Hi.
Provenance

An answer that knows what it is guessing

When a person reads something a machine wrote, the true parts and the invented parts look identical. This is the page about closing that gap: making the output say, of itself, which parts it stands behind and which it is guessing at.

After the platform answers a question, a second model reads the answer back, sentence by sentence, against the material the first one was given to work from. It sorts each sentence into three kinds: grounded, meaning it traces to the source; inferred, meaning it was reasoned to but not stated outright; and unsupported, meaning nothing in the given material backs it, so it may be the model reaching for its own general knowledge, or simply wrong.

The reader sees the result on the page. The grounded sentences sit plainly. The inferred and unsupported ones are marked, and you can click a mark to see why it was flagged and what the auditor could or could not find behind it. Every one of these audits is kept, so the judgement stays on the record.

How an answer marks its own inferred and unsupported claims A question goes to the main model, which pulls the site's own content and drafts an answer. A second model audits the draft sentence by sentence into grounded, inferred and unsupported, marks the inferred and unsupported sentences as clickable, persists the audit, and returns an answer that flags its own guesses. A question The main model answers pulling the site's own content A draft answer A second model audits it sentence by sentence, against the retrieved content grounded traces to source inferred reasoned to unsupported nothing backs it Mark the inferred and unsupported each mark clickable, opening its reasoning Persist the audit An answer that flags its own guesses the answer still comes; it no longer pretends to certainty
The tag is a second model's judgement, not a literal-copy check.

By design, the marking shows the uncertain sentences rather than removing them: better that a reader sees a flagged claim and weighs it than have it quietly deleted. The judgement is made by a model, a second and cheaper one reading the first, and making that reader stronger is where we look next.

What it changes is small and large at once. Small, because the answer still comes. Large, because the answer no longer pretends to a confidence it does not have.

Back to Reflecting  ·  All posts