agrees the brief before the AI writes. Oversee keeps AI agents in check. Recall shows what the AI remembers about you, and lets you delete it.